diff --git "a/openai_whisper-large-v3_886MB/AudioEncoder.mlmodelc/model.mil" "b/openai_whisper-large-v3_886MB/AudioEncoder.mlmodelc/model.mil" new file mode 100644--- /dev/null +++ "b/openai_whisper-large-v3_886MB/AudioEncoder.mlmodelc/model.mil" @@ -0,0 +1,3758 @@ +program(1.0) +[buildInfo = dict, tensor>({{"coremlc-component-MIL", "5.33.5"}, {"coremlc-version", "1877.40.3"}})] +{ + func main(tensor melspectrogram_features) { + tensor var_90 = const()[name = tensor("op_90"), val = tensor([1, 1])]; + tensor var_96 = const()[name = tensor("op_96"), val = tensor([1, 1])]; + tensor var_101 = const()[name = tensor("op_101"), val = tensor(1)]; + tensor var_106_pad_type_0 = const()[name = tensor("op_106_pad_type_0"), val = tensor("custom")]; + tensor var_106_pad_0 = const()[name = tensor("op_106_pad_0"), val = tensor([0, 0, 1, 1])]; + tensor op_81_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(245888))), name = tensor("op_81_to_fp16_palettized"), shape = tensor([1280, 128, 1, 3])]; + tensor var_87_to_fp16 = const()[name = tensor("op_87_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(246016)))]; + tensor var_106_cast_fp16 = conv(bias = var_87_to_fp16, dilations = var_96, groups = var_101, pad = var_106_pad_0, pad_type = var_106_pad_type_0, strides = var_90, weight = op_81_to_fp16_palettized, x = melspectrogram_features)[name = tensor("op_106_cast_fp16")]; + tensor hidden_states_1_mode_0 = const()[name = tensor("hidden_states_1_mode_0"), val = tensor("EXACT")]; + tensor hidden_states_1_cast_fp16 = gelu(mode = hidden_states_1_mode_0, x = var_106_cast_fp16)[name = tensor("hidden_states_1_cast_fp16")]; + tensor var_130 = const()[name = tensor("op_130"), val = tensor([2, 2])]; + tensor var_136 = const()[name = tensor("op_136"), val = tensor([1, 1])]; + tensor var_141 = const()[name = tensor("op_141"), val = tensor(1)]; + tensor var_146_pad_type_0 = const()[name = tensor("op_146_pad_type_0"), val = tensor("custom")]; + tensor var_146_pad_0 = const()[name = tensor("op_146_pad_0"), val = tensor([0, 0, 1, 1])]; + tensor op_121_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(248640))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3935104))), name = tensor("op_121_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 3])]; + tensor var_127_to_fp16 = const()[name = tensor("op_127_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3935296)))]; + tensor var_146_cast_fp16 = conv(bias = var_127_to_fp16, dilations = var_136, groups = var_141, pad = var_146_pad_0, pad_type = var_146_pad_type_0, strides = var_130, weight = op_121_to_fp16_palettized, x = hidden_states_1_cast_fp16)[name = tensor("op_146_cast_fp16")]; + tensor hidden_states_3_mode_0 = const()[name = tensor("hidden_states_3_mode_0"), val = tensor("EXACT")]; + tensor hidden_states_3_cast_fp16 = gelu(mode = hidden_states_3_mode_0, x = var_146_cast_fp16)[name = tensor("hidden_states_3_cast_fp16")]; + tensor var_164_to_fp16 = const()[name = tensor("op_164_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3937920)))]; + tensor inputs_1_cast_fp16 = add(x = hidden_states_3_cast_fp16, y = var_164_to_fp16)[name = tensor("inputs_1_cast_fp16")]; + tensor var_178 = const()[name = tensor("op_178"), val = tensor(3)]; + tensor var_180 = const()[name = tensor("op_180"), val = tensor(1)]; + tensor var_181 = const()[name = tensor("op_181"), val = tensor(true)]; + tensor var_191 = const()[name = tensor("op_191"), val = tensor([1])]; + tensor channels_mean_1_cast_fp16 = reduce_mean(axes = var_191, keep_dims = var_181, x = inputs_1_cast_fp16)[name = tensor("channels_mean_1_cast_fp16")]; + tensor zero_mean_1_cast_fp16 = sub(x = inputs_1_cast_fp16, y = channels_mean_1_cast_fp16)[name = tensor("zero_mean_1_cast_fp16")]; + tensor zero_mean_sq_1_cast_fp16 = mul(x = zero_mean_1_cast_fp16, y = zero_mean_1_cast_fp16)[name = tensor("zero_mean_sq_1_cast_fp16")]; + tensor var_195 = const()[name = tensor("op_195"), val = tensor([1])]; + tensor var_196_cast_fp16 = reduce_mean(axes = var_195, keep_dims = var_181, x = zero_mean_sq_1_cast_fp16)[name = tensor("op_196_cast_fp16")]; + tensor var_197_to_fp16 = const()[name = tensor("op_197_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_198_cast_fp16 = add(x = var_196_cast_fp16, y = var_197_to_fp16)[name = tensor("op_198_cast_fp16")]; + tensor denom_1_epsilon_0_to_fp16 = const()[name = tensor("denom_1_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_1_cast_fp16 = rsqrt(epsilon = denom_1_epsilon_0_to_fp16, x = var_198_cast_fp16)[name = tensor("denom_1_cast_fp16")]; + tensor out_1_cast_fp16 = mul(x = zero_mean_1_cast_fp16, y = denom_1_cast_fp16)[name = tensor("out_1_cast_fp16")]; + tensor obj_1_mean_0_to_fp16 = const()[name = tensor("obj_1_mean_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7777984)))]; + tensor obj_1_variance_0_to_fp16 = const()[name = tensor("obj_1_variance_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7780608)))]; + tensor obj_1_gamma_0_to_fp16 = const()[name = tensor("obj_1_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7783232)))]; + tensor obj_1_beta_0_to_fp16 = const()[name = tensor("obj_1_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7785856)))]; + tensor obj_1_epsilon_0_to_fp16 = const()[name = tensor("obj_1_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_1_cast_fp16 = batch_norm(beta = obj_1_beta_0_to_fp16, epsilon = obj_1_epsilon_0_to_fp16, gamma = obj_1_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_1_cast_fp16)[name = tensor("obj_1_cast_fp16")]; + tensor layers_0_self_attn_q_proj_input_shift_to_fp16 = const()[name = tensor("layers_0_self_attn_q_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7788480)))]; + tensor input_1_cast_fp16 = sub(x = obj_1_cast_fp16, y = layers_0_self_attn_q_proj_input_shift_to_fp16)[name = tensor("input_1_cast_fp16")]; + tensor var_217 = const()[name = tensor("op_217"), val = tensor([1, 1])]; + tensor var_219 = const()[name = tensor("op_219"), val = tensor([1, 1])]; + tensor x_1_pad_type_0 = const()[name = tensor("x_1_pad_type_0"), val = tensor("custom")]; + tensor x_1_pad_0 = const()[name = tensor("x_1_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_0_self_attn_q_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7791104))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8610368))), name = tensor("layers_0_self_attn_q_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_0_self_attn_q_proj_module_bias_to_fp16 = const()[name = tensor("layers_0_self_attn_q_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8610496)))]; + tensor x_1_cast_fp16 = conv(bias = layers_0_self_attn_q_proj_module_bias_to_fp16, dilations = var_219, groups = var_180, pad = x_1_pad_0, pad_type = x_1_pad_type_0, strides = var_217, weight = layers_0_self_attn_q_proj_module_weight_to_fp16_palettized, x = input_1_cast_fp16)[name = tensor("x_1_cast_fp16")]; + tensor layers_0_self_attn_q_proj_output_scale_to_fp16 = const()[name = tensor("layers_0_self_attn_q_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8613120)))]; + tensor query_1_cast_fp16 = mul(x = x_1_cast_fp16, y = layers_0_self_attn_q_proj_output_scale_to_fp16)[name = tensor("query_1_cast_fp16")]; + tensor var_229 = const()[name = tensor("op_229"), val = tensor([1, 1])]; + tensor var_231 = const()[name = tensor("op_231"), val = tensor([1, 1])]; + tensor x_3_pad_type_0 = const()[name = tensor("x_3_pad_type_0"), val = tensor("custom")]; + tensor x_3_pad_0 = const()[name = tensor("x_3_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_0_self_attn_k_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8615744))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9435008))), name = tensor("layers_0_self_attn_k_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_0_self_attn_k_proj_module_bias_to_fp16 = const()[name = tensor("layers_0_self_attn_k_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9435136)))]; + tensor x_3_cast_fp16 = conv(bias = layers_0_self_attn_k_proj_module_bias_to_fp16, dilations = var_231, groups = var_180, pad = x_3_pad_0, pad_type = x_3_pad_type_0, strides = var_229, weight = layers_0_self_attn_k_proj_module_weight_to_fp16_palettized, x = input_1_cast_fp16)[name = tensor("x_3_cast_fp16")]; + tensor layers_0_self_attn_k_proj_output_scale_to_fp16 = const()[name = tensor("layers_0_self_attn_k_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9437760)))]; + tensor key_1_cast_fp16 = mul(x = x_3_cast_fp16, y = layers_0_self_attn_k_proj_output_scale_to_fp16)[name = tensor("key_1_cast_fp16")]; + tensor var_241 = const()[name = tensor("op_241"), val = tensor([1, 1])]; + tensor var_243 = const()[name = tensor("op_243"), val = tensor([1, 1])]; + tensor x_5_pad_type_0 = const()[name = tensor("x_5_pad_type_0"), val = tensor("custom")]; + tensor x_5_pad_0 = const()[name = tensor("x_5_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_0_self_attn_v_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9440384))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10259648))), name = tensor("layers_0_self_attn_v_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_0_self_attn_v_proj_module_bias_to_fp16 = const()[name = tensor("layers_0_self_attn_v_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10259776)))]; + tensor x_5_cast_fp16 = conv(bias = layers_0_self_attn_v_proj_module_bias_to_fp16, dilations = var_243, groups = var_180, pad = x_5_pad_0, pad_type = x_5_pad_type_0, strides = var_241, weight = layers_0_self_attn_v_proj_module_weight_to_fp16_palettized, x = input_1_cast_fp16)[name = tensor("x_5_cast_fp16")]; + tensor layers_0_self_attn_v_proj_output_scale_to_fp16 = const()[name = tensor("layers_0_self_attn_v_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10262400)))]; + tensor value_1_cast_fp16 = mul(x = x_5_cast_fp16, y = layers_0_self_attn_v_proj_output_scale_to_fp16)[name = tensor("value_1_cast_fp16")]; + tensor var_248 = const()[name = tensor("op_248"), val = tensor([1, 20, 64, -1])]; + tensor var_249_cast_fp16 = reshape(shape = var_248, x = query_1_cast_fp16)[name = tensor("op_249_cast_fp16")]; + tensor var_250_to_fp16 = const()[name = tensor("op_250_to_fp16"), val = tensor(0x1p-3)]; + tensor var_251_cast_fp16 = mul(x = var_249_cast_fp16, y = var_250_to_fp16)[name = tensor("op_251_cast_fp16")]; + tensor var_252 = const()[name = tensor("op_252"), val = tensor([1, 20, 64, -1])]; + tensor var_253_cast_fp16 = reshape(shape = var_252, x = key_1_cast_fp16)[name = tensor("op_253_cast_fp16")]; + tensor mh_w_1_transpose_x_0 = const()[name = tensor("mh_w_1_transpose_x_0"), val = tensor(true)]; + tensor mh_w_1_transpose_y_0 = const()[name = tensor("mh_w_1_transpose_y_0"), val = tensor(false)]; + tensor mh_w_1_cast_fp16 = matmul(transpose_x = mh_w_1_transpose_x_0, transpose_y = mh_w_1_transpose_y_0, x = var_251_cast_fp16, y = var_253_cast_fp16)[name = tensor("mh_w_1_cast_fp16")]; + tensor var_256_cast_fp16 = softmax(axis = var_178, x = mh_w_1_cast_fp16)[name = tensor("op_256_cast_fp16")]; + tensor var_257 = const()[name = tensor("op_257"), val = tensor([1, 20, 64, -1])]; + tensor var_258_cast_fp16 = reshape(shape = var_257, x = value_1_cast_fp16)[name = tensor("op_258_cast_fp16")]; + tensor attn_1_transpose_x_0 = const()[name = tensor("attn_1_transpose_x_0"), val = tensor(false)]; + tensor attn_1_transpose_y_0 = const()[name = tensor("attn_1_transpose_y_0"), val = tensor(true)]; + tensor attn_1_cast_fp16 = matmul(transpose_x = attn_1_transpose_x_0, transpose_y = attn_1_transpose_y_0, x = var_258_cast_fp16, y = var_256_cast_fp16)[name = tensor("attn_1_cast_fp16")]; + tensor var_261 = const()[name = tensor("op_261"), val = tensor([1, 1280, 1, -1])]; + tensor x_7_cast_fp16 = reshape(shape = var_261, x = attn_1_cast_fp16)[name = tensor("x_7_cast_fp16")]; + tensor layers_0_self_attn_o_proj_input_shift_to_fp16 = const()[name = tensor("layers_0_self_attn_o_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10265024)))]; + tensor input_7_cast_fp16 = sub(x = x_7_cast_fp16, y = layers_0_self_attn_o_proj_input_shift_to_fp16)[name = tensor("input_7_cast_fp16")]; + tensor var_269 = const()[name = tensor("op_269"), val = tensor([1, 1])]; + tensor var_271 = const()[name = tensor("op_271"), val = tensor([1, 1])]; + tensor x_9_pad_type_0 = const()[name = tensor("x_9_pad_type_0"), val = tensor("custom")]; + tensor x_9_pad_0 = const()[name = tensor("x_9_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_0_self_attn_o_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10267648))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11086912))), name = tensor("layers_0_self_attn_o_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_0_self_attn_o_proj_module_bias_to_fp16 = const()[name = tensor("layers_0_self_attn_o_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11087040)))]; + tensor x_9_cast_fp16 = conv(bias = layers_0_self_attn_o_proj_module_bias_to_fp16, dilations = var_271, groups = var_180, pad = x_9_pad_0, pad_type = x_9_pad_type_0, strides = var_269, weight = layers_0_self_attn_o_proj_module_weight_to_fp16_palettized, x = input_7_cast_fp16)[name = tensor("x_9_cast_fp16")]; + tensor layers_0_self_attn_o_proj_output_scale_to_fp16 = const()[name = tensor("layers_0_self_attn_o_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11089664)))]; + tensor obj_3_cast_fp16 = mul(x = x_9_cast_fp16, y = layers_0_self_attn_o_proj_output_scale_to_fp16)[name = tensor("obj_3_cast_fp16")]; + tensor inputs_3_cast_fp16 = add(x = inputs_1_cast_fp16, y = obj_3_cast_fp16)[name = tensor("inputs_3_cast_fp16")]; + tensor var_278 = const()[name = tensor("op_278"), val = tensor([1])]; + tensor channels_mean_3_cast_fp16 = reduce_mean(axes = var_278, keep_dims = var_181, x = inputs_3_cast_fp16)[name = tensor("channels_mean_3_cast_fp16")]; + tensor zero_mean_3_cast_fp16 = sub(x = inputs_3_cast_fp16, y = channels_mean_3_cast_fp16)[name = tensor("zero_mean_3_cast_fp16")]; + tensor zero_mean_sq_3_cast_fp16 = mul(x = zero_mean_3_cast_fp16, y = zero_mean_3_cast_fp16)[name = tensor("zero_mean_sq_3_cast_fp16")]; + tensor var_282 = const()[name = tensor("op_282"), val = tensor([1])]; + tensor var_283_cast_fp16 = reduce_mean(axes = var_282, keep_dims = var_181, x = zero_mean_sq_3_cast_fp16)[name = tensor("op_283_cast_fp16")]; + tensor var_284_to_fp16 = const()[name = tensor("op_284_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_285_cast_fp16 = add(x = var_283_cast_fp16, y = var_284_to_fp16)[name = tensor("op_285_cast_fp16")]; + tensor denom_3_epsilon_0_to_fp16 = const()[name = tensor("denom_3_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_3_cast_fp16 = rsqrt(epsilon = denom_3_epsilon_0_to_fp16, x = var_285_cast_fp16)[name = tensor("denom_3_cast_fp16")]; + tensor out_3_cast_fp16 = mul(x = zero_mean_3_cast_fp16, y = denom_3_cast_fp16)[name = tensor("out_3_cast_fp16")]; + tensor x_11_gamma_0_to_fp16 = const()[name = tensor("x_11_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11092288)))]; + tensor x_11_beta_0_to_fp16 = const()[name = tensor("x_11_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11094912)))]; + tensor x_11_epsilon_0_to_fp16 = const()[name = tensor("x_11_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_11_cast_fp16 = batch_norm(beta = x_11_beta_0_to_fp16, epsilon = x_11_epsilon_0_to_fp16, gamma = x_11_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_3_cast_fp16)[name = tensor("x_11_cast_fp16")]; + tensor layers_0_fc1_input_shift_to_fp16 = const()[name = tensor("layers_0_fc1_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11097536)))]; + tensor input_9_cast_fp16 = sub(x = x_11_cast_fp16, y = layers_0_fc1_input_shift_to_fp16)[name = tensor("input_9_cast_fp16")]; + tensor var_300 = const()[name = tensor("op_300"), val = tensor([1, 1])]; + tensor var_302 = const()[name = tensor("op_302"), val = tensor([1, 1])]; + tensor x_13_pad_type_0 = const()[name = tensor("x_13_pad_type_0"), val = tensor("custom")]; + tensor x_13_pad_0 = const()[name = tensor("x_13_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_0_fc1_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11100160))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14377024))), name = tensor("layers_0_fc1_module_weight_to_fp16_palettized"), shape = tensor([5120, 1280, 1, 1])]; + tensor layers_0_fc1_module_bias_to_fp16 = const()[name = tensor("layers_0_fc1_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14377152)))]; + tensor x_13_cast_fp16 = conv(bias = layers_0_fc1_module_bias_to_fp16, dilations = var_302, groups = var_180, pad = x_13_pad_0, pad_type = x_13_pad_type_0, strides = var_300, weight = layers_0_fc1_module_weight_to_fp16_palettized, x = input_9_cast_fp16)[name = tensor("x_13_cast_fp16")]; + tensor layers_0_fc1_output_scale_to_fp16 = const()[name = tensor("layers_0_fc1_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14387456)))]; + tensor input_11_cast_fp16 = mul(x = x_13_cast_fp16, y = layers_0_fc1_output_scale_to_fp16)[name = tensor("input_11_cast_fp16")]; + tensor x_15_mode_0 = const()[name = tensor("x_15_mode_0"), val = tensor("EXACT")]; + tensor x_15_cast_fp16 = gelu(mode = x_15_mode_0, x = input_11_cast_fp16)[name = tensor("x_15_cast_fp16")]; + tensor layers_0_fc2_input_shift_to_fp16 = const()[name = tensor("layers_0_fc2_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14397760)))]; + tensor input_13_cast_fp16 = sub(x = x_15_cast_fp16, y = layers_0_fc2_input_shift_to_fp16)[name = tensor("input_13_cast_fp16")]; + tensor var_313 = const()[name = tensor("op_313"), val = tensor([1, 1])]; + tensor var_315 = const()[name = tensor("op_315"), val = tensor([1, 1])]; + tensor x_17_pad_type_0 = const()[name = tensor("x_17_pad_type_0"), val = tensor("custom")]; + tensor x_17_pad_0 = const()[name = tensor("x_17_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_0_fc2_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14408064))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(17684928))), name = tensor("layers_0_fc2_module_weight_to_fp16_palettized"), shape = tensor([1280, 5120, 1, 1])]; + tensor layers_0_fc2_module_bias_to_fp16 = const()[name = tensor("layers_0_fc2_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(17685056)))]; + tensor x_17_cast_fp16 = conv(bias = layers_0_fc2_module_bias_to_fp16, dilations = var_315, groups = var_180, pad = x_17_pad_0, pad_type = x_17_pad_type_0, strides = var_313, weight = layers_0_fc2_module_weight_to_fp16_palettized, x = input_13_cast_fp16)[name = tensor("x_17_cast_fp16")]; + tensor layers_0_fc2_output_scale_to_fp16 = const()[name = tensor("layers_0_fc2_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(17687680)))]; + tensor hidden_states_5_cast_fp16 = mul(x = x_17_cast_fp16, y = layers_0_fc2_output_scale_to_fp16)[name = tensor("hidden_states_5_cast_fp16")]; + tensor inputs_5_cast_fp16 = add(x = inputs_3_cast_fp16, y = hidden_states_5_cast_fp16)[name = tensor("inputs_5_cast_fp16")]; + tensor var_327 = const()[name = tensor("op_327"), val = tensor(3)]; + tensor var_329 = const()[name = tensor("op_329"), val = tensor(1)]; + tensor var_330 = const()[name = tensor("op_330"), val = tensor(true)]; + tensor var_340 = const()[name = tensor("op_340"), val = tensor([1])]; + tensor channels_mean_5_cast_fp16 = reduce_mean(axes = var_340, keep_dims = var_330, x = inputs_5_cast_fp16)[name = tensor("channels_mean_5_cast_fp16")]; + tensor zero_mean_5_cast_fp16 = sub(x = inputs_5_cast_fp16, y = channels_mean_5_cast_fp16)[name = tensor("zero_mean_5_cast_fp16")]; + tensor zero_mean_sq_5_cast_fp16 = mul(x = zero_mean_5_cast_fp16, y = zero_mean_5_cast_fp16)[name = tensor("zero_mean_sq_5_cast_fp16")]; + tensor var_344 = const()[name = tensor("op_344"), val = tensor([1])]; + tensor var_345_cast_fp16 = reduce_mean(axes = var_344, keep_dims = var_330, x = zero_mean_sq_5_cast_fp16)[name = tensor("op_345_cast_fp16")]; + tensor var_346_to_fp16 = const()[name = tensor("op_346_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_347_cast_fp16 = add(x = var_345_cast_fp16, y = var_346_to_fp16)[name = tensor("op_347_cast_fp16")]; + tensor denom_5_epsilon_0_to_fp16 = const()[name = tensor("denom_5_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_5_cast_fp16 = rsqrt(epsilon = denom_5_epsilon_0_to_fp16, x = var_347_cast_fp16)[name = tensor("denom_5_cast_fp16")]; + tensor out_5_cast_fp16 = mul(x = zero_mean_5_cast_fp16, y = denom_5_cast_fp16)[name = tensor("out_5_cast_fp16")]; + tensor obj_5_gamma_0_to_fp16 = const()[name = tensor("obj_5_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(17690304)))]; + tensor obj_5_beta_0_to_fp16 = const()[name = tensor("obj_5_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(17692928)))]; + tensor obj_5_epsilon_0_to_fp16 = const()[name = tensor("obj_5_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_5_cast_fp16 = batch_norm(beta = obj_5_beta_0_to_fp16, epsilon = obj_5_epsilon_0_to_fp16, gamma = obj_5_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_5_cast_fp16)[name = tensor("obj_5_cast_fp16")]; + tensor layers_1_self_attn_q_proj_input_shift_to_fp16 = const()[name = tensor("layers_1_self_attn_q_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(17695552)))]; + tensor input_15_cast_fp16 = sub(x = obj_5_cast_fp16, y = layers_1_self_attn_q_proj_input_shift_to_fp16)[name = tensor("input_15_cast_fp16")]; + tensor var_366 = const()[name = tensor("op_366"), val = tensor([1, 1])]; + tensor var_368 = const()[name = tensor("op_368"), val = tensor([1, 1])]; + tensor x_19_pad_type_0 = const()[name = tensor("x_19_pad_type_0"), val = tensor("custom")]; + tensor x_19_pad_0 = const()[name = tensor("x_19_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_1_self_attn_q_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(17698176))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18517440))), name = tensor("layers_1_self_attn_q_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_1_self_attn_q_proj_module_bias_to_fp16 = const()[name = tensor("layers_1_self_attn_q_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18517568)))]; + tensor x_19_cast_fp16 = conv(bias = layers_1_self_attn_q_proj_module_bias_to_fp16, dilations = var_368, groups = var_329, pad = x_19_pad_0, pad_type = x_19_pad_type_0, strides = var_366, weight = layers_1_self_attn_q_proj_module_weight_to_fp16_palettized, x = input_15_cast_fp16)[name = tensor("x_19_cast_fp16")]; + tensor layers_1_self_attn_q_proj_output_scale_to_fp16 = const()[name = tensor("layers_1_self_attn_q_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18520192)))]; + tensor query_3_cast_fp16 = mul(x = x_19_cast_fp16, y = layers_1_self_attn_q_proj_output_scale_to_fp16)[name = tensor("query_3_cast_fp16")]; + tensor var_378 = const()[name = tensor("op_378"), val = tensor([1, 1])]; + tensor var_380 = const()[name = tensor("op_380"), val = tensor([1, 1])]; + tensor x_21_pad_type_0 = const()[name = tensor("x_21_pad_type_0"), val = tensor("custom")]; + tensor x_21_pad_0 = const()[name = tensor("x_21_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_1_self_attn_k_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18522816))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19342080))), name = tensor("layers_1_self_attn_k_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_1_self_attn_k_proj_module_bias_to_fp16 = const()[name = tensor("layers_1_self_attn_k_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19342208)))]; + tensor x_21_cast_fp16 = conv(bias = layers_1_self_attn_k_proj_module_bias_to_fp16, dilations = var_380, groups = var_329, pad = x_21_pad_0, pad_type = x_21_pad_type_0, strides = var_378, weight = layers_1_self_attn_k_proj_module_weight_to_fp16_palettized, x = input_15_cast_fp16)[name = tensor("x_21_cast_fp16")]; + tensor layers_1_self_attn_k_proj_output_scale_to_fp16 = const()[name = tensor("layers_1_self_attn_k_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19344832)))]; + tensor key_3_cast_fp16 = mul(x = x_21_cast_fp16, y = layers_1_self_attn_k_proj_output_scale_to_fp16)[name = tensor("key_3_cast_fp16")]; + tensor var_390 = const()[name = tensor("op_390"), val = tensor([1, 1])]; + tensor var_392 = const()[name = tensor("op_392"), val = tensor([1, 1])]; + tensor x_23_pad_type_0 = const()[name = tensor("x_23_pad_type_0"), val = tensor("custom")]; + tensor x_23_pad_0 = const()[name = tensor("x_23_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_1_self_attn_v_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19347456))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20166720))), name = tensor("layers_1_self_attn_v_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_1_self_attn_v_proj_module_bias_to_fp16 = const()[name = tensor("layers_1_self_attn_v_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20166848)))]; + tensor x_23_cast_fp16 = conv(bias = layers_1_self_attn_v_proj_module_bias_to_fp16, dilations = var_392, groups = var_329, pad = x_23_pad_0, pad_type = x_23_pad_type_0, strides = var_390, weight = layers_1_self_attn_v_proj_module_weight_to_fp16_palettized, x = input_15_cast_fp16)[name = tensor("x_23_cast_fp16")]; + tensor layers_1_self_attn_v_proj_output_scale_to_fp16 = const()[name = tensor("layers_1_self_attn_v_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20169472)))]; + tensor value_3_cast_fp16 = mul(x = x_23_cast_fp16, y = layers_1_self_attn_v_proj_output_scale_to_fp16)[name = tensor("value_3_cast_fp16")]; + tensor var_397 = const()[name = tensor("op_397"), val = tensor([1, 20, 64, -1])]; + tensor var_398_cast_fp16 = reshape(shape = var_397, x = query_3_cast_fp16)[name = tensor("op_398_cast_fp16")]; + tensor var_399_to_fp16 = const()[name = tensor("op_399_to_fp16"), val = tensor(0x1p-3)]; + tensor var_400_cast_fp16 = mul(x = var_398_cast_fp16, y = var_399_to_fp16)[name = tensor("op_400_cast_fp16")]; + tensor var_401 = const()[name = tensor("op_401"), val = tensor([1, 20, 64, -1])]; + tensor var_402_cast_fp16 = reshape(shape = var_401, x = key_3_cast_fp16)[name = tensor("op_402_cast_fp16")]; + tensor mh_w_3_transpose_x_0 = const()[name = tensor("mh_w_3_transpose_x_0"), val = tensor(true)]; + tensor mh_w_3_transpose_y_0 = const()[name = tensor("mh_w_3_transpose_y_0"), val = tensor(false)]; + tensor mh_w_3_cast_fp16 = matmul(transpose_x = mh_w_3_transpose_x_0, transpose_y = mh_w_3_transpose_y_0, x = var_400_cast_fp16, y = var_402_cast_fp16)[name = tensor("mh_w_3_cast_fp16")]; + tensor var_405_cast_fp16 = softmax(axis = var_327, x = mh_w_3_cast_fp16)[name = tensor("op_405_cast_fp16")]; + tensor var_406 = const()[name = tensor("op_406"), val = tensor([1, 20, 64, -1])]; + tensor var_407_cast_fp16 = reshape(shape = var_406, x = value_3_cast_fp16)[name = tensor("op_407_cast_fp16")]; + tensor attn_3_transpose_x_0 = const()[name = tensor("attn_3_transpose_x_0"), val = tensor(false)]; + tensor attn_3_transpose_y_0 = const()[name = tensor("attn_3_transpose_y_0"), val = tensor(true)]; + tensor attn_3_cast_fp16 = matmul(transpose_x = attn_3_transpose_x_0, transpose_y = attn_3_transpose_y_0, x = var_407_cast_fp16, y = var_405_cast_fp16)[name = tensor("attn_3_cast_fp16")]; + tensor var_410 = const()[name = tensor("op_410"), val = tensor([1, 1280, 1, -1])]; + tensor x_25_cast_fp16 = reshape(shape = var_410, x = attn_3_cast_fp16)[name = tensor("x_25_cast_fp16")]; + tensor layers_1_self_attn_o_proj_input_shift_to_fp16 = const()[name = tensor("layers_1_self_attn_o_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20172096)))]; + tensor input_21_cast_fp16 = sub(x = x_25_cast_fp16, y = layers_1_self_attn_o_proj_input_shift_to_fp16)[name = tensor("input_21_cast_fp16")]; + tensor var_418 = const()[name = tensor("op_418"), val = tensor([1, 1])]; + tensor var_420 = const()[name = tensor("op_420"), val = tensor([1, 1])]; + tensor x_27_pad_type_0 = const()[name = tensor("x_27_pad_type_0"), val = tensor("custom")]; + tensor x_27_pad_0 = const()[name = tensor("x_27_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_1_self_attn_o_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20174720))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20993984))), name = tensor("layers_1_self_attn_o_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_1_self_attn_o_proj_module_bias_to_fp16 = const()[name = tensor("layers_1_self_attn_o_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20994112)))]; + tensor x_27_cast_fp16 = conv(bias = layers_1_self_attn_o_proj_module_bias_to_fp16, dilations = var_420, groups = var_329, pad = x_27_pad_0, pad_type = x_27_pad_type_0, strides = var_418, weight = layers_1_self_attn_o_proj_module_weight_to_fp16_palettized, x = input_21_cast_fp16)[name = tensor("x_27_cast_fp16")]; + tensor layers_1_self_attn_o_proj_output_scale_to_fp16 = const()[name = tensor("layers_1_self_attn_o_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20996736)))]; + tensor obj_7_cast_fp16 = mul(x = x_27_cast_fp16, y = layers_1_self_attn_o_proj_output_scale_to_fp16)[name = tensor("obj_7_cast_fp16")]; + tensor inputs_7_cast_fp16 = add(x = inputs_5_cast_fp16, y = obj_7_cast_fp16)[name = tensor("inputs_7_cast_fp16")]; + tensor var_427 = const()[name = tensor("op_427"), val = tensor([1])]; + tensor channels_mean_7_cast_fp16 = reduce_mean(axes = var_427, keep_dims = var_330, x = inputs_7_cast_fp16)[name = tensor("channels_mean_7_cast_fp16")]; + tensor zero_mean_7_cast_fp16 = sub(x = inputs_7_cast_fp16, y = channels_mean_7_cast_fp16)[name = tensor("zero_mean_7_cast_fp16")]; + tensor zero_mean_sq_7_cast_fp16 = mul(x = zero_mean_7_cast_fp16, y = zero_mean_7_cast_fp16)[name = tensor("zero_mean_sq_7_cast_fp16")]; + tensor var_431 = const()[name = tensor("op_431"), val = tensor([1])]; + tensor var_432_cast_fp16 = reduce_mean(axes = var_431, keep_dims = var_330, x = zero_mean_sq_7_cast_fp16)[name = tensor("op_432_cast_fp16")]; + tensor var_433_to_fp16 = const()[name = tensor("op_433_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_434_cast_fp16 = add(x = var_432_cast_fp16, y = var_433_to_fp16)[name = tensor("op_434_cast_fp16")]; + tensor denom_7_epsilon_0_to_fp16 = const()[name = tensor("denom_7_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_7_cast_fp16 = rsqrt(epsilon = denom_7_epsilon_0_to_fp16, x = var_434_cast_fp16)[name = tensor("denom_7_cast_fp16")]; + tensor out_7_cast_fp16 = mul(x = zero_mean_7_cast_fp16, y = denom_7_cast_fp16)[name = tensor("out_7_cast_fp16")]; + tensor x_29_gamma_0_to_fp16 = const()[name = tensor("x_29_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20999360)))]; + tensor x_29_beta_0_to_fp16 = const()[name = tensor("x_29_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21001984)))]; + tensor x_29_epsilon_0_to_fp16 = const()[name = tensor("x_29_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_29_cast_fp16 = batch_norm(beta = x_29_beta_0_to_fp16, epsilon = x_29_epsilon_0_to_fp16, gamma = x_29_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_7_cast_fp16)[name = tensor("x_29_cast_fp16")]; + tensor layers_1_fc1_input_shift_to_fp16 = const()[name = tensor("layers_1_fc1_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21004608)))]; + tensor input_23_cast_fp16 = sub(x = x_29_cast_fp16, y = layers_1_fc1_input_shift_to_fp16)[name = tensor("input_23_cast_fp16")]; + tensor var_449 = const()[name = tensor("op_449"), val = tensor([1, 1])]; + tensor var_451 = const()[name = tensor("op_451"), val = tensor([1, 1])]; + tensor x_31_pad_type_0 = const()[name = tensor("x_31_pad_type_0"), val = tensor("custom")]; + tensor x_31_pad_0 = const()[name = tensor("x_31_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_1_fc1_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21007232))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24284096))), name = tensor("layers_1_fc1_module_weight_to_fp16_palettized"), shape = tensor([5120, 1280, 1, 1])]; + tensor layers_1_fc1_module_bias_to_fp16 = const()[name = tensor("layers_1_fc1_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24284224)))]; + tensor x_31_cast_fp16 = conv(bias = layers_1_fc1_module_bias_to_fp16, dilations = var_451, groups = var_329, pad = x_31_pad_0, pad_type = x_31_pad_type_0, strides = var_449, weight = layers_1_fc1_module_weight_to_fp16_palettized, x = input_23_cast_fp16)[name = tensor("x_31_cast_fp16")]; + tensor layers_1_fc1_output_scale_to_fp16 = const()[name = tensor("layers_1_fc1_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24294528)))]; + tensor input_25_cast_fp16 = mul(x = x_31_cast_fp16, y = layers_1_fc1_output_scale_to_fp16)[name = tensor("input_25_cast_fp16")]; + tensor x_33_mode_0 = const()[name = tensor("x_33_mode_0"), val = tensor("EXACT")]; + tensor x_33_cast_fp16 = gelu(mode = x_33_mode_0, x = input_25_cast_fp16)[name = tensor("x_33_cast_fp16")]; + tensor layers_1_fc2_input_shift_to_fp16 = const()[name = tensor("layers_1_fc2_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24304832)))]; + tensor input_27_cast_fp16 = sub(x = x_33_cast_fp16, y = layers_1_fc2_input_shift_to_fp16)[name = tensor("input_27_cast_fp16")]; + tensor var_462 = const()[name = tensor("op_462"), val = tensor([1, 1])]; + tensor var_464 = const()[name = tensor("op_464"), val = tensor([1, 1])]; + tensor x_35_pad_type_0 = const()[name = tensor("x_35_pad_type_0"), val = tensor("custom")]; + tensor x_35_pad_0 = const()[name = tensor("x_35_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_1_fc2_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24315136))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27592000))), name = tensor("layers_1_fc2_module_weight_to_fp16_palettized"), shape = tensor([1280, 5120, 1, 1])]; + tensor layers_1_fc2_module_bias_to_fp16 = const()[name = tensor("layers_1_fc2_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27592128)))]; + tensor x_35_cast_fp16 = conv(bias = layers_1_fc2_module_bias_to_fp16, dilations = var_464, groups = var_329, pad = x_35_pad_0, pad_type = x_35_pad_type_0, strides = var_462, weight = layers_1_fc2_module_weight_to_fp16_palettized, x = input_27_cast_fp16)[name = tensor("x_35_cast_fp16")]; + tensor layers_1_fc2_output_scale_to_fp16 = const()[name = tensor("layers_1_fc2_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27594752)))]; + tensor hidden_states_7_cast_fp16 = mul(x = x_35_cast_fp16, y = layers_1_fc2_output_scale_to_fp16)[name = tensor("hidden_states_7_cast_fp16")]; + tensor inputs_9_cast_fp16 = add(x = inputs_7_cast_fp16, y = hidden_states_7_cast_fp16)[name = tensor("inputs_9_cast_fp16")]; + tensor var_476 = const()[name = tensor("op_476"), val = tensor(3)]; + tensor var_478 = const()[name = tensor("op_478"), val = tensor(1)]; + tensor var_479 = const()[name = tensor("op_479"), val = tensor(true)]; + tensor var_489 = const()[name = tensor("op_489"), val = tensor([1])]; + tensor channels_mean_9_cast_fp16 = reduce_mean(axes = var_489, keep_dims = var_479, x = inputs_9_cast_fp16)[name = tensor("channels_mean_9_cast_fp16")]; + tensor zero_mean_9_cast_fp16 = sub(x = inputs_9_cast_fp16, y = channels_mean_9_cast_fp16)[name = tensor("zero_mean_9_cast_fp16")]; + tensor zero_mean_sq_9_cast_fp16 = mul(x = zero_mean_9_cast_fp16, y = zero_mean_9_cast_fp16)[name = tensor("zero_mean_sq_9_cast_fp16")]; + tensor var_493 = const()[name = tensor("op_493"), val = tensor([1])]; + tensor var_494_cast_fp16 = reduce_mean(axes = var_493, keep_dims = var_479, x = zero_mean_sq_9_cast_fp16)[name = tensor("op_494_cast_fp16")]; + tensor var_495_to_fp16 = const()[name = tensor("op_495_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_496_cast_fp16 = add(x = var_494_cast_fp16, y = var_495_to_fp16)[name = tensor("op_496_cast_fp16")]; + tensor denom_9_epsilon_0_to_fp16 = const()[name = tensor("denom_9_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_9_cast_fp16 = rsqrt(epsilon = denom_9_epsilon_0_to_fp16, x = var_496_cast_fp16)[name = tensor("denom_9_cast_fp16")]; + tensor out_9_cast_fp16 = mul(x = zero_mean_9_cast_fp16, y = denom_9_cast_fp16)[name = tensor("out_9_cast_fp16")]; + tensor obj_9_gamma_0_to_fp16 = const()[name = tensor("obj_9_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27597376)))]; + tensor obj_9_beta_0_to_fp16 = const()[name = tensor("obj_9_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27600000)))]; + tensor obj_9_epsilon_0_to_fp16 = const()[name = tensor("obj_9_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_9_cast_fp16 = batch_norm(beta = obj_9_beta_0_to_fp16, epsilon = obj_9_epsilon_0_to_fp16, gamma = obj_9_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_9_cast_fp16)[name = tensor("obj_9_cast_fp16")]; + tensor layers_2_self_attn_q_proj_input_shift_to_fp16 = const()[name = tensor("layers_2_self_attn_q_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27602624)))]; + tensor input_29_cast_fp16 = sub(x = obj_9_cast_fp16, y = layers_2_self_attn_q_proj_input_shift_to_fp16)[name = tensor("input_29_cast_fp16")]; + tensor var_515 = const()[name = tensor("op_515"), val = tensor([1, 1])]; + tensor var_517 = const()[name = tensor("op_517"), val = tensor([1, 1])]; + tensor x_37_pad_type_0 = const()[name = tensor("x_37_pad_type_0"), val = tensor("custom")]; + tensor x_37_pad_0 = const()[name = tensor("x_37_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_2_self_attn_q_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27605248))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28424512))), name = tensor("layers_2_self_attn_q_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_2_self_attn_q_proj_module_bias_to_fp16 = const()[name = tensor("layers_2_self_attn_q_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28424640)))]; + tensor x_37_cast_fp16 = conv(bias = layers_2_self_attn_q_proj_module_bias_to_fp16, dilations = var_517, groups = var_478, pad = x_37_pad_0, pad_type = x_37_pad_type_0, strides = var_515, weight = layers_2_self_attn_q_proj_module_weight_to_fp16_palettized, x = input_29_cast_fp16)[name = tensor("x_37_cast_fp16")]; + tensor layers_2_self_attn_q_proj_output_scale_to_fp16 = const()[name = tensor("layers_2_self_attn_q_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28427264)))]; + tensor query_5_cast_fp16 = mul(x = x_37_cast_fp16, y = layers_2_self_attn_q_proj_output_scale_to_fp16)[name = tensor("query_5_cast_fp16")]; + tensor var_527 = const()[name = tensor("op_527"), val = tensor([1, 1])]; + tensor var_529 = const()[name = tensor("op_529"), val = tensor([1, 1])]; + tensor x_39_pad_type_0 = const()[name = tensor("x_39_pad_type_0"), val = tensor("custom")]; + tensor x_39_pad_0 = const()[name = tensor("x_39_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_2_self_attn_k_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28429888))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29249152))), name = tensor("layers_2_self_attn_k_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_2_self_attn_k_proj_module_bias_to_fp16 = const()[name = tensor("layers_2_self_attn_k_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29249280)))]; + tensor x_39_cast_fp16 = conv(bias = layers_2_self_attn_k_proj_module_bias_to_fp16, dilations = var_529, groups = var_478, pad = x_39_pad_0, pad_type = x_39_pad_type_0, strides = var_527, weight = layers_2_self_attn_k_proj_module_weight_to_fp16_palettized, x = input_29_cast_fp16)[name = tensor("x_39_cast_fp16")]; + tensor layers_2_self_attn_k_proj_output_scale_to_fp16 = const()[name = tensor("layers_2_self_attn_k_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29251904)))]; + tensor key_5_cast_fp16 = mul(x = x_39_cast_fp16, y = layers_2_self_attn_k_proj_output_scale_to_fp16)[name = tensor("key_5_cast_fp16")]; + tensor var_539 = const()[name = tensor("op_539"), val = tensor([1, 1])]; + tensor var_541 = const()[name = tensor("op_541"), val = tensor([1, 1])]; + tensor x_41_pad_type_0 = const()[name = tensor("x_41_pad_type_0"), val = tensor("custom")]; + tensor x_41_pad_0 = const()[name = tensor("x_41_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_2_self_attn_v_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29254528))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30073792))), name = tensor("layers_2_self_attn_v_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_2_self_attn_v_proj_module_bias_to_fp16 = const()[name = tensor("layers_2_self_attn_v_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30073920)))]; + tensor x_41_cast_fp16 = conv(bias = layers_2_self_attn_v_proj_module_bias_to_fp16, dilations = var_541, groups = var_478, pad = x_41_pad_0, pad_type = x_41_pad_type_0, strides = var_539, weight = layers_2_self_attn_v_proj_module_weight_to_fp16_palettized, x = input_29_cast_fp16)[name = tensor("x_41_cast_fp16")]; + tensor layers_2_self_attn_v_proj_output_scale_to_fp16 = const()[name = tensor("layers_2_self_attn_v_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30076544)))]; + tensor value_5_cast_fp16 = mul(x = x_41_cast_fp16, y = layers_2_self_attn_v_proj_output_scale_to_fp16)[name = tensor("value_5_cast_fp16")]; + tensor var_546 = const()[name = tensor("op_546"), val = tensor([1, 20, 64, -1])]; + tensor var_547_cast_fp16 = reshape(shape = var_546, x = query_5_cast_fp16)[name = tensor("op_547_cast_fp16")]; + tensor var_548_to_fp16 = const()[name = tensor("op_548_to_fp16"), val = tensor(0x1p-3)]; + tensor var_549_cast_fp16 = mul(x = var_547_cast_fp16, y = var_548_to_fp16)[name = tensor("op_549_cast_fp16")]; + tensor var_550 = const()[name = tensor("op_550"), val = tensor([1, 20, 64, -1])]; + tensor var_551_cast_fp16 = reshape(shape = var_550, x = key_5_cast_fp16)[name = tensor("op_551_cast_fp16")]; + tensor mh_w_5_transpose_x_0 = const()[name = tensor("mh_w_5_transpose_x_0"), val = tensor(true)]; + tensor mh_w_5_transpose_y_0 = const()[name = tensor("mh_w_5_transpose_y_0"), val = tensor(false)]; + tensor mh_w_5_cast_fp16 = matmul(transpose_x = mh_w_5_transpose_x_0, transpose_y = mh_w_5_transpose_y_0, x = var_549_cast_fp16, y = var_551_cast_fp16)[name = tensor("mh_w_5_cast_fp16")]; + tensor var_554_cast_fp16 = softmax(axis = var_476, x = mh_w_5_cast_fp16)[name = tensor("op_554_cast_fp16")]; + tensor var_555 = const()[name = tensor("op_555"), val = tensor([1, 20, 64, -1])]; + tensor var_556_cast_fp16 = reshape(shape = var_555, x = value_5_cast_fp16)[name = tensor("op_556_cast_fp16")]; + tensor attn_5_transpose_x_0 = const()[name = tensor("attn_5_transpose_x_0"), val = tensor(false)]; + tensor attn_5_transpose_y_0 = const()[name = tensor("attn_5_transpose_y_0"), val = tensor(true)]; + tensor attn_5_cast_fp16 = matmul(transpose_x = attn_5_transpose_x_0, transpose_y = attn_5_transpose_y_0, x = var_556_cast_fp16, y = var_554_cast_fp16)[name = tensor("attn_5_cast_fp16")]; + tensor var_559 = const()[name = tensor("op_559"), val = tensor([1, 1280, 1, -1])]; + tensor x_43_cast_fp16 = reshape(shape = var_559, x = attn_5_cast_fp16)[name = tensor("x_43_cast_fp16")]; + tensor layers_2_self_attn_o_proj_input_shift_to_fp16 = const()[name = tensor("layers_2_self_attn_o_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30079168)))]; + tensor input_35_cast_fp16 = sub(x = x_43_cast_fp16, y = layers_2_self_attn_o_proj_input_shift_to_fp16)[name = tensor("input_35_cast_fp16")]; + tensor var_567 = const()[name = tensor("op_567"), val = tensor([1, 1])]; + tensor var_569 = const()[name = tensor("op_569"), val = tensor([1, 1])]; + tensor x_45_pad_type_0 = const()[name = tensor("x_45_pad_type_0"), val = tensor("custom")]; + tensor x_45_pad_0 = const()[name = tensor("x_45_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_2_self_attn_o_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30081792))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30901056))), name = tensor("layers_2_self_attn_o_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_2_self_attn_o_proj_module_bias_to_fp16 = const()[name = tensor("layers_2_self_attn_o_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30901184)))]; + tensor x_45_cast_fp16 = conv(bias = layers_2_self_attn_o_proj_module_bias_to_fp16, dilations = var_569, groups = var_478, pad = x_45_pad_0, pad_type = x_45_pad_type_0, strides = var_567, weight = layers_2_self_attn_o_proj_module_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = tensor("x_45_cast_fp16")]; + tensor layers_2_self_attn_o_proj_output_scale_to_fp16 = const()[name = tensor("layers_2_self_attn_o_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30903808)))]; + tensor obj_11_cast_fp16 = mul(x = x_45_cast_fp16, y = layers_2_self_attn_o_proj_output_scale_to_fp16)[name = tensor("obj_11_cast_fp16")]; + tensor inputs_11_cast_fp16 = add(x = inputs_9_cast_fp16, y = obj_11_cast_fp16)[name = tensor("inputs_11_cast_fp16")]; + tensor var_576 = const()[name = tensor("op_576"), val = tensor([1])]; + tensor channels_mean_11_cast_fp16 = reduce_mean(axes = var_576, keep_dims = var_479, x = inputs_11_cast_fp16)[name = tensor("channels_mean_11_cast_fp16")]; + tensor zero_mean_11_cast_fp16 = sub(x = inputs_11_cast_fp16, y = channels_mean_11_cast_fp16)[name = tensor("zero_mean_11_cast_fp16")]; + tensor zero_mean_sq_11_cast_fp16 = mul(x = zero_mean_11_cast_fp16, y = zero_mean_11_cast_fp16)[name = tensor("zero_mean_sq_11_cast_fp16")]; + tensor var_580 = const()[name = tensor("op_580"), val = tensor([1])]; + tensor var_581_cast_fp16 = reduce_mean(axes = var_580, keep_dims = var_479, x = zero_mean_sq_11_cast_fp16)[name = tensor("op_581_cast_fp16")]; + tensor var_582_to_fp16 = const()[name = tensor("op_582_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_583_cast_fp16 = add(x = var_581_cast_fp16, y = var_582_to_fp16)[name = tensor("op_583_cast_fp16")]; + tensor denom_11_epsilon_0_to_fp16 = const()[name = tensor("denom_11_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_11_cast_fp16 = rsqrt(epsilon = denom_11_epsilon_0_to_fp16, x = var_583_cast_fp16)[name = tensor("denom_11_cast_fp16")]; + tensor out_11_cast_fp16 = mul(x = zero_mean_11_cast_fp16, y = denom_11_cast_fp16)[name = tensor("out_11_cast_fp16")]; + tensor x_47_gamma_0_to_fp16 = const()[name = tensor("x_47_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30906432)))]; + tensor x_47_beta_0_to_fp16 = const()[name = tensor("x_47_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30909056)))]; + tensor x_47_epsilon_0_to_fp16 = const()[name = tensor("x_47_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_47_cast_fp16 = batch_norm(beta = x_47_beta_0_to_fp16, epsilon = x_47_epsilon_0_to_fp16, gamma = x_47_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_11_cast_fp16)[name = tensor("x_47_cast_fp16")]; + tensor layers_2_fc1_input_shift_to_fp16 = const()[name = tensor("layers_2_fc1_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30911680)))]; + tensor input_37_cast_fp16 = sub(x = x_47_cast_fp16, y = layers_2_fc1_input_shift_to_fp16)[name = tensor("input_37_cast_fp16")]; + tensor var_598 = const()[name = tensor("op_598"), val = tensor([1, 1])]; + tensor var_600 = const()[name = tensor("op_600"), val = tensor([1, 1])]; + tensor x_49_pad_type_0 = const()[name = tensor("x_49_pad_type_0"), val = tensor("custom")]; + tensor x_49_pad_0 = const()[name = tensor("x_49_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_2_fc1_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30914304))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(34191168))), name = tensor("layers_2_fc1_module_weight_to_fp16_palettized"), shape = tensor([5120, 1280, 1, 1])]; + tensor layers_2_fc1_module_bias_to_fp16 = const()[name = tensor("layers_2_fc1_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(34191296)))]; + tensor x_49_cast_fp16 = conv(bias = layers_2_fc1_module_bias_to_fp16, dilations = var_600, groups = var_478, pad = x_49_pad_0, pad_type = x_49_pad_type_0, strides = var_598, weight = layers_2_fc1_module_weight_to_fp16_palettized, x = input_37_cast_fp16)[name = tensor("x_49_cast_fp16")]; + tensor layers_2_fc1_output_scale_to_fp16 = const()[name = tensor("layers_2_fc1_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(34201600)))]; + tensor input_39_cast_fp16 = mul(x = x_49_cast_fp16, y = layers_2_fc1_output_scale_to_fp16)[name = tensor("input_39_cast_fp16")]; + tensor x_51_mode_0 = const()[name = tensor("x_51_mode_0"), val = tensor("EXACT")]; + tensor x_51_cast_fp16 = gelu(mode = x_51_mode_0, x = input_39_cast_fp16)[name = tensor("x_51_cast_fp16")]; + tensor layers_2_fc2_input_shift_to_fp16 = const()[name = tensor("layers_2_fc2_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(34211904)))]; + tensor input_41_cast_fp16 = sub(x = x_51_cast_fp16, y = layers_2_fc2_input_shift_to_fp16)[name = tensor("input_41_cast_fp16")]; + tensor var_611 = const()[name = tensor("op_611"), val = tensor([1, 1])]; + tensor var_613 = const()[name = tensor("op_613"), val = tensor([1, 1])]; + tensor x_53_pad_type_0 = const()[name = tensor("x_53_pad_type_0"), val = tensor("custom")]; + tensor x_53_pad_0 = const()[name = tensor("x_53_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_2_fc2_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(34222208))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37499072))), name = tensor("layers_2_fc2_module_weight_to_fp16_palettized"), shape = tensor([1280, 5120, 1, 1])]; + tensor layers_2_fc2_module_bias_to_fp16 = const()[name = tensor("layers_2_fc2_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37499200)))]; + tensor x_53_cast_fp16 = conv(bias = layers_2_fc2_module_bias_to_fp16, dilations = var_613, groups = var_478, pad = x_53_pad_0, pad_type = x_53_pad_type_0, strides = var_611, weight = layers_2_fc2_module_weight_to_fp16_palettized, x = input_41_cast_fp16)[name = tensor("x_53_cast_fp16")]; + tensor layers_2_fc2_output_scale_to_fp16 = const()[name = tensor("layers_2_fc2_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37501824)))]; + tensor hidden_states_9_cast_fp16 = mul(x = x_53_cast_fp16, y = layers_2_fc2_output_scale_to_fp16)[name = tensor("hidden_states_9_cast_fp16")]; + tensor inputs_13_cast_fp16 = add(x = inputs_11_cast_fp16, y = hidden_states_9_cast_fp16)[name = tensor("inputs_13_cast_fp16")]; + tensor var_625 = const()[name = tensor("op_625"), val = tensor(3)]; + tensor var_627 = const()[name = tensor("op_627"), val = tensor(1)]; + tensor var_628 = const()[name = tensor("op_628"), val = tensor(true)]; + tensor var_638 = const()[name = tensor("op_638"), val = tensor([1])]; + tensor channels_mean_13_cast_fp16 = reduce_mean(axes = var_638, keep_dims = var_628, x = inputs_13_cast_fp16)[name = tensor("channels_mean_13_cast_fp16")]; + tensor zero_mean_13_cast_fp16 = sub(x = inputs_13_cast_fp16, y = channels_mean_13_cast_fp16)[name = tensor("zero_mean_13_cast_fp16")]; + tensor zero_mean_sq_13_cast_fp16 = mul(x = zero_mean_13_cast_fp16, y = zero_mean_13_cast_fp16)[name = tensor("zero_mean_sq_13_cast_fp16")]; + tensor var_642 = const()[name = tensor("op_642"), val = tensor([1])]; + tensor var_643_cast_fp16 = reduce_mean(axes = var_642, keep_dims = var_628, x = zero_mean_sq_13_cast_fp16)[name = tensor("op_643_cast_fp16")]; + tensor var_644_to_fp16 = const()[name = tensor("op_644_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_645_cast_fp16 = add(x = var_643_cast_fp16, y = var_644_to_fp16)[name = tensor("op_645_cast_fp16")]; + tensor denom_13_epsilon_0_to_fp16 = const()[name = tensor("denom_13_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_13_cast_fp16 = rsqrt(epsilon = denom_13_epsilon_0_to_fp16, x = var_645_cast_fp16)[name = tensor("denom_13_cast_fp16")]; + tensor out_13_cast_fp16 = mul(x = zero_mean_13_cast_fp16, y = denom_13_cast_fp16)[name = tensor("out_13_cast_fp16")]; + tensor obj_13_gamma_0_to_fp16 = const()[name = tensor("obj_13_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37504448)))]; + tensor obj_13_beta_0_to_fp16 = const()[name = tensor("obj_13_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37507072)))]; + tensor obj_13_epsilon_0_to_fp16 = const()[name = tensor("obj_13_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_13_cast_fp16 = batch_norm(beta = obj_13_beta_0_to_fp16, epsilon = obj_13_epsilon_0_to_fp16, gamma = obj_13_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_13_cast_fp16)[name = tensor("obj_13_cast_fp16")]; + tensor layers_3_self_attn_q_proj_input_shift_to_fp16 = const()[name = tensor("layers_3_self_attn_q_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37509696)))]; + tensor input_43_cast_fp16 = sub(x = obj_13_cast_fp16, y = layers_3_self_attn_q_proj_input_shift_to_fp16)[name = tensor("input_43_cast_fp16")]; + tensor var_664 = const()[name = tensor("op_664"), val = tensor([1, 1])]; + tensor var_666 = const()[name = tensor("op_666"), val = tensor([1, 1])]; + tensor x_55_pad_type_0 = const()[name = tensor("x_55_pad_type_0"), val = tensor("custom")]; + tensor x_55_pad_0 = const()[name = tensor("x_55_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_3_self_attn_q_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37512320))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38331584))), name = tensor("layers_3_self_attn_q_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_3_self_attn_q_proj_module_bias_to_fp16 = const()[name = tensor("layers_3_self_attn_q_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38331712)))]; + tensor x_55_cast_fp16 = conv(bias = layers_3_self_attn_q_proj_module_bias_to_fp16, dilations = var_666, groups = var_627, pad = x_55_pad_0, pad_type = x_55_pad_type_0, strides = var_664, weight = layers_3_self_attn_q_proj_module_weight_to_fp16_palettized, x = input_43_cast_fp16)[name = tensor("x_55_cast_fp16")]; + tensor layers_3_self_attn_q_proj_output_scale_to_fp16 = const()[name = tensor("layers_3_self_attn_q_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38334336)))]; + tensor query_7_cast_fp16 = mul(x = x_55_cast_fp16, y = layers_3_self_attn_q_proj_output_scale_to_fp16)[name = tensor("query_7_cast_fp16")]; + tensor var_676 = const()[name = tensor("op_676"), val = tensor([1, 1])]; + tensor var_678 = const()[name = tensor("op_678"), val = tensor([1, 1])]; + tensor x_57_pad_type_0 = const()[name = tensor("x_57_pad_type_0"), val = tensor("custom")]; + tensor x_57_pad_0 = const()[name = tensor("x_57_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_3_self_attn_k_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38336960))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39156224))), name = tensor("layers_3_self_attn_k_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_3_self_attn_k_proj_module_bias_to_fp16 = const()[name = tensor("layers_3_self_attn_k_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39156352)))]; + tensor x_57_cast_fp16 = conv(bias = layers_3_self_attn_k_proj_module_bias_to_fp16, dilations = var_678, groups = var_627, pad = x_57_pad_0, pad_type = x_57_pad_type_0, strides = var_676, weight = layers_3_self_attn_k_proj_module_weight_to_fp16_palettized, x = input_43_cast_fp16)[name = tensor("x_57_cast_fp16")]; + tensor layers_3_self_attn_k_proj_output_scale_to_fp16 = const()[name = tensor("layers_3_self_attn_k_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39158976)))]; + tensor key_7_cast_fp16 = mul(x = x_57_cast_fp16, y = layers_3_self_attn_k_proj_output_scale_to_fp16)[name = tensor("key_7_cast_fp16")]; + tensor var_688 = const()[name = tensor("op_688"), val = tensor([1, 1])]; + tensor var_690 = const()[name = tensor("op_690"), val = tensor([1, 1])]; + tensor x_59_pad_type_0 = const()[name = tensor("x_59_pad_type_0"), val = tensor("custom")]; + tensor x_59_pad_0 = const()[name = tensor("x_59_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_3_self_attn_v_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39161600))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39980864))), name = tensor("layers_3_self_attn_v_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_3_self_attn_v_proj_module_bias_to_fp16 = const()[name = tensor("layers_3_self_attn_v_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39980992)))]; + tensor x_59_cast_fp16 = conv(bias = layers_3_self_attn_v_proj_module_bias_to_fp16, dilations = var_690, groups = var_627, pad = x_59_pad_0, pad_type = x_59_pad_type_0, strides = var_688, weight = layers_3_self_attn_v_proj_module_weight_to_fp16_palettized, x = input_43_cast_fp16)[name = tensor("x_59_cast_fp16")]; + tensor layers_3_self_attn_v_proj_output_scale_to_fp16 = const()[name = tensor("layers_3_self_attn_v_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39983616)))]; + tensor value_7_cast_fp16 = mul(x = x_59_cast_fp16, y = layers_3_self_attn_v_proj_output_scale_to_fp16)[name = tensor("value_7_cast_fp16")]; + tensor var_695 = const()[name = tensor("op_695"), val = tensor([1, 20, 64, -1])]; + tensor var_696_cast_fp16 = reshape(shape = var_695, x = query_7_cast_fp16)[name = tensor("op_696_cast_fp16")]; + tensor var_697_to_fp16 = const()[name = tensor("op_697_to_fp16"), val = tensor(0x1p-3)]; + tensor var_698_cast_fp16 = mul(x = var_696_cast_fp16, y = var_697_to_fp16)[name = tensor("op_698_cast_fp16")]; + tensor var_699 = const()[name = tensor("op_699"), val = tensor([1, 20, 64, -1])]; + tensor var_700_cast_fp16 = reshape(shape = var_699, x = key_7_cast_fp16)[name = tensor("op_700_cast_fp16")]; + tensor mh_w_7_transpose_x_0 = const()[name = tensor("mh_w_7_transpose_x_0"), val = tensor(true)]; + tensor mh_w_7_transpose_y_0 = const()[name = tensor("mh_w_7_transpose_y_0"), val = tensor(false)]; + tensor mh_w_7_cast_fp16 = matmul(transpose_x = mh_w_7_transpose_x_0, transpose_y = mh_w_7_transpose_y_0, x = var_698_cast_fp16, y = var_700_cast_fp16)[name = tensor("mh_w_7_cast_fp16")]; + tensor var_703_cast_fp16 = softmax(axis = var_625, x = mh_w_7_cast_fp16)[name = tensor("op_703_cast_fp16")]; + tensor var_704 = const()[name = tensor("op_704"), val = tensor([1, 20, 64, -1])]; + tensor var_705_cast_fp16 = reshape(shape = var_704, x = value_7_cast_fp16)[name = tensor("op_705_cast_fp16")]; + tensor attn_7_transpose_x_0 = const()[name = tensor("attn_7_transpose_x_0"), val = tensor(false)]; + tensor attn_7_transpose_y_0 = const()[name = tensor("attn_7_transpose_y_0"), val = tensor(true)]; + tensor attn_7_cast_fp16 = matmul(transpose_x = attn_7_transpose_x_0, transpose_y = attn_7_transpose_y_0, x = var_705_cast_fp16, y = var_703_cast_fp16)[name = tensor("attn_7_cast_fp16")]; + tensor var_708 = const()[name = tensor("op_708"), val = tensor([1, 1280, 1, -1])]; + tensor x_61_cast_fp16 = reshape(shape = var_708, x = attn_7_cast_fp16)[name = tensor("x_61_cast_fp16")]; + tensor layers_3_self_attn_o_proj_input_shift_to_fp16 = const()[name = tensor("layers_3_self_attn_o_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39986240)))]; + tensor input_49_cast_fp16 = sub(x = x_61_cast_fp16, y = layers_3_self_attn_o_proj_input_shift_to_fp16)[name = tensor("input_49_cast_fp16")]; + tensor var_716 = const()[name = tensor("op_716"), val = tensor([1, 1])]; + tensor var_718 = const()[name = tensor("op_718"), val = tensor([1, 1])]; + tensor x_63_pad_type_0 = const()[name = tensor("x_63_pad_type_0"), val = tensor("custom")]; + tensor x_63_pad_0 = const()[name = tensor("x_63_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_3_self_attn_o_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39988864))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40808128))), name = tensor("layers_3_self_attn_o_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_3_self_attn_o_proj_module_bias_to_fp16 = const()[name = tensor("layers_3_self_attn_o_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40808256)))]; + tensor x_63_cast_fp16 = conv(bias = layers_3_self_attn_o_proj_module_bias_to_fp16, dilations = var_718, groups = var_627, pad = x_63_pad_0, pad_type = x_63_pad_type_0, strides = var_716, weight = layers_3_self_attn_o_proj_module_weight_to_fp16_palettized, x = input_49_cast_fp16)[name = tensor("x_63_cast_fp16")]; + tensor layers_3_self_attn_o_proj_output_scale_to_fp16 = const()[name = tensor("layers_3_self_attn_o_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40810880)))]; + tensor obj_15_cast_fp16 = mul(x = x_63_cast_fp16, y = layers_3_self_attn_o_proj_output_scale_to_fp16)[name = tensor("obj_15_cast_fp16")]; + tensor inputs_15_cast_fp16 = add(x = inputs_13_cast_fp16, y = obj_15_cast_fp16)[name = tensor("inputs_15_cast_fp16")]; + tensor var_725 = const()[name = tensor("op_725"), val = tensor([1])]; + tensor channels_mean_15_cast_fp16 = reduce_mean(axes = var_725, keep_dims = var_628, x = inputs_15_cast_fp16)[name = tensor("channels_mean_15_cast_fp16")]; + tensor zero_mean_15_cast_fp16 = sub(x = inputs_15_cast_fp16, y = channels_mean_15_cast_fp16)[name = tensor("zero_mean_15_cast_fp16")]; + tensor zero_mean_sq_15_cast_fp16 = mul(x = zero_mean_15_cast_fp16, y = zero_mean_15_cast_fp16)[name = tensor("zero_mean_sq_15_cast_fp16")]; + tensor var_729 = const()[name = tensor("op_729"), val = tensor([1])]; + tensor var_730_cast_fp16 = reduce_mean(axes = var_729, keep_dims = var_628, x = zero_mean_sq_15_cast_fp16)[name = tensor("op_730_cast_fp16")]; + tensor var_731_to_fp16 = const()[name = tensor("op_731_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_732_cast_fp16 = add(x = var_730_cast_fp16, y = var_731_to_fp16)[name = tensor("op_732_cast_fp16")]; + tensor denom_15_epsilon_0_to_fp16 = const()[name = tensor("denom_15_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_15_cast_fp16 = rsqrt(epsilon = denom_15_epsilon_0_to_fp16, x = var_732_cast_fp16)[name = tensor("denom_15_cast_fp16")]; + tensor out_15_cast_fp16 = mul(x = zero_mean_15_cast_fp16, y = denom_15_cast_fp16)[name = tensor("out_15_cast_fp16")]; + tensor x_65_gamma_0_to_fp16 = const()[name = tensor("x_65_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40813504)))]; + tensor x_65_beta_0_to_fp16 = const()[name = tensor("x_65_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40816128)))]; + tensor x_65_epsilon_0_to_fp16 = const()[name = tensor("x_65_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_65_cast_fp16 = batch_norm(beta = x_65_beta_0_to_fp16, epsilon = x_65_epsilon_0_to_fp16, gamma = x_65_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_15_cast_fp16)[name = tensor("x_65_cast_fp16")]; + tensor layers_3_fc1_input_shift_to_fp16 = const()[name = tensor("layers_3_fc1_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40818752)))]; + tensor input_51_cast_fp16 = sub(x = x_65_cast_fp16, y = layers_3_fc1_input_shift_to_fp16)[name = tensor("input_51_cast_fp16")]; + tensor var_747 = const()[name = tensor("op_747"), val = tensor([1, 1])]; + tensor var_749 = const()[name = tensor("op_749"), val = tensor([1, 1])]; + tensor x_67_pad_type_0 = const()[name = tensor("x_67_pad_type_0"), val = tensor("custom")]; + tensor x_67_pad_0 = const()[name = tensor("x_67_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_3_fc1_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40821376))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44098240))), name = tensor("layers_3_fc1_module_weight_to_fp16_palettized"), shape = tensor([5120, 1280, 1, 1])]; + tensor layers_3_fc1_module_bias_to_fp16 = const()[name = tensor("layers_3_fc1_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44098368)))]; + tensor x_67_cast_fp16 = conv(bias = layers_3_fc1_module_bias_to_fp16, dilations = var_749, groups = var_627, pad = x_67_pad_0, pad_type = x_67_pad_type_0, strides = var_747, weight = layers_3_fc1_module_weight_to_fp16_palettized, x = input_51_cast_fp16)[name = tensor("x_67_cast_fp16")]; + tensor layers_3_fc1_output_scale_to_fp16 = const()[name = tensor("layers_3_fc1_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44108672)))]; + tensor input_53_cast_fp16 = mul(x = x_67_cast_fp16, y = layers_3_fc1_output_scale_to_fp16)[name = tensor("input_53_cast_fp16")]; + tensor x_69_mode_0 = const()[name = tensor("x_69_mode_0"), val = tensor("EXACT")]; + tensor x_69_cast_fp16 = gelu(mode = x_69_mode_0, x = input_53_cast_fp16)[name = tensor("x_69_cast_fp16")]; + tensor layers_3_fc2_input_shift_to_fp16 = const()[name = tensor("layers_3_fc2_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44118976)))]; + tensor input_55_cast_fp16 = sub(x = x_69_cast_fp16, y = layers_3_fc2_input_shift_to_fp16)[name = tensor("input_55_cast_fp16")]; + tensor var_760 = const()[name = tensor("op_760"), val = tensor([1, 1])]; + tensor var_762 = const()[name = tensor("op_762"), val = tensor([1, 1])]; + tensor x_71_pad_type_0 = const()[name = tensor("x_71_pad_type_0"), val = tensor("custom")]; + tensor x_71_pad_0 = const()[name = tensor("x_71_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_3_fc2_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44129280))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(47406144))), name = tensor("layers_3_fc2_module_weight_to_fp16_palettized"), shape = tensor([1280, 5120, 1, 1])]; + tensor layers_3_fc2_module_bias_to_fp16 = const()[name = tensor("layers_3_fc2_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(47406272)))]; + tensor x_71_cast_fp16 = conv(bias = layers_3_fc2_module_bias_to_fp16, dilations = var_762, groups = var_627, pad = x_71_pad_0, pad_type = x_71_pad_type_0, strides = var_760, weight = layers_3_fc2_module_weight_to_fp16_palettized, x = input_55_cast_fp16)[name = tensor("x_71_cast_fp16")]; + tensor layers_3_fc2_output_scale_to_fp16 = const()[name = tensor("layers_3_fc2_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(47408896)))]; + tensor hidden_states_11_cast_fp16 = mul(x = x_71_cast_fp16, y = layers_3_fc2_output_scale_to_fp16)[name = tensor("hidden_states_11_cast_fp16")]; + tensor inputs_17_cast_fp16 = add(x = inputs_15_cast_fp16, y = hidden_states_11_cast_fp16)[name = tensor("inputs_17_cast_fp16")]; + tensor var_774 = const()[name = tensor("op_774"), val = tensor(3)]; + tensor var_776 = const()[name = tensor("op_776"), val = tensor(1)]; + tensor var_777 = const()[name = tensor("op_777"), val = tensor(true)]; + tensor var_787 = const()[name = tensor("op_787"), val = tensor([1])]; + tensor channels_mean_17_cast_fp16 = reduce_mean(axes = var_787, keep_dims = var_777, x = inputs_17_cast_fp16)[name = tensor("channels_mean_17_cast_fp16")]; + tensor zero_mean_17_cast_fp16 = sub(x = inputs_17_cast_fp16, y = channels_mean_17_cast_fp16)[name = tensor("zero_mean_17_cast_fp16")]; + tensor zero_mean_sq_17_cast_fp16 = mul(x = zero_mean_17_cast_fp16, y = zero_mean_17_cast_fp16)[name = tensor("zero_mean_sq_17_cast_fp16")]; + tensor var_791 = const()[name = tensor("op_791"), val = tensor([1])]; + tensor var_792_cast_fp16 = reduce_mean(axes = var_791, keep_dims = var_777, x = zero_mean_sq_17_cast_fp16)[name = tensor("op_792_cast_fp16")]; + tensor var_793_to_fp16 = const()[name = tensor("op_793_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_794_cast_fp16 = add(x = var_792_cast_fp16, y = var_793_to_fp16)[name = tensor("op_794_cast_fp16")]; + tensor denom_17_epsilon_0_to_fp16 = const()[name = tensor("denom_17_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_17_cast_fp16 = rsqrt(epsilon = denom_17_epsilon_0_to_fp16, x = var_794_cast_fp16)[name = tensor("denom_17_cast_fp16")]; + tensor out_17_cast_fp16 = mul(x = zero_mean_17_cast_fp16, y = denom_17_cast_fp16)[name = tensor("out_17_cast_fp16")]; + tensor obj_17_gamma_0_to_fp16 = const()[name = tensor("obj_17_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(47411520)))]; + tensor obj_17_beta_0_to_fp16 = const()[name = tensor("obj_17_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(47414144)))]; + tensor obj_17_epsilon_0_to_fp16 = const()[name = tensor("obj_17_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_17_cast_fp16 = batch_norm(beta = obj_17_beta_0_to_fp16, epsilon = obj_17_epsilon_0_to_fp16, gamma = obj_17_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_17_cast_fp16)[name = tensor("obj_17_cast_fp16")]; + tensor layers_4_self_attn_q_proj_input_shift_to_fp16 = const()[name = tensor("layers_4_self_attn_q_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(47416768)))]; + tensor input_57_cast_fp16 = sub(x = obj_17_cast_fp16, y = layers_4_self_attn_q_proj_input_shift_to_fp16)[name = tensor("input_57_cast_fp16")]; + tensor var_813 = const()[name = tensor("op_813"), val = tensor([1, 1])]; + tensor var_815 = const()[name = tensor("op_815"), val = tensor([1, 1])]; + tensor x_73_pad_type_0 = const()[name = tensor("x_73_pad_type_0"), val = tensor("custom")]; + tensor x_73_pad_0 = const()[name = tensor("x_73_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_4_self_attn_q_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(47419392))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(48238656))), name = tensor("layers_4_self_attn_q_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_4_self_attn_q_proj_module_bias_to_fp16 = const()[name = tensor("layers_4_self_attn_q_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(48238784)))]; + tensor x_73_cast_fp16 = conv(bias = layers_4_self_attn_q_proj_module_bias_to_fp16, dilations = var_815, groups = var_776, pad = x_73_pad_0, pad_type = x_73_pad_type_0, strides = var_813, weight = layers_4_self_attn_q_proj_module_weight_to_fp16_palettized, x = input_57_cast_fp16)[name = tensor("x_73_cast_fp16")]; + tensor layers_4_self_attn_q_proj_output_scale_to_fp16 = const()[name = tensor("layers_4_self_attn_q_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(48241408)))]; + tensor query_9_cast_fp16 = mul(x = x_73_cast_fp16, y = layers_4_self_attn_q_proj_output_scale_to_fp16)[name = tensor("query_9_cast_fp16")]; + tensor var_825 = const()[name = tensor("op_825"), val = tensor([1, 1])]; + tensor var_827 = const()[name = tensor("op_827"), val = tensor([1, 1])]; + tensor x_75_pad_type_0 = const()[name = tensor("x_75_pad_type_0"), val = tensor("custom")]; + tensor x_75_pad_0 = const()[name = tensor("x_75_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_4_self_attn_k_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(48244032))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(49063296))), name = tensor("layers_4_self_attn_k_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_4_self_attn_k_proj_module_bias_to_fp16 = const()[name = tensor("layers_4_self_attn_k_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(49063424)))]; + tensor x_75_cast_fp16 = conv(bias = layers_4_self_attn_k_proj_module_bias_to_fp16, dilations = var_827, groups = var_776, pad = x_75_pad_0, pad_type = x_75_pad_type_0, strides = var_825, weight = layers_4_self_attn_k_proj_module_weight_to_fp16_palettized, x = input_57_cast_fp16)[name = tensor("x_75_cast_fp16")]; + tensor layers_4_self_attn_k_proj_output_scale_to_fp16 = const()[name = tensor("layers_4_self_attn_k_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(49066048)))]; + tensor key_9_cast_fp16 = mul(x = x_75_cast_fp16, y = layers_4_self_attn_k_proj_output_scale_to_fp16)[name = tensor("key_9_cast_fp16")]; + tensor var_837 = const()[name = tensor("op_837"), val = tensor([1, 1])]; + tensor var_839 = const()[name = tensor("op_839"), val = tensor([1, 1])]; + tensor x_77_pad_type_0 = const()[name = tensor("x_77_pad_type_0"), val = tensor("custom")]; + tensor x_77_pad_0 = const()[name = tensor("x_77_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_4_self_attn_v_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(49068672))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(49887936))), name = tensor("layers_4_self_attn_v_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_4_self_attn_v_proj_module_bias_to_fp16 = const()[name = tensor("layers_4_self_attn_v_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(49888064)))]; + tensor x_77_cast_fp16 = conv(bias = layers_4_self_attn_v_proj_module_bias_to_fp16, dilations = var_839, groups = var_776, pad = x_77_pad_0, pad_type = x_77_pad_type_0, strides = var_837, weight = layers_4_self_attn_v_proj_module_weight_to_fp16_palettized, x = input_57_cast_fp16)[name = tensor("x_77_cast_fp16")]; + tensor layers_4_self_attn_v_proj_output_scale_to_fp16 = const()[name = tensor("layers_4_self_attn_v_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(49890688)))]; + tensor value_9_cast_fp16 = mul(x = x_77_cast_fp16, y = layers_4_self_attn_v_proj_output_scale_to_fp16)[name = tensor("value_9_cast_fp16")]; + tensor var_844 = const()[name = tensor("op_844"), val = tensor([1, 20, 64, -1])]; + tensor var_845_cast_fp16 = reshape(shape = var_844, x = query_9_cast_fp16)[name = tensor("op_845_cast_fp16")]; + tensor var_846_to_fp16 = const()[name = tensor("op_846_to_fp16"), val = tensor(0x1p-3)]; + tensor var_847_cast_fp16 = mul(x = var_845_cast_fp16, y = var_846_to_fp16)[name = tensor("op_847_cast_fp16")]; + tensor var_848 = const()[name = tensor("op_848"), val = tensor([1, 20, 64, -1])]; + tensor var_849_cast_fp16 = reshape(shape = var_848, x = key_9_cast_fp16)[name = tensor("op_849_cast_fp16")]; + tensor mh_w_9_transpose_x_0 = const()[name = tensor("mh_w_9_transpose_x_0"), val = tensor(true)]; + tensor mh_w_9_transpose_y_0 = const()[name = tensor("mh_w_9_transpose_y_0"), val = tensor(false)]; + tensor mh_w_9_cast_fp16 = matmul(transpose_x = mh_w_9_transpose_x_0, transpose_y = mh_w_9_transpose_y_0, x = var_847_cast_fp16, y = var_849_cast_fp16)[name = tensor("mh_w_9_cast_fp16")]; + tensor var_852_cast_fp16 = softmax(axis = var_774, x = mh_w_9_cast_fp16)[name = tensor("op_852_cast_fp16")]; + tensor var_853 = const()[name = tensor("op_853"), val = tensor([1, 20, 64, -1])]; + tensor var_854_cast_fp16 = reshape(shape = var_853, x = value_9_cast_fp16)[name = tensor("op_854_cast_fp16")]; + tensor attn_9_transpose_x_0 = const()[name = tensor("attn_9_transpose_x_0"), val = tensor(false)]; + tensor attn_9_transpose_y_0 = const()[name = tensor("attn_9_transpose_y_0"), val = tensor(true)]; + tensor attn_9_cast_fp16 = matmul(transpose_x = attn_9_transpose_x_0, transpose_y = attn_9_transpose_y_0, x = var_854_cast_fp16, y = var_852_cast_fp16)[name = tensor("attn_9_cast_fp16")]; + tensor var_857 = const()[name = tensor("op_857"), val = tensor([1, 1280, 1, -1])]; + tensor x_79_cast_fp16 = reshape(shape = var_857, x = attn_9_cast_fp16)[name = tensor("x_79_cast_fp16")]; + tensor layers_4_self_attn_o_proj_input_shift_to_fp16 = const()[name = tensor("layers_4_self_attn_o_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(49893312)))]; + tensor input_63_cast_fp16 = sub(x = x_79_cast_fp16, y = layers_4_self_attn_o_proj_input_shift_to_fp16)[name = tensor("input_63_cast_fp16")]; + tensor var_865 = const()[name = tensor("op_865"), val = tensor([1, 1])]; + tensor var_867 = const()[name = tensor("op_867"), val = tensor([1, 1])]; + tensor x_81_pad_type_0 = const()[name = tensor("x_81_pad_type_0"), val = tensor("custom")]; + tensor x_81_pad_0 = const()[name = tensor("x_81_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_4_self_attn_o_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(49895936))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(50715200))), name = tensor("layers_4_self_attn_o_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_4_self_attn_o_proj_module_bias_to_fp16 = const()[name = tensor("layers_4_self_attn_o_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(50715328)))]; + tensor x_81_cast_fp16 = conv(bias = layers_4_self_attn_o_proj_module_bias_to_fp16, dilations = var_867, groups = var_776, pad = x_81_pad_0, pad_type = x_81_pad_type_0, strides = var_865, weight = layers_4_self_attn_o_proj_module_weight_to_fp16_palettized, x = input_63_cast_fp16)[name = tensor("x_81_cast_fp16")]; + tensor layers_4_self_attn_o_proj_output_scale_to_fp16 = const()[name = tensor("layers_4_self_attn_o_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(50717952)))]; + tensor obj_19_cast_fp16 = mul(x = x_81_cast_fp16, y = layers_4_self_attn_o_proj_output_scale_to_fp16)[name = tensor("obj_19_cast_fp16")]; + tensor inputs_19_cast_fp16 = add(x = inputs_17_cast_fp16, y = obj_19_cast_fp16)[name = tensor("inputs_19_cast_fp16")]; + tensor var_874 = const()[name = tensor("op_874"), val = tensor([1])]; + tensor channels_mean_19_cast_fp16 = reduce_mean(axes = var_874, keep_dims = var_777, x = inputs_19_cast_fp16)[name = tensor("channels_mean_19_cast_fp16")]; + tensor zero_mean_19_cast_fp16 = sub(x = inputs_19_cast_fp16, y = channels_mean_19_cast_fp16)[name = tensor("zero_mean_19_cast_fp16")]; + tensor zero_mean_sq_19_cast_fp16 = mul(x = zero_mean_19_cast_fp16, y = zero_mean_19_cast_fp16)[name = tensor("zero_mean_sq_19_cast_fp16")]; + tensor var_878 = const()[name = tensor("op_878"), val = tensor([1])]; + tensor var_879_cast_fp16 = reduce_mean(axes = var_878, keep_dims = var_777, x = zero_mean_sq_19_cast_fp16)[name = tensor("op_879_cast_fp16")]; + tensor var_880_to_fp16 = const()[name = tensor("op_880_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_881_cast_fp16 = add(x = var_879_cast_fp16, y = var_880_to_fp16)[name = tensor("op_881_cast_fp16")]; + tensor denom_19_epsilon_0_to_fp16 = const()[name = tensor("denom_19_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_19_cast_fp16 = rsqrt(epsilon = denom_19_epsilon_0_to_fp16, x = var_881_cast_fp16)[name = tensor("denom_19_cast_fp16")]; + tensor out_19_cast_fp16 = mul(x = zero_mean_19_cast_fp16, y = denom_19_cast_fp16)[name = tensor("out_19_cast_fp16")]; + tensor x_83_gamma_0_to_fp16 = const()[name = tensor("x_83_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(50720576)))]; + tensor x_83_beta_0_to_fp16 = const()[name = tensor("x_83_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(50723200)))]; + tensor x_83_epsilon_0_to_fp16 = const()[name = tensor("x_83_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_83_cast_fp16 = batch_norm(beta = x_83_beta_0_to_fp16, epsilon = x_83_epsilon_0_to_fp16, gamma = x_83_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_19_cast_fp16)[name = tensor("x_83_cast_fp16")]; + tensor layers_4_fc1_input_shift_to_fp16 = const()[name = tensor("layers_4_fc1_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(50725824)))]; + tensor input_65_cast_fp16 = sub(x = x_83_cast_fp16, y = layers_4_fc1_input_shift_to_fp16)[name = tensor("input_65_cast_fp16")]; + tensor var_896 = const()[name = tensor("op_896"), val = tensor([1, 1])]; + tensor var_898 = const()[name = tensor("op_898"), val = tensor([1, 1])]; + tensor x_85_pad_type_0 = const()[name = tensor("x_85_pad_type_0"), val = tensor("custom")]; + tensor x_85_pad_0 = const()[name = tensor("x_85_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_4_fc1_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(50728448))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(54005312))), name = tensor("layers_4_fc1_module_weight_to_fp16_palettized"), shape = tensor([5120, 1280, 1, 1])]; + tensor layers_4_fc1_module_bias_to_fp16 = const()[name = tensor("layers_4_fc1_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(54005440)))]; + tensor x_85_cast_fp16 = conv(bias = layers_4_fc1_module_bias_to_fp16, dilations = var_898, groups = var_776, pad = x_85_pad_0, pad_type = x_85_pad_type_0, strides = var_896, weight = layers_4_fc1_module_weight_to_fp16_palettized, x = input_65_cast_fp16)[name = tensor("x_85_cast_fp16")]; + tensor layers_4_fc1_output_scale_to_fp16 = const()[name = tensor("layers_4_fc1_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(54015744)))]; + tensor input_67_cast_fp16 = mul(x = x_85_cast_fp16, y = layers_4_fc1_output_scale_to_fp16)[name = tensor("input_67_cast_fp16")]; + tensor x_87_mode_0 = const()[name = tensor("x_87_mode_0"), val = tensor("EXACT")]; + tensor x_87_cast_fp16 = gelu(mode = x_87_mode_0, x = input_67_cast_fp16)[name = tensor("x_87_cast_fp16")]; + tensor layers_4_fc2_input_shift_to_fp16 = const()[name = tensor("layers_4_fc2_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(54026048)))]; + tensor input_69_cast_fp16 = sub(x = x_87_cast_fp16, y = layers_4_fc2_input_shift_to_fp16)[name = tensor("input_69_cast_fp16")]; + tensor var_909 = const()[name = tensor("op_909"), val = tensor([1, 1])]; + tensor var_911 = const()[name = tensor("op_911"), val = tensor([1, 1])]; + tensor x_89_pad_type_0 = const()[name = tensor("x_89_pad_type_0"), val = tensor("custom")]; + tensor x_89_pad_0 = const()[name = tensor("x_89_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_4_fc2_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(54036352))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(57313216))), name = tensor("layers_4_fc2_module_weight_to_fp16_palettized"), shape = tensor([1280, 5120, 1, 1])]; + tensor layers_4_fc2_module_bias_to_fp16 = const()[name = tensor("layers_4_fc2_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(57313344)))]; + tensor x_89_cast_fp16 = conv(bias = layers_4_fc2_module_bias_to_fp16, dilations = var_911, groups = var_776, pad = x_89_pad_0, pad_type = x_89_pad_type_0, strides = var_909, weight = layers_4_fc2_module_weight_to_fp16_palettized, x = input_69_cast_fp16)[name = tensor("x_89_cast_fp16")]; + tensor layers_4_fc2_output_scale_to_fp16 = const()[name = tensor("layers_4_fc2_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(57315968)))]; + tensor hidden_states_13_cast_fp16 = mul(x = x_89_cast_fp16, y = layers_4_fc2_output_scale_to_fp16)[name = tensor("hidden_states_13_cast_fp16")]; + tensor inputs_21_cast_fp16 = add(x = inputs_19_cast_fp16, y = hidden_states_13_cast_fp16)[name = tensor("inputs_21_cast_fp16")]; + tensor var_923 = const()[name = tensor("op_923"), val = tensor(3)]; + tensor var_925 = const()[name = tensor("op_925"), val = tensor(1)]; + tensor var_926 = const()[name = tensor("op_926"), val = tensor(true)]; + tensor var_936 = const()[name = tensor("op_936"), val = tensor([1])]; + tensor channels_mean_21_cast_fp16 = reduce_mean(axes = var_936, keep_dims = var_926, x = inputs_21_cast_fp16)[name = tensor("channels_mean_21_cast_fp16")]; + tensor zero_mean_21_cast_fp16 = sub(x = inputs_21_cast_fp16, y = channels_mean_21_cast_fp16)[name = tensor("zero_mean_21_cast_fp16")]; + tensor zero_mean_sq_21_cast_fp16 = mul(x = zero_mean_21_cast_fp16, y = zero_mean_21_cast_fp16)[name = tensor("zero_mean_sq_21_cast_fp16")]; + tensor var_940 = const()[name = tensor("op_940"), val = tensor([1])]; + tensor var_941_cast_fp16 = reduce_mean(axes = var_940, keep_dims = var_926, x = zero_mean_sq_21_cast_fp16)[name = tensor("op_941_cast_fp16")]; + tensor var_942_to_fp16 = const()[name = tensor("op_942_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_943_cast_fp16 = add(x = var_941_cast_fp16, y = var_942_to_fp16)[name = tensor("op_943_cast_fp16")]; + tensor denom_21_epsilon_0_to_fp16 = const()[name = tensor("denom_21_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_21_cast_fp16 = rsqrt(epsilon = denom_21_epsilon_0_to_fp16, x = var_943_cast_fp16)[name = tensor("denom_21_cast_fp16")]; + tensor out_21_cast_fp16 = mul(x = zero_mean_21_cast_fp16, y = denom_21_cast_fp16)[name = tensor("out_21_cast_fp16")]; + tensor obj_21_gamma_0_to_fp16 = const()[name = tensor("obj_21_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(57318592)))]; + tensor obj_21_beta_0_to_fp16 = const()[name = tensor("obj_21_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(57321216)))]; + tensor obj_21_epsilon_0_to_fp16 = const()[name = tensor("obj_21_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_21_cast_fp16 = batch_norm(beta = obj_21_beta_0_to_fp16, epsilon = obj_21_epsilon_0_to_fp16, gamma = obj_21_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_21_cast_fp16)[name = tensor("obj_21_cast_fp16")]; + tensor layers_5_self_attn_q_proj_input_shift_to_fp16 = const()[name = tensor("layers_5_self_attn_q_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(57323840)))]; + tensor input_71_cast_fp16 = sub(x = obj_21_cast_fp16, y = layers_5_self_attn_q_proj_input_shift_to_fp16)[name = tensor("input_71_cast_fp16")]; + tensor var_962 = const()[name = tensor("op_962"), val = tensor([1, 1])]; + tensor var_964 = const()[name = tensor("op_964"), val = tensor([1, 1])]; + tensor x_91_pad_type_0 = const()[name = tensor("x_91_pad_type_0"), val = tensor("custom")]; + tensor x_91_pad_0 = const()[name = tensor("x_91_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_5_self_attn_q_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(57326464))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(58145728))), name = tensor("layers_5_self_attn_q_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_5_self_attn_q_proj_module_bias_to_fp16 = const()[name = tensor("layers_5_self_attn_q_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(58145856)))]; + tensor x_91_cast_fp16 = conv(bias = layers_5_self_attn_q_proj_module_bias_to_fp16, dilations = var_964, groups = var_925, pad = x_91_pad_0, pad_type = x_91_pad_type_0, strides = var_962, weight = layers_5_self_attn_q_proj_module_weight_to_fp16_palettized, x = input_71_cast_fp16)[name = tensor("x_91_cast_fp16")]; + tensor layers_5_self_attn_q_proj_output_scale_to_fp16 = const()[name = tensor("layers_5_self_attn_q_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(58148480)))]; + tensor query_11_cast_fp16 = mul(x = x_91_cast_fp16, y = layers_5_self_attn_q_proj_output_scale_to_fp16)[name = tensor("query_11_cast_fp16")]; + tensor var_974 = const()[name = tensor("op_974"), val = tensor([1, 1])]; + tensor var_976 = const()[name = tensor("op_976"), val = tensor([1, 1])]; + tensor x_93_pad_type_0 = const()[name = tensor("x_93_pad_type_0"), val = tensor("custom")]; + tensor x_93_pad_0 = const()[name = tensor("x_93_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_5_self_attn_k_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(58151104))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(58970368))), name = tensor("layers_5_self_attn_k_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_5_self_attn_k_proj_module_bias_to_fp16 = const()[name = tensor("layers_5_self_attn_k_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(58970496)))]; + tensor x_93_cast_fp16 = conv(bias = layers_5_self_attn_k_proj_module_bias_to_fp16, dilations = var_976, groups = var_925, pad = x_93_pad_0, pad_type = x_93_pad_type_0, strides = var_974, weight = layers_5_self_attn_k_proj_module_weight_to_fp16_palettized, x = input_71_cast_fp16)[name = tensor("x_93_cast_fp16")]; + tensor layers_5_self_attn_k_proj_output_scale_to_fp16 = const()[name = tensor("layers_5_self_attn_k_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(58973120)))]; + tensor key_11_cast_fp16 = mul(x = x_93_cast_fp16, y = layers_5_self_attn_k_proj_output_scale_to_fp16)[name = tensor("key_11_cast_fp16")]; + tensor var_986 = const()[name = tensor("op_986"), val = tensor([1, 1])]; + tensor var_988 = const()[name = tensor("op_988"), val = tensor([1, 1])]; + tensor x_95_pad_type_0 = const()[name = tensor("x_95_pad_type_0"), val = tensor("custom")]; + tensor x_95_pad_0 = const()[name = tensor("x_95_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_5_self_attn_v_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(58975744))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(59795008))), name = tensor("layers_5_self_attn_v_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_5_self_attn_v_proj_module_bias_to_fp16 = const()[name = tensor("layers_5_self_attn_v_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(59795136)))]; + tensor x_95_cast_fp16 = conv(bias = layers_5_self_attn_v_proj_module_bias_to_fp16, dilations = var_988, groups = var_925, pad = x_95_pad_0, pad_type = x_95_pad_type_0, strides = var_986, weight = layers_5_self_attn_v_proj_module_weight_to_fp16_palettized, x = input_71_cast_fp16)[name = tensor("x_95_cast_fp16")]; + tensor layers_5_self_attn_v_proj_output_scale_to_fp16 = const()[name = tensor("layers_5_self_attn_v_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(59797760)))]; + tensor value_11_cast_fp16 = mul(x = x_95_cast_fp16, y = layers_5_self_attn_v_proj_output_scale_to_fp16)[name = tensor("value_11_cast_fp16")]; + tensor var_993 = const()[name = tensor("op_993"), val = tensor([1, 20, 64, -1])]; + tensor var_994_cast_fp16 = reshape(shape = var_993, x = query_11_cast_fp16)[name = tensor("op_994_cast_fp16")]; + tensor var_995_to_fp16 = const()[name = tensor("op_995_to_fp16"), val = tensor(0x1p-3)]; + tensor var_996_cast_fp16 = mul(x = var_994_cast_fp16, y = var_995_to_fp16)[name = tensor("op_996_cast_fp16")]; + tensor var_997 = const()[name = tensor("op_997"), val = tensor([1, 20, 64, -1])]; + tensor var_998_cast_fp16 = reshape(shape = var_997, x = key_11_cast_fp16)[name = tensor("op_998_cast_fp16")]; + tensor mh_w_11_transpose_x_0 = const()[name = tensor("mh_w_11_transpose_x_0"), val = tensor(true)]; + tensor mh_w_11_transpose_y_0 = const()[name = tensor("mh_w_11_transpose_y_0"), val = tensor(false)]; + tensor mh_w_11_cast_fp16 = matmul(transpose_x = mh_w_11_transpose_x_0, transpose_y = mh_w_11_transpose_y_0, x = var_996_cast_fp16, y = var_998_cast_fp16)[name = tensor("mh_w_11_cast_fp16")]; + tensor var_1001_cast_fp16 = softmax(axis = var_923, x = mh_w_11_cast_fp16)[name = tensor("op_1001_cast_fp16")]; + tensor var_1002 = const()[name = tensor("op_1002"), val = tensor([1, 20, 64, -1])]; + tensor var_1003_cast_fp16 = reshape(shape = var_1002, x = value_11_cast_fp16)[name = tensor("op_1003_cast_fp16")]; + tensor attn_11_transpose_x_0 = const()[name = tensor("attn_11_transpose_x_0"), val = tensor(false)]; + tensor attn_11_transpose_y_0 = const()[name = tensor("attn_11_transpose_y_0"), val = tensor(true)]; + tensor attn_11_cast_fp16 = matmul(transpose_x = attn_11_transpose_x_0, transpose_y = attn_11_transpose_y_0, x = var_1003_cast_fp16, y = var_1001_cast_fp16)[name = tensor("attn_11_cast_fp16")]; + tensor var_1006 = const()[name = tensor("op_1006"), val = tensor([1, 1280, 1, -1])]; + tensor x_97_cast_fp16 = reshape(shape = var_1006, x = attn_11_cast_fp16)[name = tensor("x_97_cast_fp16")]; + tensor layers_5_self_attn_o_proj_input_shift_to_fp16 = const()[name = tensor("layers_5_self_attn_o_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(59800384)))]; + tensor input_77_cast_fp16 = sub(x = x_97_cast_fp16, y = layers_5_self_attn_o_proj_input_shift_to_fp16)[name = tensor("input_77_cast_fp16")]; + tensor var_1014 = const()[name = tensor("op_1014"), val = tensor([1, 1])]; + tensor var_1016 = const()[name = tensor("op_1016"), val = tensor([1, 1])]; + tensor x_99_pad_type_0 = const()[name = tensor("x_99_pad_type_0"), val = tensor("custom")]; + tensor x_99_pad_0 = const()[name = tensor("x_99_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_5_self_attn_o_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(59803008))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(60622272))), name = tensor("layers_5_self_attn_o_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_5_self_attn_o_proj_module_bias_to_fp16 = const()[name = tensor("layers_5_self_attn_o_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(60622400)))]; + tensor x_99_cast_fp16 = conv(bias = layers_5_self_attn_o_proj_module_bias_to_fp16, dilations = var_1016, groups = var_925, pad = x_99_pad_0, pad_type = x_99_pad_type_0, strides = var_1014, weight = layers_5_self_attn_o_proj_module_weight_to_fp16_palettized, x = input_77_cast_fp16)[name = tensor("x_99_cast_fp16")]; + tensor layers_5_self_attn_o_proj_output_scale_to_fp16 = const()[name = tensor("layers_5_self_attn_o_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(60625024)))]; + tensor obj_23_cast_fp16 = mul(x = x_99_cast_fp16, y = layers_5_self_attn_o_proj_output_scale_to_fp16)[name = tensor("obj_23_cast_fp16")]; + tensor inputs_23_cast_fp16 = add(x = inputs_21_cast_fp16, y = obj_23_cast_fp16)[name = tensor("inputs_23_cast_fp16")]; + tensor var_1023 = const()[name = tensor("op_1023"), val = tensor([1])]; + tensor channels_mean_23_cast_fp16 = reduce_mean(axes = var_1023, keep_dims = var_926, x = inputs_23_cast_fp16)[name = tensor("channels_mean_23_cast_fp16")]; + tensor zero_mean_23_cast_fp16 = sub(x = inputs_23_cast_fp16, y = channels_mean_23_cast_fp16)[name = tensor("zero_mean_23_cast_fp16")]; + tensor zero_mean_sq_23_cast_fp16 = mul(x = zero_mean_23_cast_fp16, y = zero_mean_23_cast_fp16)[name = tensor("zero_mean_sq_23_cast_fp16")]; + tensor var_1027 = const()[name = tensor("op_1027"), val = tensor([1])]; + tensor var_1028_cast_fp16 = reduce_mean(axes = var_1027, keep_dims = var_926, x = zero_mean_sq_23_cast_fp16)[name = tensor("op_1028_cast_fp16")]; + tensor var_1029_to_fp16 = const()[name = tensor("op_1029_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_1030_cast_fp16 = add(x = var_1028_cast_fp16, y = var_1029_to_fp16)[name = tensor("op_1030_cast_fp16")]; + tensor denom_23_epsilon_0_to_fp16 = const()[name = tensor("denom_23_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_23_cast_fp16 = rsqrt(epsilon = denom_23_epsilon_0_to_fp16, x = var_1030_cast_fp16)[name = tensor("denom_23_cast_fp16")]; + tensor out_23_cast_fp16 = mul(x = zero_mean_23_cast_fp16, y = denom_23_cast_fp16)[name = tensor("out_23_cast_fp16")]; + tensor x_101_gamma_0_to_fp16 = const()[name = tensor("x_101_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(60627648)))]; + tensor x_101_beta_0_to_fp16 = const()[name = tensor("x_101_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(60630272)))]; + tensor x_101_epsilon_0_to_fp16 = const()[name = tensor("x_101_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_101_cast_fp16 = batch_norm(beta = x_101_beta_0_to_fp16, epsilon = x_101_epsilon_0_to_fp16, gamma = x_101_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_23_cast_fp16)[name = tensor("x_101_cast_fp16")]; + tensor layers_5_fc1_input_shift_to_fp16 = const()[name = tensor("layers_5_fc1_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(60632896)))]; + tensor input_79_cast_fp16 = sub(x = x_101_cast_fp16, y = layers_5_fc1_input_shift_to_fp16)[name = tensor("input_79_cast_fp16")]; + tensor var_1045 = const()[name = tensor("op_1045"), val = tensor([1, 1])]; + tensor var_1047 = const()[name = tensor("op_1047"), val = tensor([1, 1])]; + tensor x_103_pad_type_0 = const()[name = tensor("x_103_pad_type_0"), val = tensor("custom")]; + tensor x_103_pad_0 = const()[name = tensor("x_103_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_5_fc1_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(60635520))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(63912384))), name = tensor("layers_5_fc1_module_weight_to_fp16_palettized"), shape = tensor([5120, 1280, 1, 1])]; + tensor layers_5_fc1_module_bias_to_fp16 = const()[name = tensor("layers_5_fc1_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(63912512)))]; + tensor x_103_cast_fp16 = conv(bias = layers_5_fc1_module_bias_to_fp16, dilations = var_1047, groups = var_925, pad = x_103_pad_0, pad_type = x_103_pad_type_0, strides = var_1045, weight = layers_5_fc1_module_weight_to_fp16_palettized, x = input_79_cast_fp16)[name = tensor("x_103_cast_fp16")]; + tensor layers_5_fc1_output_scale_to_fp16 = const()[name = tensor("layers_5_fc1_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(63922816)))]; + tensor input_81_cast_fp16 = mul(x = x_103_cast_fp16, y = layers_5_fc1_output_scale_to_fp16)[name = tensor("input_81_cast_fp16")]; + tensor x_105_mode_0 = const()[name = tensor("x_105_mode_0"), val = tensor("EXACT")]; + tensor x_105_cast_fp16 = gelu(mode = x_105_mode_0, x = input_81_cast_fp16)[name = tensor("x_105_cast_fp16")]; + tensor layers_5_fc2_input_shift_to_fp16 = const()[name = tensor("layers_5_fc2_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(63933120)))]; + tensor input_83_cast_fp16 = sub(x = x_105_cast_fp16, y = layers_5_fc2_input_shift_to_fp16)[name = tensor("input_83_cast_fp16")]; + tensor var_1058 = const()[name = tensor("op_1058"), val = tensor([1, 1])]; + tensor var_1060 = const()[name = tensor("op_1060"), val = tensor([1, 1])]; + tensor x_107_pad_type_0 = const()[name = tensor("x_107_pad_type_0"), val = tensor("custom")]; + tensor x_107_pad_0 = const()[name = tensor("x_107_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_5_fc2_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(63943424))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(67220288))), name = tensor("layers_5_fc2_module_weight_to_fp16_palettized"), shape = tensor([1280, 5120, 1, 1])]; + tensor layers_5_fc2_module_bias_to_fp16 = const()[name = tensor("layers_5_fc2_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(67220416)))]; + tensor x_107_cast_fp16 = conv(bias = layers_5_fc2_module_bias_to_fp16, dilations = var_1060, groups = var_925, pad = x_107_pad_0, pad_type = x_107_pad_type_0, strides = var_1058, weight = layers_5_fc2_module_weight_to_fp16_palettized, x = input_83_cast_fp16)[name = tensor("x_107_cast_fp16")]; + tensor layers_5_fc2_output_scale_to_fp16 = const()[name = tensor("layers_5_fc2_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(67223040)))]; + tensor hidden_states_15_cast_fp16 = mul(x = x_107_cast_fp16, y = layers_5_fc2_output_scale_to_fp16)[name = tensor("hidden_states_15_cast_fp16")]; + tensor inputs_25_cast_fp16 = add(x = inputs_23_cast_fp16, y = hidden_states_15_cast_fp16)[name = tensor("inputs_25_cast_fp16")]; + tensor var_1072 = const()[name = tensor("op_1072"), val = tensor(3)]; + tensor var_1074 = const()[name = tensor("op_1074"), val = tensor(1)]; + tensor var_1075 = const()[name = tensor("op_1075"), val = tensor(true)]; + tensor var_1085 = const()[name = tensor("op_1085"), val = tensor([1])]; + tensor channels_mean_25_cast_fp16 = reduce_mean(axes = var_1085, keep_dims = var_1075, x = inputs_25_cast_fp16)[name = tensor("channels_mean_25_cast_fp16")]; + tensor zero_mean_25_cast_fp16 = sub(x = inputs_25_cast_fp16, y = channels_mean_25_cast_fp16)[name = tensor("zero_mean_25_cast_fp16")]; + tensor zero_mean_sq_25_cast_fp16 = mul(x = zero_mean_25_cast_fp16, y = zero_mean_25_cast_fp16)[name = tensor("zero_mean_sq_25_cast_fp16")]; + tensor var_1089 = const()[name = tensor("op_1089"), val = tensor([1])]; + tensor var_1090_cast_fp16 = reduce_mean(axes = var_1089, keep_dims = var_1075, x = zero_mean_sq_25_cast_fp16)[name = tensor("op_1090_cast_fp16")]; + tensor var_1091_to_fp16 = const()[name = tensor("op_1091_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_1092_cast_fp16 = add(x = var_1090_cast_fp16, y = var_1091_to_fp16)[name = tensor("op_1092_cast_fp16")]; + tensor denom_25_epsilon_0_to_fp16 = const()[name = tensor("denom_25_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_25_cast_fp16 = rsqrt(epsilon = denom_25_epsilon_0_to_fp16, x = var_1092_cast_fp16)[name = tensor("denom_25_cast_fp16")]; + tensor out_25_cast_fp16 = mul(x = zero_mean_25_cast_fp16, y = denom_25_cast_fp16)[name = tensor("out_25_cast_fp16")]; + tensor obj_25_gamma_0_to_fp16 = const()[name = tensor("obj_25_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(67225664)))]; + tensor obj_25_beta_0_to_fp16 = const()[name = tensor("obj_25_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(67228288)))]; + tensor obj_25_epsilon_0_to_fp16 = const()[name = tensor("obj_25_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_25_cast_fp16 = batch_norm(beta = obj_25_beta_0_to_fp16, epsilon = obj_25_epsilon_0_to_fp16, gamma = obj_25_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_25_cast_fp16)[name = tensor("obj_25_cast_fp16")]; + tensor layers_6_self_attn_q_proj_input_shift_to_fp16 = const()[name = tensor("layers_6_self_attn_q_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(67230912)))]; + tensor input_85_cast_fp16 = sub(x = obj_25_cast_fp16, y = layers_6_self_attn_q_proj_input_shift_to_fp16)[name = tensor("input_85_cast_fp16")]; + tensor var_1111 = const()[name = tensor("op_1111"), val = tensor([1, 1])]; + tensor var_1113 = const()[name = tensor("op_1113"), val = tensor([1, 1])]; + tensor x_109_pad_type_0 = const()[name = tensor("x_109_pad_type_0"), val = tensor("custom")]; + tensor x_109_pad_0 = const()[name = tensor("x_109_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_6_self_attn_q_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(67233536))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(68052800))), name = tensor("layers_6_self_attn_q_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_6_self_attn_q_proj_module_bias_to_fp16 = const()[name = tensor("layers_6_self_attn_q_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(68052928)))]; + tensor x_109_cast_fp16 = conv(bias = layers_6_self_attn_q_proj_module_bias_to_fp16, dilations = var_1113, groups = var_1074, pad = x_109_pad_0, pad_type = x_109_pad_type_0, strides = var_1111, weight = layers_6_self_attn_q_proj_module_weight_to_fp16_palettized, x = input_85_cast_fp16)[name = tensor("x_109_cast_fp16")]; + tensor layers_6_self_attn_q_proj_output_scale_to_fp16 = const()[name = tensor("layers_6_self_attn_q_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(68055552)))]; + tensor query_13_cast_fp16 = mul(x = x_109_cast_fp16, y = layers_6_self_attn_q_proj_output_scale_to_fp16)[name = tensor("query_13_cast_fp16")]; + tensor var_1123 = const()[name = tensor("op_1123"), val = tensor([1, 1])]; + tensor var_1125 = const()[name = tensor("op_1125"), val = tensor([1, 1])]; + tensor x_111_pad_type_0 = const()[name = tensor("x_111_pad_type_0"), val = tensor("custom")]; + tensor x_111_pad_0 = const()[name = tensor("x_111_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_6_self_attn_k_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(68058176))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(68877440))), name = tensor("layers_6_self_attn_k_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_6_self_attn_k_proj_module_bias_to_fp16 = const()[name = tensor("layers_6_self_attn_k_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(68877568)))]; + tensor x_111_cast_fp16 = conv(bias = layers_6_self_attn_k_proj_module_bias_to_fp16, dilations = var_1125, groups = var_1074, pad = x_111_pad_0, pad_type = x_111_pad_type_0, strides = var_1123, weight = layers_6_self_attn_k_proj_module_weight_to_fp16_palettized, x = input_85_cast_fp16)[name = tensor("x_111_cast_fp16")]; + tensor layers_6_self_attn_k_proj_output_scale_to_fp16 = const()[name = tensor("layers_6_self_attn_k_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(68880192)))]; + tensor key_13_cast_fp16 = mul(x = x_111_cast_fp16, y = layers_6_self_attn_k_proj_output_scale_to_fp16)[name = tensor("key_13_cast_fp16")]; + tensor var_1135 = const()[name = tensor("op_1135"), val = tensor([1, 1])]; + tensor var_1137 = const()[name = tensor("op_1137"), val = tensor([1, 1])]; + tensor x_113_pad_type_0 = const()[name = tensor("x_113_pad_type_0"), val = tensor("custom")]; + tensor x_113_pad_0 = const()[name = tensor("x_113_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_6_self_attn_v_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(68882816))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(69702080))), name = tensor("layers_6_self_attn_v_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_6_self_attn_v_proj_module_bias_to_fp16 = const()[name = tensor("layers_6_self_attn_v_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(69702208)))]; + tensor x_113_cast_fp16 = conv(bias = layers_6_self_attn_v_proj_module_bias_to_fp16, dilations = var_1137, groups = var_1074, pad = x_113_pad_0, pad_type = x_113_pad_type_0, strides = var_1135, weight = layers_6_self_attn_v_proj_module_weight_to_fp16_palettized, x = input_85_cast_fp16)[name = tensor("x_113_cast_fp16")]; + tensor layers_6_self_attn_v_proj_output_scale_to_fp16 = const()[name = tensor("layers_6_self_attn_v_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(69704832)))]; + tensor value_13_cast_fp16 = mul(x = x_113_cast_fp16, y = layers_6_self_attn_v_proj_output_scale_to_fp16)[name = tensor("value_13_cast_fp16")]; + tensor var_1142 = const()[name = tensor("op_1142"), val = tensor([1, 20, 64, -1])]; + tensor var_1143_cast_fp16 = reshape(shape = var_1142, x = query_13_cast_fp16)[name = tensor("op_1143_cast_fp16")]; + tensor var_1144_to_fp16 = const()[name = tensor("op_1144_to_fp16"), val = tensor(0x1p-3)]; + tensor var_1145_cast_fp16 = mul(x = var_1143_cast_fp16, y = var_1144_to_fp16)[name = tensor("op_1145_cast_fp16")]; + tensor var_1146 = const()[name = tensor("op_1146"), val = tensor([1, 20, 64, -1])]; + tensor var_1147_cast_fp16 = reshape(shape = var_1146, x = key_13_cast_fp16)[name = tensor("op_1147_cast_fp16")]; + tensor mh_w_13_transpose_x_0 = const()[name = tensor("mh_w_13_transpose_x_0"), val = tensor(true)]; + tensor mh_w_13_transpose_y_0 = const()[name = tensor("mh_w_13_transpose_y_0"), val = tensor(false)]; + tensor mh_w_13_cast_fp16 = matmul(transpose_x = mh_w_13_transpose_x_0, transpose_y = mh_w_13_transpose_y_0, x = var_1145_cast_fp16, y = var_1147_cast_fp16)[name = tensor("mh_w_13_cast_fp16")]; + tensor var_1150_cast_fp16 = softmax(axis = var_1072, x = mh_w_13_cast_fp16)[name = tensor("op_1150_cast_fp16")]; + tensor var_1151 = const()[name = tensor("op_1151"), val = tensor([1, 20, 64, -1])]; + tensor var_1152_cast_fp16 = reshape(shape = var_1151, x = value_13_cast_fp16)[name = tensor("op_1152_cast_fp16")]; + tensor attn_13_transpose_x_0 = const()[name = tensor("attn_13_transpose_x_0"), val = tensor(false)]; + tensor attn_13_transpose_y_0 = const()[name = tensor("attn_13_transpose_y_0"), val = tensor(true)]; + tensor attn_13_cast_fp16 = matmul(transpose_x = attn_13_transpose_x_0, transpose_y = attn_13_transpose_y_0, x = var_1152_cast_fp16, y = var_1150_cast_fp16)[name = tensor("attn_13_cast_fp16")]; + tensor var_1155 = const()[name = tensor("op_1155"), val = tensor([1, 1280, 1, -1])]; + tensor x_115_cast_fp16 = reshape(shape = var_1155, x = attn_13_cast_fp16)[name = tensor("x_115_cast_fp16")]; + tensor layers_6_self_attn_o_proj_input_shift_to_fp16 = const()[name = tensor("layers_6_self_attn_o_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(69707456)))]; + tensor input_91_cast_fp16 = sub(x = x_115_cast_fp16, y = layers_6_self_attn_o_proj_input_shift_to_fp16)[name = tensor("input_91_cast_fp16")]; + tensor var_1163 = const()[name = tensor("op_1163"), val = tensor([1, 1])]; + tensor var_1165 = const()[name = tensor("op_1165"), val = tensor([1, 1])]; + tensor x_117_pad_type_0 = const()[name = tensor("x_117_pad_type_0"), val = tensor("custom")]; + tensor x_117_pad_0 = const()[name = tensor("x_117_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_6_self_attn_o_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(69710080))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(70529344))), name = tensor("layers_6_self_attn_o_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_6_self_attn_o_proj_module_bias_to_fp16 = const()[name = tensor("layers_6_self_attn_o_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(70529472)))]; + tensor x_117_cast_fp16 = conv(bias = layers_6_self_attn_o_proj_module_bias_to_fp16, dilations = var_1165, groups = var_1074, pad = x_117_pad_0, pad_type = x_117_pad_type_0, strides = var_1163, weight = layers_6_self_attn_o_proj_module_weight_to_fp16_palettized, x = input_91_cast_fp16)[name = tensor("x_117_cast_fp16")]; + tensor layers_6_self_attn_o_proj_output_scale_to_fp16 = const()[name = tensor("layers_6_self_attn_o_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(70532096)))]; + tensor obj_27_cast_fp16 = mul(x = x_117_cast_fp16, y = layers_6_self_attn_o_proj_output_scale_to_fp16)[name = tensor("obj_27_cast_fp16")]; + tensor inputs_27_cast_fp16 = add(x = inputs_25_cast_fp16, y = obj_27_cast_fp16)[name = tensor("inputs_27_cast_fp16")]; + tensor var_1172 = const()[name = tensor("op_1172"), val = tensor([1])]; + tensor channels_mean_27_cast_fp16 = reduce_mean(axes = var_1172, keep_dims = var_1075, x = inputs_27_cast_fp16)[name = tensor("channels_mean_27_cast_fp16")]; + tensor zero_mean_27_cast_fp16 = sub(x = inputs_27_cast_fp16, y = channels_mean_27_cast_fp16)[name = tensor("zero_mean_27_cast_fp16")]; + tensor zero_mean_sq_27_cast_fp16 = mul(x = zero_mean_27_cast_fp16, y = zero_mean_27_cast_fp16)[name = tensor("zero_mean_sq_27_cast_fp16")]; + tensor var_1176 = const()[name = tensor("op_1176"), val = tensor([1])]; + tensor var_1177_cast_fp16 = reduce_mean(axes = var_1176, keep_dims = var_1075, x = zero_mean_sq_27_cast_fp16)[name = tensor("op_1177_cast_fp16")]; + tensor var_1178_to_fp16 = const()[name = tensor("op_1178_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_1179_cast_fp16 = add(x = var_1177_cast_fp16, y = var_1178_to_fp16)[name = tensor("op_1179_cast_fp16")]; + tensor denom_27_epsilon_0_to_fp16 = const()[name = tensor("denom_27_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_27_cast_fp16 = rsqrt(epsilon = denom_27_epsilon_0_to_fp16, x = var_1179_cast_fp16)[name = tensor("denom_27_cast_fp16")]; + tensor out_27_cast_fp16 = mul(x = zero_mean_27_cast_fp16, y = denom_27_cast_fp16)[name = tensor("out_27_cast_fp16")]; + tensor x_119_gamma_0_to_fp16 = const()[name = tensor("x_119_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(70534720)))]; + tensor x_119_beta_0_to_fp16 = const()[name = tensor("x_119_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(70537344)))]; + tensor x_119_epsilon_0_to_fp16 = const()[name = tensor("x_119_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_119_cast_fp16 = batch_norm(beta = x_119_beta_0_to_fp16, epsilon = x_119_epsilon_0_to_fp16, gamma = x_119_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_27_cast_fp16)[name = tensor("x_119_cast_fp16")]; + tensor layers_6_fc1_input_shift_to_fp16 = const()[name = tensor("layers_6_fc1_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(70539968)))]; + tensor input_93_cast_fp16 = sub(x = x_119_cast_fp16, y = layers_6_fc1_input_shift_to_fp16)[name = tensor("input_93_cast_fp16")]; + tensor var_1194 = const()[name = tensor("op_1194"), val = tensor([1, 1])]; + tensor var_1196 = const()[name = tensor("op_1196"), val = tensor([1, 1])]; + tensor x_121_pad_type_0 = const()[name = tensor("x_121_pad_type_0"), val = tensor("custom")]; + tensor x_121_pad_0 = const()[name = tensor("x_121_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_6_fc1_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(70542592))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(73819456))), name = tensor("layers_6_fc1_module_weight_to_fp16_palettized"), shape = tensor([5120, 1280, 1, 1])]; + tensor layers_6_fc1_module_bias_to_fp16 = const()[name = tensor("layers_6_fc1_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(73819584)))]; + tensor x_121_cast_fp16 = conv(bias = layers_6_fc1_module_bias_to_fp16, dilations = var_1196, groups = var_1074, pad = x_121_pad_0, pad_type = x_121_pad_type_0, strides = var_1194, weight = layers_6_fc1_module_weight_to_fp16_palettized, x = input_93_cast_fp16)[name = tensor("x_121_cast_fp16")]; + tensor layers_6_fc1_output_scale_to_fp16 = const()[name = tensor("layers_6_fc1_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(73829888)))]; + tensor input_95_cast_fp16 = mul(x = x_121_cast_fp16, y = layers_6_fc1_output_scale_to_fp16)[name = tensor("input_95_cast_fp16")]; + tensor x_123_mode_0 = const()[name = tensor("x_123_mode_0"), val = tensor("EXACT")]; + tensor x_123_cast_fp16 = gelu(mode = x_123_mode_0, x = input_95_cast_fp16)[name = tensor("x_123_cast_fp16")]; + tensor layers_6_fc2_input_shift_to_fp16 = const()[name = tensor("layers_6_fc2_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(73840192)))]; + tensor input_97_cast_fp16 = sub(x = x_123_cast_fp16, y = layers_6_fc2_input_shift_to_fp16)[name = tensor("input_97_cast_fp16")]; + tensor var_1207 = const()[name = tensor("op_1207"), val = tensor([1, 1])]; + tensor var_1209 = const()[name = tensor("op_1209"), val = tensor([1, 1])]; + tensor x_125_pad_type_0 = const()[name = tensor("x_125_pad_type_0"), val = tensor("custom")]; + tensor x_125_pad_0 = const()[name = tensor("x_125_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_6_fc2_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(73850496))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(77127360))), name = tensor("layers_6_fc2_module_weight_to_fp16_palettized"), shape = tensor([1280, 5120, 1, 1])]; + tensor layers_6_fc2_module_bias_to_fp16 = const()[name = tensor("layers_6_fc2_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(77127488)))]; + tensor x_125_cast_fp16 = conv(bias = layers_6_fc2_module_bias_to_fp16, dilations = var_1209, groups = var_1074, pad = x_125_pad_0, pad_type = x_125_pad_type_0, strides = var_1207, weight = layers_6_fc2_module_weight_to_fp16_palettized, x = input_97_cast_fp16)[name = tensor("x_125_cast_fp16")]; + tensor layers_6_fc2_output_scale_to_fp16 = const()[name = tensor("layers_6_fc2_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(77130112)))]; + tensor hidden_states_17_cast_fp16 = mul(x = x_125_cast_fp16, y = layers_6_fc2_output_scale_to_fp16)[name = tensor("hidden_states_17_cast_fp16")]; + tensor inputs_29_cast_fp16 = add(x = inputs_27_cast_fp16, y = hidden_states_17_cast_fp16)[name = tensor("inputs_29_cast_fp16")]; + tensor var_1221 = const()[name = tensor("op_1221"), val = tensor(3)]; + tensor var_1223 = const()[name = tensor("op_1223"), val = tensor(1)]; + tensor var_1224 = const()[name = tensor("op_1224"), val = tensor(true)]; + tensor var_1234 = const()[name = tensor("op_1234"), val = tensor([1])]; + tensor channels_mean_29_cast_fp16 = reduce_mean(axes = var_1234, keep_dims = var_1224, x = inputs_29_cast_fp16)[name = tensor("channels_mean_29_cast_fp16")]; + tensor zero_mean_29_cast_fp16 = sub(x = inputs_29_cast_fp16, y = channels_mean_29_cast_fp16)[name = tensor("zero_mean_29_cast_fp16")]; + tensor zero_mean_sq_29_cast_fp16 = mul(x = zero_mean_29_cast_fp16, y = zero_mean_29_cast_fp16)[name = tensor("zero_mean_sq_29_cast_fp16")]; + tensor var_1238 = const()[name = tensor("op_1238"), val = tensor([1])]; + tensor var_1239_cast_fp16 = reduce_mean(axes = var_1238, keep_dims = var_1224, x = zero_mean_sq_29_cast_fp16)[name = tensor("op_1239_cast_fp16")]; + tensor var_1240_to_fp16 = const()[name = tensor("op_1240_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_1241_cast_fp16 = add(x = var_1239_cast_fp16, y = var_1240_to_fp16)[name = tensor("op_1241_cast_fp16")]; + tensor denom_29_epsilon_0_to_fp16 = const()[name = tensor("denom_29_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_29_cast_fp16 = rsqrt(epsilon = denom_29_epsilon_0_to_fp16, x = var_1241_cast_fp16)[name = tensor("denom_29_cast_fp16")]; + tensor out_29_cast_fp16 = mul(x = zero_mean_29_cast_fp16, y = denom_29_cast_fp16)[name = tensor("out_29_cast_fp16")]; + tensor obj_29_gamma_0_to_fp16 = const()[name = tensor("obj_29_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(77132736)))]; + tensor obj_29_beta_0_to_fp16 = const()[name = tensor("obj_29_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(77135360)))]; + tensor obj_29_epsilon_0_to_fp16 = const()[name = tensor("obj_29_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_29_cast_fp16 = batch_norm(beta = obj_29_beta_0_to_fp16, epsilon = obj_29_epsilon_0_to_fp16, gamma = obj_29_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_29_cast_fp16)[name = tensor("obj_29_cast_fp16")]; + tensor layers_7_self_attn_q_proj_input_shift_to_fp16 = const()[name = tensor("layers_7_self_attn_q_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(77137984)))]; + tensor input_99_cast_fp16 = sub(x = obj_29_cast_fp16, y = layers_7_self_attn_q_proj_input_shift_to_fp16)[name = tensor("input_99_cast_fp16")]; + tensor var_1260 = const()[name = tensor("op_1260"), val = tensor([1, 1])]; + tensor var_1262 = const()[name = tensor("op_1262"), val = tensor([1, 1])]; + tensor x_127_pad_type_0 = const()[name = tensor("x_127_pad_type_0"), val = tensor("custom")]; + tensor x_127_pad_0 = const()[name = tensor("x_127_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_7_self_attn_q_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(77140608))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(77959872))), name = tensor("layers_7_self_attn_q_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_7_self_attn_q_proj_module_bias_to_fp16 = const()[name = tensor("layers_7_self_attn_q_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(77960000)))]; + tensor x_127_cast_fp16 = conv(bias = layers_7_self_attn_q_proj_module_bias_to_fp16, dilations = var_1262, groups = var_1223, pad = x_127_pad_0, pad_type = x_127_pad_type_0, strides = var_1260, weight = layers_7_self_attn_q_proj_module_weight_to_fp16_palettized, x = input_99_cast_fp16)[name = tensor("x_127_cast_fp16")]; + tensor layers_7_self_attn_q_proj_output_scale_to_fp16 = const()[name = tensor("layers_7_self_attn_q_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(77962624)))]; + tensor query_15_cast_fp16 = mul(x = x_127_cast_fp16, y = layers_7_self_attn_q_proj_output_scale_to_fp16)[name = tensor("query_15_cast_fp16")]; + tensor var_1272 = const()[name = tensor("op_1272"), val = tensor([1, 1])]; + tensor var_1274 = const()[name = tensor("op_1274"), val = tensor([1, 1])]; + tensor x_129_pad_type_0 = const()[name = tensor("x_129_pad_type_0"), val = tensor("custom")]; + tensor x_129_pad_0 = const()[name = tensor("x_129_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_7_self_attn_k_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(77965248))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(78784512))), name = tensor("layers_7_self_attn_k_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_7_self_attn_k_proj_module_bias_to_fp16 = const()[name = tensor("layers_7_self_attn_k_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(78784640)))]; + tensor x_129_cast_fp16 = conv(bias = layers_7_self_attn_k_proj_module_bias_to_fp16, dilations = var_1274, groups = var_1223, pad = x_129_pad_0, pad_type = x_129_pad_type_0, strides = var_1272, weight = layers_7_self_attn_k_proj_module_weight_to_fp16_palettized, x = input_99_cast_fp16)[name = tensor("x_129_cast_fp16")]; + tensor layers_7_self_attn_k_proj_output_scale_to_fp16 = const()[name = tensor("layers_7_self_attn_k_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(78787264)))]; + tensor key_15_cast_fp16 = mul(x = x_129_cast_fp16, y = layers_7_self_attn_k_proj_output_scale_to_fp16)[name = tensor("key_15_cast_fp16")]; + tensor var_1284 = const()[name = tensor("op_1284"), val = tensor([1, 1])]; + tensor var_1286 = const()[name = tensor("op_1286"), val = tensor([1, 1])]; + tensor x_131_pad_type_0 = const()[name = tensor("x_131_pad_type_0"), val = tensor("custom")]; + tensor x_131_pad_0 = const()[name = tensor("x_131_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_7_self_attn_v_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(78789888))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(79609152))), name = tensor("layers_7_self_attn_v_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_7_self_attn_v_proj_module_bias_to_fp16 = const()[name = tensor("layers_7_self_attn_v_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(79609280)))]; + tensor x_131_cast_fp16 = conv(bias = layers_7_self_attn_v_proj_module_bias_to_fp16, dilations = var_1286, groups = var_1223, pad = x_131_pad_0, pad_type = x_131_pad_type_0, strides = var_1284, weight = layers_7_self_attn_v_proj_module_weight_to_fp16_palettized, x = input_99_cast_fp16)[name = tensor("x_131_cast_fp16")]; + tensor layers_7_self_attn_v_proj_output_scale_to_fp16 = const()[name = tensor("layers_7_self_attn_v_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(79611904)))]; + tensor value_15_cast_fp16 = mul(x = x_131_cast_fp16, y = layers_7_self_attn_v_proj_output_scale_to_fp16)[name = tensor("value_15_cast_fp16")]; + tensor var_1291 = const()[name = tensor("op_1291"), val = tensor([1, 20, 64, -1])]; + tensor var_1292_cast_fp16 = reshape(shape = var_1291, x = query_15_cast_fp16)[name = tensor("op_1292_cast_fp16")]; + tensor var_1293_to_fp16 = const()[name = tensor("op_1293_to_fp16"), val = tensor(0x1p-3)]; + tensor var_1294_cast_fp16 = mul(x = var_1292_cast_fp16, y = var_1293_to_fp16)[name = tensor("op_1294_cast_fp16")]; + tensor var_1295 = const()[name = tensor("op_1295"), val = tensor([1, 20, 64, -1])]; + tensor var_1296_cast_fp16 = reshape(shape = var_1295, x = key_15_cast_fp16)[name = tensor("op_1296_cast_fp16")]; + tensor mh_w_15_transpose_x_0 = const()[name = tensor("mh_w_15_transpose_x_0"), val = tensor(true)]; + tensor mh_w_15_transpose_y_0 = const()[name = tensor("mh_w_15_transpose_y_0"), val = tensor(false)]; + tensor mh_w_15_cast_fp16 = matmul(transpose_x = mh_w_15_transpose_x_0, transpose_y = mh_w_15_transpose_y_0, x = var_1294_cast_fp16, y = var_1296_cast_fp16)[name = tensor("mh_w_15_cast_fp16")]; + tensor var_1299_cast_fp16 = softmax(axis = var_1221, x = mh_w_15_cast_fp16)[name = tensor("op_1299_cast_fp16")]; + tensor var_1300 = const()[name = tensor("op_1300"), val = tensor([1, 20, 64, -1])]; + tensor var_1301_cast_fp16 = reshape(shape = var_1300, x = value_15_cast_fp16)[name = tensor("op_1301_cast_fp16")]; + tensor attn_15_transpose_x_0 = const()[name = tensor("attn_15_transpose_x_0"), val = tensor(false)]; + tensor attn_15_transpose_y_0 = const()[name = tensor("attn_15_transpose_y_0"), val = tensor(true)]; + tensor attn_15_cast_fp16 = matmul(transpose_x = attn_15_transpose_x_0, transpose_y = attn_15_transpose_y_0, x = var_1301_cast_fp16, y = var_1299_cast_fp16)[name = tensor("attn_15_cast_fp16")]; + tensor var_1304 = const()[name = tensor("op_1304"), val = tensor([1, 1280, 1, -1])]; + tensor x_133_cast_fp16 = reshape(shape = var_1304, x = attn_15_cast_fp16)[name = tensor("x_133_cast_fp16")]; + tensor layers_7_self_attn_o_proj_input_shift_to_fp16 = const()[name = tensor("layers_7_self_attn_o_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(79614528)))]; + tensor input_105_cast_fp16 = sub(x = x_133_cast_fp16, y = layers_7_self_attn_o_proj_input_shift_to_fp16)[name = tensor("input_105_cast_fp16")]; + tensor var_1312 = const()[name = tensor("op_1312"), val = tensor([1, 1])]; + tensor var_1314 = const()[name = tensor("op_1314"), val = tensor([1, 1])]; + tensor x_135_pad_type_0 = const()[name = tensor("x_135_pad_type_0"), val = tensor("custom")]; + tensor x_135_pad_0 = const()[name = tensor("x_135_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_7_self_attn_o_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(79617152))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(80436416))), name = tensor("layers_7_self_attn_o_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_7_self_attn_o_proj_module_bias_to_fp16 = const()[name = tensor("layers_7_self_attn_o_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(80436544)))]; + tensor x_135_cast_fp16 = conv(bias = layers_7_self_attn_o_proj_module_bias_to_fp16, dilations = var_1314, groups = var_1223, pad = x_135_pad_0, pad_type = x_135_pad_type_0, strides = var_1312, weight = layers_7_self_attn_o_proj_module_weight_to_fp16_palettized, x = input_105_cast_fp16)[name = tensor("x_135_cast_fp16")]; + tensor layers_7_self_attn_o_proj_output_scale_to_fp16 = const()[name = tensor("layers_7_self_attn_o_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(80439168)))]; + tensor obj_31_cast_fp16 = mul(x = x_135_cast_fp16, y = layers_7_self_attn_o_proj_output_scale_to_fp16)[name = tensor("obj_31_cast_fp16")]; + tensor inputs_31_cast_fp16 = add(x = inputs_29_cast_fp16, y = obj_31_cast_fp16)[name = tensor("inputs_31_cast_fp16")]; + tensor var_1321 = const()[name = tensor("op_1321"), val = tensor([1])]; + tensor channels_mean_31_cast_fp16 = reduce_mean(axes = var_1321, keep_dims = var_1224, x = inputs_31_cast_fp16)[name = tensor("channels_mean_31_cast_fp16")]; + tensor zero_mean_31_cast_fp16 = sub(x = inputs_31_cast_fp16, y = channels_mean_31_cast_fp16)[name = tensor("zero_mean_31_cast_fp16")]; + tensor zero_mean_sq_31_cast_fp16 = mul(x = zero_mean_31_cast_fp16, y = zero_mean_31_cast_fp16)[name = tensor("zero_mean_sq_31_cast_fp16")]; + tensor var_1325 = const()[name = tensor("op_1325"), val = tensor([1])]; + tensor var_1326_cast_fp16 = reduce_mean(axes = var_1325, keep_dims = var_1224, x = zero_mean_sq_31_cast_fp16)[name = tensor("op_1326_cast_fp16")]; + tensor var_1327_to_fp16 = const()[name = tensor("op_1327_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_1328_cast_fp16 = add(x = var_1326_cast_fp16, y = var_1327_to_fp16)[name = tensor("op_1328_cast_fp16")]; + tensor denom_31_epsilon_0_to_fp16 = const()[name = tensor("denom_31_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_31_cast_fp16 = rsqrt(epsilon = denom_31_epsilon_0_to_fp16, x = var_1328_cast_fp16)[name = tensor("denom_31_cast_fp16")]; + tensor out_31_cast_fp16 = mul(x = zero_mean_31_cast_fp16, y = denom_31_cast_fp16)[name = tensor("out_31_cast_fp16")]; + tensor x_137_gamma_0_to_fp16 = const()[name = tensor("x_137_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(80441792)))]; + tensor x_137_beta_0_to_fp16 = const()[name = tensor("x_137_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(80444416)))]; + tensor x_137_epsilon_0_to_fp16 = const()[name = tensor("x_137_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_137_cast_fp16 = batch_norm(beta = x_137_beta_0_to_fp16, epsilon = x_137_epsilon_0_to_fp16, gamma = x_137_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_31_cast_fp16)[name = tensor("x_137_cast_fp16")]; + tensor layers_7_fc1_input_shift_to_fp16 = const()[name = tensor("layers_7_fc1_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(80447040)))]; + tensor input_107_cast_fp16 = sub(x = x_137_cast_fp16, y = layers_7_fc1_input_shift_to_fp16)[name = tensor("input_107_cast_fp16")]; + tensor var_1343 = const()[name = tensor("op_1343"), val = tensor([1, 1])]; + tensor var_1345 = const()[name = tensor("op_1345"), val = tensor([1, 1])]; + tensor x_139_pad_type_0 = const()[name = tensor("x_139_pad_type_0"), val = tensor("custom")]; + tensor x_139_pad_0 = const()[name = tensor("x_139_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_7_fc1_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(80449664))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(83726528))), name = tensor("layers_7_fc1_module_weight_to_fp16_palettized"), shape = tensor([5120, 1280, 1, 1])]; + tensor layers_7_fc1_module_bias_to_fp16 = const()[name = tensor("layers_7_fc1_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(83726656)))]; + tensor x_139_cast_fp16 = conv(bias = layers_7_fc1_module_bias_to_fp16, dilations = var_1345, groups = var_1223, pad = x_139_pad_0, pad_type = x_139_pad_type_0, strides = var_1343, weight = layers_7_fc1_module_weight_to_fp16_palettized, x = input_107_cast_fp16)[name = tensor("x_139_cast_fp16")]; + tensor layers_7_fc1_output_scale_to_fp16 = const()[name = tensor("layers_7_fc1_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(83736960)))]; + tensor input_109_cast_fp16 = mul(x = x_139_cast_fp16, y = layers_7_fc1_output_scale_to_fp16)[name = tensor("input_109_cast_fp16")]; + tensor x_141_mode_0 = const()[name = tensor("x_141_mode_0"), val = tensor("EXACT")]; + tensor x_141_cast_fp16 = gelu(mode = x_141_mode_0, x = input_109_cast_fp16)[name = tensor("x_141_cast_fp16")]; + tensor layers_7_fc2_input_shift_to_fp16 = const()[name = tensor("layers_7_fc2_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(83747264)))]; + tensor input_111_cast_fp16 = sub(x = x_141_cast_fp16, y = layers_7_fc2_input_shift_to_fp16)[name = tensor("input_111_cast_fp16")]; + tensor var_1356 = const()[name = tensor("op_1356"), val = tensor([1, 1])]; + tensor var_1358 = const()[name = tensor("op_1358"), val = tensor([1, 1])]; + tensor x_143_pad_type_0 = const()[name = tensor("x_143_pad_type_0"), val = tensor("custom")]; + tensor x_143_pad_0 = const()[name = tensor("x_143_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_7_fc2_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(83757568))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(87034432))), name = tensor("layers_7_fc2_module_weight_to_fp16_palettized"), shape = tensor([1280, 5120, 1, 1])]; + tensor layers_7_fc2_module_bias_to_fp16 = const()[name = tensor("layers_7_fc2_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(87034560)))]; + tensor x_143_cast_fp16 = conv(bias = layers_7_fc2_module_bias_to_fp16, dilations = var_1358, groups = var_1223, pad = x_143_pad_0, pad_type = x_143_pad_type_0, strides = var_1356, weight = layers_7_fc2_module_weight_to_fp16_palettized, x = input_111_cast_fp16)[name = tensor("x_143_cast_fp16")]; + tensor layers_7_fc2_output_scale_to_fp16 = const()[name = tensor("layers_7_fc2_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(87037184)))]; + tensor hidden_states_19_cast_fp16 = mul(x = x_143_cast_fp16, y = layers_7_fc2_output_scale_to_fp16)[name = tensor("hidden_states_19_cast_fp16")]; + tensor inputs_33_cast_fp16 = add(x = inputs_31_cast_fp16, y = hidden_states_19_cast_fp16)[name = tensor("inputs_33_cast_fp16")]; + tensor var_1370 = const()[name = tensor("op_1370"), val = tensor(3)]; + tensor var_1372 = const()[name = tensor("op_1372"), val = tensor(1)]; + tensor var_1373 = const()[name = tensor("op_1373"), val = tensor(true)]; + tensor var_1383 = const()[name = tensor("op_1383"), val = tensor([1])]; + tensor channels_mean_33_cast_fp16 = reduce_mean(axes = var_1383, keep_dims = var_1373, x = inputs_33_cast_fp16)[name = tensor("channels_mean_33_cast_fp16")]; + tensor zero_mean_33_cast_fp16 = sub(x = inputs_33_cast_fp16, y = channels_mean_33_cast_fp16)[name = tensor("zero_mean_33_cast_fp16")]; + tensor zero_mean_sq_33_cast_fp16 = mul(x = zero_mean_33_cast_fp16, y = zero_mean_33_cast_fp16)[name = tensor("zero_mean_sq_33_cast_fp16")]; + tensor var_1387 = const()[name = tensor("op_1387"), val = tensor([1])]; + tensor var_1388_cast_fp16 = reduce_mean(axes = var_1387, keep_dims = var_1373, x = zero_mean_sq_33_cast_fp16)[name = tensor("op_1388_cast_fp16")]; + tensor var_1389_to_fp16 = const()[name = tensor("op_1389_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_1390_cast_fp16 = add(x = var_1388_cast_fp16, y = var_1389_to_fp16)[name = tensor("op_1390_cast_fp16")]; + tensor denom_33_epsilon_0_to_fp16 = const()[name = tensor("denom_33_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_33_cast_fp16 = rsqrt(epsilon = denom_33_epsilon_0_to_fp16, x = var_1390_cast_fp16)[name = tensor("denom_33_cast_fp16")]; + tensor out_33_cast_fp16 = mul(x = zero_mean_33_cast_fp16, y = denom_33_cast_fp16)[name = tensor("out_33_cast_fp16")]; + tensor obj_33_gamma_0_to_fp16 = const()[name = tensor("obj_33_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(87039808)))]; + tensor obj_33_beta_0_to_fp16 = const()[name = tensor("obj_33_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(87042432)))]; + tensor obj_33_epsilon_0_to_fp16 = const()[name = tensor("obj_33_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_33_cast_fp16 = batch_norm(beta = obj_33_beta_0_to_fp16, epsilon = obj_33_epsilon_0_to_fp16, gamma = obj_33_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_33_cast_fp16)[name = tensor("obj_33_cast_fp16")]; + tensor layers_8_self_attn_q_proj_input_shift_to_fp16 = const()[name = tensor("layers_8_self_attn_q_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(87045056)))]; + tensor input_113_cast_fp16 = sub(x = obj_33_cast_fp16, y = layers_8_self_attn_q_proj_input_shift_to_fp16)[name = tensor("input_113_cast_fp16")]; + tensor var_1409 = const()[name = tensor("op_1409"), val = tensor([1, 1])]; + tensor var_1411 = const()[name = tensor("op_1411"), val = tensor([1, 1])]; + tensor x_145_pad_type_0 = const()[name = tensor("x_145_pad_type_0"), val = tensor("custom")]; + tensor x_145_pad_0 = const()[name = tensor("x_145_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_8_self_attn_q_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(87047680))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(87866944))), name = tensor("layers_8_self_attn_q_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_8_self_attn_q_proj_module_bias_to_fp16 = const()[name = tensor("layers_8_self_attn_q_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(87867072)))]; + tensor x_145_cast_fp16 = conv(bias = layers_8_self_attn_q_proj_module_bias_to_fp16, dilations = var_1411, groups = var_1372, pad = x_145_pad_0, pad_type = x_145_pad_type_0, strides = var_1409, weight = layers_8_self_attn_q_proj_module_weight_to_fp16_palettized, x = input_113_cast_fp16)[name = tensor("x_145_cast_fp16")]; + tensor layers_8_self_attn_q_proj_output_scale_to_fp16 = const()[name = tensor("layers_8_self_attn_q_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(87869696)))]; + tensor query_17_cast_fp16 = mul(x = x_145_cast_fp16, y = layers_8_self_attn_q_proj_output_scale_to_fp16)[name = tensor("query_17_cast_fp16")]; + tensor var_1421 = const()[name = tensor("op_1421"), val = tensor([1, 1])]; + tensor var_1423 = const()[name = tensor("op_1423"), val = tensor([1, 1])]; + tensor x_147_pad_type_0 = const()[name = tensor("x_147_pad_type_0"), val = tensor("custom")]; + tensor x_147_pad_0 = const()[name = tensor("x_147_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_8_self_attn_k_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(87872320))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(88691584))), name = tensor("layers_8_self_attn_k_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_8_self_attn_k_proj_module_bias_to_fp16 = const()[name = tensor("layers_8_self_attn_k_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(88691712)))]; + tensor x_147_cast_fp16 = conv(bias = layers_8_self_attn_k_proj_module_bias_to_fp16, dilations = var_1423, groups = var_1372, pad = x_147_pad_0, pad_type = x_147_pad_type_0, strides = var_1421, weight = layers_8_self_attn_k_proj_module_weight_to_fp16_palettized, x = input_113_cast_fp16)[name = tensor("x_147_cast_fp16")]; + tensor layers_8_self_attn_k_proj_output_scale_to_fp16 = const()[name = tensor("layers_8_self_attn_k_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(88694336)))]; + tensor key_17_cast_fp16 = mul(x = x_147_cast_fp16, y = layers_8_self_attn_k_proj_output_scale_to_fp16)[name = tensor("key_17_cast_fp16")]; + tensor var_1433 = const()[name = tensor("op_1433"), val = tensor([1, 1])]; + tensor var_1435 = const()[name = tensor("op_1435"), val = tensor([1, 1])]; + tensor x_149_pad_type_0 = const()[name = tensor("x_149_pad_type_0"), val = tensor("custom")]; + tensor x_149_pad_0 = const()[name = tensor("x_149_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_8_self_attn_v_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(88696960))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(89516224))), name = tensor("layers_8_self_attn_v_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_8_self_attn_v_proj_module_bias_to_fp16 = const()[name = tensor("layers_8_self_attn_v_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(89516352)))]; + tensor x_149_cast_fp16 = conv(bias = layers_8_self_attn_v_proj_module_bias_to_fp16, dilations = var_1435, groups = var_1372, pad = x_149_pad_0, pad_type = x_149_pad_type_0, strides = var_1433, weight = layers_8_self_attn_v_proj_module_weight_to_fp16_palettized, x = input_113_cast_fp16)[name = tensor("x_149_cast_fp16")]; + tensor layers_8_self_attn_v_proj_output_scale_to_fp16 = const()[name = tensor("layers_8_self_attn_v_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(89518976)))]; + tensor value_17_cast_fp16 = mul(x = x_149_cast_fp16, y = layers_8_self_attn_v_proj_output_scale_to_fp16)[name = tensor("value_17_cast_fp16")]; + tensor var_1440 = const()[name = tensor("op_1440"), val = tensor([1, 20, 64, -1])]; + tensor var_1441_cast_fp16 = reshape(shape = var_1440, x = query_17_cast_fp16)[name = tensor("op_1441_cast_fp16")]; + tensor var_1442_to_fp16 = const()[name = tensor("op_1442_to_fp16"), val = tensor(0x1p-3)]; + tensor var_1443_cast_fp16 = mul(x = var_1441_cast_fp16, y = var_1442_to_fp16)[name = tensor("op_1443_cast_fp16")]; + tensor var_1444 = const()[name = tensor("op_1444"), val = tensor([1, 20, 64, -1])]; + tensor var_1445_cast_fp16 = reshape(shape = var_1444, x = key_17_cast_fp16)[name = tensor("op_1445_cast_fp16")]; + tensor mh_w_17_transpose_x_0 = const()[name = tensor("mh_w_17_transpose_x_0"), val = tensor(true)]; + tensor mh_w_17_transpose_y_0 = const()[name = tensor("mh_w_17_transpose_y_0"), val = tensor(false)]; + tensor mh_w_17_cast_fp16 = matmul(transpose_x = mh_w_17_transpose_x_0, transpose_y = mh_w_17_transpose_y_0, x = var_1443_cast_fp16, y = var_1445_cast_fp16)[name = tensor("mh_w_17_cast_fp16")]; + tensor var_1448_cast_fp16 = softmax(axis = var_1370, x = mh_w_17_cast_fp16)[name = tensor("op_1448_cast_fp16")]; + tensor var_1449 = const()[name = tensor("op_1449"), val = tensor([1, 20, 64, -1])]; + tensor var_1450_cast_fp16 = reshape(shape = var_1449, x = value_17_cast_fp16)[name = tensor("op_1450_cast_fp16")]; + tensor attn_17_transpose_x_0 = const()[name = tensor("attn_17_transpose_x_0"), val = tensor(false)]; + tensor attn_17_transpose_y_0 = const()[name = tensor("attn_17_transpose_y_0"), val = tensor(true)]; + tensor attn_17_cast_fp16 = matmul(transpose_x = attn_17_transpose_x_0, transpose_y = attn_17_transpose_y_0, x = var_1450_cast_fp16, y = var_1448_cast_fp16)[name = tensor("attn_17_cast_fp16")]; + tensor var_1453 = const()[name = tensor("op_1453"), val = tensor([1, 1280, 1, -1])]; + tensor x_151_cast_fp16 = reshape(shape = var_1453, x = attn_17_cast_fp16)[name = tensor("x_151_cast_fp16")]; + tensor layers_8_self_attn_o_proj_input_shift_to_fp16 = const()[name = tensor("layers_8_self_attn_o_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(89521600)))]; + tensor input_119_cast_fp16 = sub(x = x_151_cast_fp16, y = layers_8_self_attn_o_proj_input_shift_to_fp16)[name = tensor("input_119_cast_fp16")]; + tensor var_1461 = const()[name = tensor("op_1461"), val = tensor([1, 1])]; + tensor var_1463 = const()[name = tensor("op_1463"), val = tensor([1, 1])]; + tensor x_153_pad_type_0 = const()[name = tensor("x_153_pad_type_0"), val = tensor("custom")]; + tensor x_153_pad_0 = const()[name = tensor("x_153_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_8_self_attn_o_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(89524224))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(90343488))), name = tensor("layers_8_self_attn_o_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_8_self_attn_o_proj_module_bias_to_fp16 = const()[name = tensor("layers_8_self_attn_o_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(90343616)))]; + tensor x_153_cast_fp16 = conv(bias = layers_8_self_attn_o_proj_module_bias_to_fp16, dilations = var_1463, groups = var_1372, pad = x_153_pad_0, pad_type = x_153_pad_type_0, strides = var_1461, weight = layers_8_self_attn_o_proj_module_weight_to_fp16_palettized, x = input_119_cast_fp16)[name = tensor("x_153_cast_fp16")]; + tensor layers_8_self_attn_o_proj_output_scale_to_fp16 = const()[name = tensor("layers_8_self_attn_o_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(90346240)))]; + tensor obj_35_cast_fp16 = mul(x = x_153_cast_fp16, y = layers_8_self_attn_o_proj_output_scale_to_fp16)[name = tensor("obj_35_cast_fp16")]; + tensor inputs_35_cast_fp16 = add(x = inputs_33_cast_fp16, y = obj_35_cast_fp16)[name = tensor("inputs_35_cast_fp16")]; + tensor var_1470 = const()[name = tensor("op_1470"), val = tensor([1])]; + tensor channels_mean_35_cast_fp16 = reduce_mean(axes = var_1470, keep_dims = var_1373, x = inputs_35_cast_fp16)[name = tensor("channels_mean_35_cast_fp16")]; + tensor zero_mean_35_cast_fp16 = sub(x = inputs_35_cast_fp16, y = channels_mean_35_cast_fp16)[name = tensor("zero_mean_35_cast_fp16")]; + tensor zero_mean_sq_35_cast_fp16 = mul(x = zero_mean_35_cast_fp16, y = zero_mean_35_cast_fp16)[name = tensor("zero_mean_sq_35_cast_fp16")]; + tensor var_1474 = const()[name = tensor("op_1474"), val = tensor([1])]; + tensor var_1475_cast_fp16 = reduce_mean(axes = var_1474, keep_dims = var_1373, x = zero_mean_sq_35_cast_fp16)[name = tensor("op_1475_cast_fp16")]; + tensor var_1476_to_fp16 = const()[name = tensor("op_1476_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_1477_cast_fp16 = add(x = var_1475_cast_fp16, y = var_1476_to_fp16)[name = tensor("op_1477_cast_fp16")]; + tensor denom_35_epsilon_0_to_fp16 = const()[name = tensor("denom_35_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_35_cast_fp16 = rsqrt(epsilon = denom_35_epsilon_0_to_fp16, x = var_1477_cast_fp16)[name = tensor("denom_35_cast_fp16")]; + tensor out_35_cast_fp16 = mul(x = zero_mean_35_cast_fp16, y = denom_35_cast_fp16)[name = tensor("out_35_cast_fp16")]; + tensor x_155_gamma_0_to_fp16 = const()[name = tensor("x_155_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(90348864)))]; + tensor x_155_beta_0_to_fp16 = const()[name = tensor("x_155_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(90351488)))]; + tensor x_155_epsilon_0_to_fp16 = const()[name = tensor("x_155_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_155_cast_fp16 = batch_norm(beta = x_155_beta_0_to_fp16, epsilon = x_155_epsilon_0_to_fp16, gamma = x_155_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_35_cast_fp16)[name = tensor("x_155_cast_fp16")]; + tensor layers_8_fc1_input_shift_to_fp16 = const()[name = tensor("layers_8_fc1_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(90354112)))]; + tensor input_121_cast_fp16 = sub(x = x_155_cast_fp16, y = layers_8_fc1_input_shift_to_fp16)[name = tensor("input_121_cast_fp16")]; + tensor var_1492 = const()[name = tensor("op_1492"), val = tensor([1, 1])]; + tensor var_1494 = const()[name = tensor("op_1494"), val = tensor([1, 1])]; + tensor x_157_pad_type_0 = const()[name = tensor("x_157_pad_type_0"), val = tensor("custom")]; + tensor x_157_pad_0 = const()[name = tensor("x_157_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_8_fc1_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(90356736))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(93633600))), name = tensor("layers_8_fc1_module_weight_to_fp16_palettized"), shape = tensor([5120, 1280, 1, 1])]; + tensor layers_8_fc1_module_bias_to_fp16 = const()[name = tensor("layers_8_fc1_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(93633728)))]; + tensor x_157_cast_fp16 = conv(bias = layers_8_fc1_module_bias_to_fp16, dilations = var_1494, groups = var_1372, pad = x_157_pad_0, pad_type = x_157_pad_type_0, strides = var_1492, weight = layers_8_fc1_module_weight_to_fp16_palettized, x = input_121_cast_fp16)[name = tensor("x_157_cast_fp16")]; + tensor layers_8_fc1_output_scale_to_fp16 = const()[name = tensor("layers_8_fc1_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(93644032)))]; + tensor input_123_cast_fp16 = mul(x = x_157_cast_fp16, y = layers_8_fc1_output_scale_to_fp16)[name = tensor("input_123_cast_fp16")]; + tensor x_159_mode_0 = const()[name = tensor("x_159_mode_0"), val = tensor("EXACT")]; + tensor x_159_cast_fp16 = gelu(mode = x_159_mode_0, x = input_123_cast_fp16)[name = tensor("x_159_cast_fp16")]; + tensor layers_8_fc2_input_shift_to_fp16 = const()[name = tensor("layers_8_fc2_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(93654336)))]; + tensor input_125_cast_fp16 = sub(x = x_159_cast_fp16, y = layers_8_fc2_input_shift_to_fp16)[name = tensor("input_125_cast_fp16")]; + tensor var_1505 = const()[name = tensor("op_1505"), val = tensor([1, 1])]; + tensor var_1507 = const()[name = tensor("op_1507"), val = tensor([1, 1])]; + tensor x_161_pad_type_0 = const()[name = tensor("x_161_pad_type_0"), val = tensor("custom")]; + tensor x_161_pad_0 = const()[name = tensor("x_161_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_8_fc2_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(93664640))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(96941504))), name = tensor("layers_8_fc2_module_weight_to_fp16_palettized"), shape = tensor([1280, 5120, 1, 1])]; + tensor layers_8_fc2_module_bias_to_fp16 = const()[name = tensor("layers_8_fc2_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(96941632)))]; + tensor x_161_cast_fp16 = conv(bias = layers_8_fc2_module_bias_to_fp16, dilations = var_1507, groups = var_1372, pad = x_161_pad_0, pad_type = x_161_pad_type_0, strides = var_1505, weight = layers_8_fc2_module_weight_to_fp16_palettized, x = input_125_cast_fp16)[name = tensor("x_161_cast_fp16")]; + tensor layers_8_fc2_output_scale_to_fp16 = const()[name = tensor("layers_8_fc2_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(96944256)))]; + tensor hidden_states_21_cast_fp16 = mul(x = x_161_cast_fp16, y = layers_8_fc2_output_scale_to_fp16)[name = tensor("hidden_states_21_cast_fp16")]; + tensor inputs_37_cast_fp16 = add(x = inputs_35_cast_fp16, y = hidden_states_21_cast_fp16)[name = tensor("inputs_37_cast_fp16")]; + tensor var_1519 = const()[name = tensor("op_1519"), val = tensor(3)]; + tensor var_1521 = const()[name = tensor("op_1521"), val = tensor(1)]; + tensor var_1522 = const()[name = tensor("op_1522"), val = tensor(true)]; + tensor var_1532 = const()[name = tensor("op_1532"), val = tensor([1])]; + tensor channels_mean_37_cast_fp16 = reduce_mean(axes = var_1532, keep_dims = var_1522, x = inputs_37_cast_fp16)[name = tensor("channels_mean_37_cast_fp16")]; + tensor zero_mean_37_cast_fp16 = sub(x = inputs_37_cast_fp16, y = channels_mean_37_cast_fp16)[name = tensor("zero_mean_37_cast_fp16")]; + tensor zero_mean_sq_37_cast_fp16 = mul(x = zero_mean_37_cast_fp16, y = zero_mean_37_cast_fp16)[name = tensor("zero_mean_sq_37_cast_fp16")]; + tensor var_1536 = const()[name = tensor("op_1536"), val = tensor([1])]; + tensor var_1537_cast_fp16 = reduce_mean(axes = var_1536, keep_dims = var_1522, x = zero_mean_sq_37_cast_fp16)[name = tensor("op_1537_cast_fp16")]; + tensor var_1538_to_fp16 = const()[name = tensor("op_1538_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_1539_cast_fp16 = add(x = var_1537_cast_fp16, y = var_1538_to_fp16)[name = tensor("op_1539_cast_fp16")]; + tensor denom_37_epsilon_0_to_fp16 = const()[name = tensor("denom_37_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_37_cast_fp16 = rsqrt(epsilon = denom_37_epsilon_0_to_fp16, x = var_1539_cast_fp16)[name = tensor("denom_37_cast_fp16")]; + tensor out_37_cast_fp16 = mul(x = zero_mean_37_cast_fp16, y = denom_37_cast_fp16)[name = tensor("out_37_cast_fp16")]; + tensor obj_37_gamma_0_to_fp16 = const()[name = tensor("obj_37_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(96946880)))]; + tensor obj_37_beta_0_to_fp16 = const()[name = tensor("obj_37_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(96949504)))]; + tensor obj_37_epsilon_0_to_fp16 = const()[name = tensor("obj_37_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_37_cast_fp16 = batch_norm(beta = obj_37_beta_0_to_fp16, epsilon = obj_37_epsilon_0_to_fp16, gamma = obj_37_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_37_cast_fp16)[name = tensor("obj_37_cast_fp16")]; + tensor layers_9_self_attn_q_proj_input_shift_to_fp16 = const()[name = tensor("layers_9_self_attn_q_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(96952128)))]; + tensor input_127_cast_fp16 = sub(x = obj_37_cast_fp16, y = layers_9_self_attn_q_proj_input_shift_to_fp16)[name = tensor("input_127_cast_fp16")]; + tensor var_1558 = const()[name = tensor("op_1558"), val = tensor([1, 1])]; + tensor var_1560 = const()[name = tensor("op_1560"), val = tensor([1, 1])]; + tensor x_163_pad_type_0 = const()[name = tensor("x_163_pad_type_0"), val = tensor("custom")]; + tensor x_163_pad_0 = const()[name = tensor("x_163_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_9_self_attn_q_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(96954752))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(97774016))), name = tensor("layers_9_self_attn_q_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_9_self_attn_q_proj_module_bias_to_fp16 = const()[name = tensor("layers_9_self_attn_q_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(97774144)))]; + tensor x_163_cast_fp16 = conv(bias = layers_9_self_attn_q_proj_module_bias_to_fp16, dilations = var_1560, groups = var_1521, pad = x_163_pad_0, pad_type = x_163_pad_type_0, strides = var_1558, weight = layers_9_self_attn_q_proj_module_weight_to_fp16_palettized, x = input_127_cast_fp16)[name = tensor("x_163_cast_fp16")]; + tensor layers_9_self_attn_q_proj_output_scale_to_fp16 = const()[name = tensor("layers_9_self_attn_q_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(97776768)))]; + tensor query_19_cast_fp16 = mul(x = x_163_cast_fp16, y = layers_9_self_attn_q_proj_output_scale_to_fp16)[name = tensor("query_19_cast_fp16")]; + tensor var_1570 = const()[name = tensor("op_1570"), val = tensor([1, 1])]; + tensor var_1572 = const()[name = tensor("op_1572"), val = tensor([1, 1])]; + tensor x_165_pad_type_0 = const()[name = tensor("x_165_pad_type_0"), val = tensor("custom")]; + tensor x_165_pad_0 = const()[name = tensor("x_165_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_9_self_attn_k_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(97779392))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(98598656))), name = tensor("layers_9_self_attn_k_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_9_self_attn_k_proj_module_bias_to_fp16 = const()[name = tensor("layers_9_self_attn_k_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(98598784)))]; + tensor x_165_cast_fp16 = conv(bias = layers_9_self_attn_k_proj_module_bias_to_fp16, dilations = var_1572, groups = var_1521, pad = x_165_pad_0, pad_type = x_165_pad_type_0, strides = var_1570, weight = layers_9_self_attn_k_proj_module_weight_to_fp16_palettized, x = input_127_cast_fp16)[name = tensor("x_165_cast_fp16")]; + tensor layers_9_self_attn_k_proj_output_scale_to_fp16 = const()[name = tensor("layers_9_self_attn_k_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(98601408)))]; + tensor key_19_cast_fp16 = mul(x = x_165_cast_fp16, y = layers_9_self_attn_k_proj_output_scale_to_fp16)[name = tensor("key_19_cast_fp16")]; + tensor var_1582 = const()[name = tensor("op_1582"), val = tensor([1, 1])]; + tensor var_1584 = const()[name = tensor("op_1584"), val = tensor([1, 1])]; + tensor x_167_pad_type_0 = const()[name = tensor("x_167_pad_type_0"), val = tensor("custom")]; + tensor x_167_pad_0 = const()[name = tensor("x_167_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_9_self_attn_v_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(98604032))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(99423296))), name = tensor("layers_9_self_attn_v_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_9_self_attn_v_proj_module_bias_to_fp16 = const()[name = tensor("layers_9_self_attn_v_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(99423424)))]; + tensor x_167_cast_fp16 = conv(bias = layers_9_self_attn_v_proj_module_bias_to_fp16, dilations = var_1584, groups = var_1521, pad = x_167_pad_0, pad_type = x_167_pad_type_0, strides = var_1582, weight = layers_9_self_attn_v_proj_module_weight_to_fp16_palettized, x = input_127_cast_fp16)[name = tensor("x_167_cast_fp16")]; + tensor layers_9_self_attn_v_proj_output_scale_to_fp16 = const()[name = tensor("layers_9_self_attn_v_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(99426048)))]; + tensor value_19_cast_fp16 = mul(x = x_167_cast_fp16, y = layers_9_self_attn_v_proj_output_scale_to_fp16)[name = tensor("value_19_cast_fp16")]; + tensor var_1589 = const()[name = tensor("op_1589"), val = tensor([1, 20, 64, -1])]; + tensor var_1590_cast_fp16 = reshape(shape = var_1589, x = query_19_cast_fp16)[name = tensor("op_1590_cast_fp16")]; + tensor var_1591_to_fp16 = const()[name = tensor("op_1591_to_fp16"), val = tensor(0x1p-3)]; + tensor var_1592_cast_fp16 = mul(x = var_1590_cast_fp16, y = var_1591_to_fp16)[name = tensor("op_1592_cast_fp16")]; + tensor var_1593 = const()[name = tensor("op_1593"), val = tensor([1, 20, 64, -1])]; + tensor var_1594_cast_fp16 = reshape(shape = var_1593, x = key_19_cast_fp16)[name = tensor("op_1594_cast_fp16")]; + tensor mh_w_19_transpose_x_0 = const()[name = tensor("mh_w_19_transpose_x_0"), val = tensor(true)]; + tensor mh_w_19_transpose_y_0 = const()[name = tensor("mh_w_19_transpose_y_0"), val = tensor(false)]; + tensor mh_w_19_cast_fp16 = matmul(transpose_x = mh_w_19_transpose_x_0, transpose_y = mh_w_19_transpose_y_0, x = var_1592_cast_fp16, y = var_1594_cast_fp16)[name = tensor("mh_w_19_cast_fp16")]; + tensor var_1597_cast_fp16 = softmax(axis = var_1519, x = mh_w_19_cast_fp16)[name = tensor("op_1597_cast_fp16")]; + tensor var_1598 = const()[name = tensor("op_1598"), val = tensor([1, 20, 64, -1])]; + tensor var_1599_cast_fp16 = reshape(shape = var_1598, x = value_19_cast_fp16)[name = tensor("op_1599_cast_fp16")]; + tensor attn_19_transpose_x_0 = const()[name = tensor("attn_19_transpose_x_0"), val = tensor(false)]; + tensor attn_19_transpose_y_0 = const()[name = tensor("attn_19_transpose_y_0"), val = tensor(true)]; + tensor attn_19_cast_fp16 = matmul(transpose_x = attn_19_transpose_x_0, transpose_y = attn_19_transpose_y_0, x = var_1599_cast_fp16, y = var_1597_cast_fp16)[name = tensor("attn_19_cast_fp16")]; + tensor var_1602 = const()[name = tensor("op_1602"), val = tensor([1, 1280, 1, -1])]; + tensor x_169_cast_fp16 = reshape(shape = var_1602, x = attn_19_cast_fp16)[name = tensor("x_169_cast_fp16")]; + tensor layers_9_self_attn_o_proj_input_shift_to_fp16 = const()[name = tensor("layers_9_self_attn_o_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(99428672)))]; + tensor input_133_cast_fp16 = sub(x = x_169_cast_fp16, y = layers_9_self_attn_o_proj_input_shift_to_fp16)[name = tensor("input_133_cast_fp16")]; + tensor var_1610 = const()[name = tensor("op_1610"), val = tensor([1, 1])]; + tensor var_1612 = const()[name = tensor("op_1612"), val = tensor([1, 1])]; + tensor x_171_pad_type_0 = const()[name = tensor("x_171_pad_type_0"), val = tensor("custom")]; + tensor x_171_pad_0 = const()[name = tensor("x_171_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_9_self_attn_o_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(99431296))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(100250560))), name = tensor("layers_9_self_attn_o_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_9_self_attn_o_proj_module_bias_to_fp16 = const()[name = tensor("layers_9_self_attn_o_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(100250688)))]; + tensor x_171_cast_fp16 = conv(bias = layers_9_self_attn_o_proj_module_bias_to_fp16, dilations = var_1612, groups = var_1521, pad = x_171_pad_0, pad_type = x_171_pad_type_0, strides = var_1610, weight = layers_9_self_attn_o_proj_module_weight_to_fp16_palettized, x = input_133_cast_fp16)[name = tensor("x_171_cast_fp16")]; + tensor layers_9_self_attn_o_proj_output_scale_to_fp16 = const()[name = tensor("layers_9_self_attn_o_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(100253312)))]; + tensor obj_39_cast_fp16 = mul(x = x_171_cast_fp16, y = layers_9_self_attn_o_proj_output_scale_to_fp16)[name = tensor("obj_39_cast_fp16")]; + tensor inputs_39_cast_fp16 = add(x = inputs_37_cast_fp16, y = obj_39_cast_fp16)[name = tensor("inputs_39_cast_fp16")]; + tensor var_1619 = const()[name = tensor("op_1619"), val = tensor([1])]; + tensor channels_mean_39_cast_fp16 = reduce_mean(axes = var_1619, keep_dims = var_1522, x = inputs_39_cast_fp16)[name = tensor("channels_mean_39_cast_fp16")]; + tensor zero_mean_39_cast_fp16 = sub(x = inputs_39_cast_fp16, y = channels_mean_39_cast_fp16)[name = tensor("zero_mean_39_cast_fp16")]; + tensor zero_mean_sq_39_cast_fp16 = mul(x = zero_mean_39_cast_fp16, y = zero_mean_39_cast_fp16)[name = tensor("zero_mean_sq_39_cast_fp16")]; + tensor var_1623 = const()[name = tensor("op_1623"), val = tensor([1])]; + tensor var_1624_cast_fp16 = reduce_mean(axes = var_1623, keep_dims = var_1522, x = zero_mean_sq_39_cast_fp16)[name = tensor("op_1624_cast_fp16")]; + tensor var_1625_to_fp16 = const()[name = tensor("op_1625_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_1626_cast_fp16 = add(x = var_1624_cast_fp16, y = var_1625_to_fp16)[name = tensor("op_1626_cast_fp16")]; + tensor denom_39_epsilon_0_to_fp16 = const()[name = tensor("denom_39_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_39_cast_fp16 = rsqrt(epsilon = denom_39_epsilon_0_to_fp16, x = var_1626_cast_fp16)[name = tensor("denom_39_cast_fp16")]; + tensor out_39_cast_fp16 = mul(x = zero_mean_39_cast_fp16, y = denom_39_cast_fp16)[name = tensor("out_39_cast_fp16")]; + tensor x_173_gamma_0_to_fp16 = const()[name = tensor("x_173_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(100255936)))]; + tensor x_173_beta_0_to_fp16 = const()[name = tensor("x_173_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(100258560)))]; + tensor x_173_epsilon_0_to_fp16 = const()[name = tensor("x_173_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_173_cast_fp16 = batch_norm(beta = x_173_beta_0_to_fp16, epsilon = x_173_epsilon_0_to_fp16, gamma = x_173_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_39_cast_fp16)[name = tensor("x_173_cast_fp16")]; + tensor layers_9_fc1_input_shift_to_fp16 = const()[name = tensor("layers_9_fc1_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(100261184)))]; + tensor input_135_cast_fp16 = sub(x = x_173_cast_fp16, y = layers_9_fc1_input_shift_to_fp16)[name = tensor("input_135_cast_fp16")]; + tensor var_1641 = const()[name = tensor("op_1641"), val = tensor([1, 1])]; + tensor var_1643 = const()[name = tensor("op_1643"), val = tensor([1, 1])]; + tensor x_175_pad_type_0 = const()[name = tensor("x_175_pad_type_0"), val = tensor("custom")]; + tensor x_175_pad_0 = const()[name = tensor("x_175_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_9_fc1_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(100263808))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(103540672))), name = tensor("layers_9_fc1_module_weight_to_fp16_palettized"), shape = tensor([5120, 1280, 1, 1])]; + tensor layers_9_fc1_module_bias_to_fp16 = const()[name = tensor("layers_9_fc1_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(103540800)))]; + tensor x_175_cast_fp16 = conv(bias = layers_9_fc1_module_bias_to_fp16, dilations = var_1643, groups = var_1521, pad = x_175_pad_0, pad_type = x_175_pad_type_0, strides = var_1641, weight = layers_9_fc1_module_weight_to_fp16_palettized, x = input_135_cast_fp16)[name = tensor("x_175_cast_fp16")]; + tensor layers_9_fc1_output_scale_to_fp16 = const()[name = tensor("layers_9_fc1_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(103551104)))]; + tensor input_137_cast_fp16 = mul(x = x_175_cast_fp16, y = layers_9_fc1_output_scale_to_fp16)[name = tensor("input_137_cast_fp16")]; + tensor x_177_mode_0 = const()[name = tensor("x_177_mode_0"), val = tensor("EXACT")]; + tensor x_177_cast_fp16 = gelu(mode = x_177_mode_0, x = input_137_cast_fp16)[name = tensor("x_177_cast_fp16")]; + tensor layers_9_fc2_input_shift_to_fp16 = const()[name = tensor("layers_9_fc2_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(103561408)))]; + tensor input_139_cast_fp16 = sub(x = x_177_cast_fp16, y = layers_9_fc2_input_shift_to_fp16)[name = tensor("input_139_cast_fp16")]; + tensor var_1654 = const()[name = tensor("op_1654"), val = tensor([1, 1])]; + tensor var_1656 = const()[name = tensor("op_1656"), val = tensor([1, 1])]; + tensor x_179_pad_type_0 = const()[name = tensor("x_179_pad_type_0"), val = tensor("custom")]; + tensor x_179_pad_0 = const()[name = tensor("x_179_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_9_fc2_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(103571712))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(106848576))), name = tensor("layers_9_fc2_module_weight_to_fp16_palettized"), shape = tensor([1280, 5120, 1, 1])]; + tensor layers_9_fc2_module_bias_to_fp16 = const()[name = tensor("layers_9_fc2_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(106848704)))]; + tensor x_179_cast_fp16 = conv(bias = layers_9_fc2_module_bias_to_fp16, dilations = var_1656, groups = var_1521, pad = x_179_pad_0, pad_type = x_179_pad_type_0, strides = var_1654, weight = layers_9_fc2_module_weight_to_fp16_palettized, x = input_139_cast_fp16)[name = tensor("x_179_cast_fp16")]; + tensor layers_9_fc2_output_scale_to_fp16 = const()[name = tensor("layers_9_fc2_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(106851328)))]; + tensor hidden_states_23_cast_fp16 = mul(x = x_179_cast_fp16, y = layers_9_fc2_output_scale_to_fp16)[name = tensor("hidden_states_23_cast_fp16")]; + tensor inputs_41_cast_fp16 = add(x = inputs_39_cast_fp16, y = hidden_states_23_cast_fp16)[name = tensor("inputs_41_cast_fp16")]; + tensor var_1668 = const()[name = tensor("op_1668"), val = tensor(3)]; + tensor var_1670 = const()[name = tensor("op_1670"), val = tensor(1)]; + tensor var_1671 = const()[name = tensor("op_1671"), val = tensor(true)]; + tensor var_1681 = const()[name = tensor("op_1681"), val = tensor([1])]; + tensor channels_mean_41_cast_fp16 = reduce_mean(axes = var_1681, keep_dims = var_1671, x = inputs_41_cast_fp16)[name = tensor("channels_mean_41_cast_fp16")]; + tensor zero_mean_41_cast_fp16 = sub(x = inputs_41_cast_fp16, y = channels_mean_41_cast_fp16)[name = tensor("zero_mean_41_cast_fp16")]; + tensor zero_mean_sq_41_cast_fp16 = mul(x = zero_mean_41_cast_fp16, y = zero_mean_41_cast_fp16)[name = tensor("zero_mean_sq_41_cast_fp16")]; + tensor var_1685 = const()[name = tensor("op_1685"), val = tensor([1])]; + tensor var_1686_cast_fp16 = reduce_mean(axes = var_1685, keep_dims = var_1671, x = zero_mean_sq_41_cast_fp16)[name = tensor("op_1686_cast_fp16")]; + tensor var_1687_to_fp16 = const()[name = tensor("op_1687_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_1688_cast_fp16 = add(x = var_1686_cast_fp16, y = var_1687_to_fp16)[name = tensor("op_1688_cast_fp16")]; + tensor denom_41_epsilon_0_to_fp16 = const()[name = tensor("denom_41_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_41_cast_fp16 = rsqrt(epsilon = denom_41_epsilon_0_to_fp16, x = var_1688_cast_fp16)[name = tensor("denom_41_cast_fp16")]; + tensor out_41_cast_fp16 = mul(x = zero_mean_41_cast_fp16, y = denom_41_cast_fp16)[name = tensor("out_41_cast_fp16")]; + tensor obj_41_gamma_0_to_fp16 = const()[name = tensor("obj_41_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(106853952)))]; + tensor obj_41_beta_0_to_fp16 = const()[name = tensor("obj_41_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(106856576)))]; + tensor obj_41_epsilon_0_to_fp16 = const()[name = tensor("obj_41_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_41_cast_fp16 = batch_norm(beta = obj_41_beta_0_to_fp16, epsilon = obj_41_epsilon_0_to_fp16, gamma = obj_41_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_41_cast_fp16)[name = tensor("obj_41_cast_fp16")]; + tensor layers_10_self_attn_q_proj_input_shift_to_fp16 = const()[name = tensor("layers_10_self_attn_q_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(106859200)))]; + tensor input_141_cast_fp16 = sub(x = obj_41_cast_fp16, y = layers_10_self_attn_q_proj_input_shift_to_fp16)[name = tensor("input_141_cast_fp16")]; + tensor var_1707 = const()[name = tensor("op_1707"), val = tensor([1, 1])]; + tensor var_1709 = const()[name = tensor("op_1709"), val = tensor([1, 1])]; + tensor x_181_pad_type_0 = const()[name = tensor("x_181_pad_type_0"), val = tensor("custom")]; + tensor x_181_pad_0 = const()[name = tensor("x_181_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_10_self_attn_q_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(106861824))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(107681088))), name = tensor("layers_10_self_attn_q_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_10_self_attn_q_proj_module_bias_to_fp16 = const()[name = tensor("layers_10_self_attn_q_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(107681216)))]; + tensor x_181_cast_fp16 = conv(bias = layers_10_self_attn_q_proj_module_bias_to_fp16, dilations = var_1709, groups = var_1670, pad = x_181_pad_0, pad_type = x_181_pad_type_0, strides = var_1707, weight = layers_10_self_attn_q_proj_module_weight_to_fp16_palettized, x = input_141_cast_fp16)[name = tensor("x_181_cast_fp16")]; + tensor layers_10_self_attn_q_proj_output_scale_to_fp16 = const()[name = tensor("layers_10_self_attn_q_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(107683840)))]; + tensor query_21_cast_fp16 = mul(x = x_181_cast_fp16, y = layers_10_self_attn_q_proj_output_scale_to_fp16)[name = tensor("query_21_cast_fp16")]; + tensor var_1719 = const()[name = tensor("op_1719"), val = tensor([1, 1])]; + tensor var_1721 = const()[name = tensor("op_1721"), val = tensor([1, 1])]; + tensor x_183_pad_type_0 = const()[name = tensor("x_183_pad_type_0"), val = tensor("custom")]; + tensor x_183_pad_0 = const()[name = tensor("x_183_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_10_self_attn_k_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(107686464))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(108505728))), name = tensor("layers_10_self_attn_k_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_10_self_attn_k_proj_module_bias_to_fp16 = const()[name = tensor("layers_10_self_attn_k_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(108505856)))]; + tensor x_183_cast_fp16 = conv(bias = layers_10_self_attn_k_proj_module_bias_to_fp16, dilations = var_1721, groups = var_1670, pad = x_183_pad_0, pad_type = x_183_pad_type_0, strides = var_1719, weight = layers_10_self_attn_k_proj_module_weight_to_fp16_palettized, x = input_141_cast_fp16)[name = tensor("x_183_cast_fp16")]; + tensor layers_10_self_attn_k_proj_output_scale_to_fp16 = const()[name = tensor("layers_10_self_attn_k_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(108508480)))]; + tensor key_21_cast_fp16 = mul(x = x_183_cast_fp16, y = layers_10_self_attn_k_proj_output_scale_to_fp16)[name = tensor("key_21_cast_fp16")]; + tensor var_1731 = const()[name = tensor("op_1731"), val = tensor([1, 1])]; + tensor var_1733 = const()[name = tensor("op_1733"), val = tensor([1, 1])]; + tensor x_185_pad_type_0 = const()[name = tensor("x_185_pad_type_0"), val = tensor("custom")]; + tensor x_185_pad_0 = const()[name = tensor("x_185_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_10_self_attn_v_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(108511104))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(109330368))), name = tensor("layers_10_self_attn_v_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_10_self_attn_v_proj_module_bias_to_fp16 = const()[name = tensor("layers_10_self_attn_v_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(109330496)))]; + tensor x_185_cast_fp16 = conv(bias = layers_10_self_attn_v_proj_module_bias_to_fp16, dilations = var_1733, groups = var_1670, pad = x_185_pad_0, pad_type = x_185_pad_type_0, strides = var_1731, weight = layers_10_self_attn_v_proj_module_weight_to_fp16_palettized, x = input_141_cast_fp16)[name = tensor("x_185_cast_fp16")]; + tensor layers_10_self_attn_v_proj_output_scale_to_fp16 = const()[name = tensor("layers_10_self_attn_v_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(109333120)))]; + tensor value_21_cast_fp16 = mul(x = x_185_cast_fp16, y = layers_10_self_attn_v_proj_output_scale_to_fp16)[name = tensor("value_21_cast_fp16")]; + tensor var_1738 = const()[name = tensor("op_1738"), val = tensor([1, 20, 64, -1])]; + tensor var_1739_cast_fp16 = reshape(shape = var_1738, x = query_21_cast_fp16)[name = tensor("op_1739_cast_fp16")]; + tensor var_1740_to_fp16 = const()[name = tensor("op_1740_to_fp16"), val = tensor(0x1p-3)]; + tensor var_1741_cast_fp16 = mul(x = var_1739_cast_fp16, y = var_1740_to_fp16)[name = tensor("op_1741_cast_fp16")]; + tensor var_1742 = const()[name = tensor("op_1742"), val = tensor([1, 20, 64, -1])]; + tensor var_1743_cast_fp16 = reshape(shape = var_1742, x = key_21_cast_fp16)[name = tensor("op_1743_cast_fp16")]; + tensor mh_w_21_transpose_x_0 = const()[name = tensor("mh_w_21_transpose_x_0"), val = tensor(true)]; + tensor mh_w_21_transpose_y_0 = const()[name = tensor("mh_w_21_transpose_y_0"), val = tensor(false)]; + tensor mh_w_21_cast_fp16 = matmul(transpose_x = mh_w_21_transpose_x_0, transpose_y = mh_w_21_transpose_y_0, x = var_1741_cast_fp16, y = var_1743_cast_fp16)[name = tensor("mh_w_21_cast_fp16")]; + tensor var_1746_cast_fp16 = softmax(axis = var_1668, x = mh_w_21_cast_fp16)[name = tensor("op_1746_cast_fp16")]; + tensor var_1747 = const()[name = tensor("op_1747"), val = tensor([1, 20, 64, -1])]; + tensor var_1748_cast_fp16 = reshape(shape = var_1747, x = value_21_cast_fp16)[name = tensor("op_1748_cast_fp16")]; + tensor attn_21_transpose_x_0 = const()[name = tensor("attn_21_transpose_x_0"), val = tensor(false)]; + tensor attn_21_transpose_y_0 = const()[name = tensor("attn_21_transpose_y_0"), val = tensor(true)]; + tensor attn_21_cast_fp16 = matmul(transpose_x = attn_21_transpose_x_0, transpose_y = attn_21_transpose_y_0, x = var_1748_cast_fp16, y = var_1746_cast_fp16)[name = tensor("attn_21_cast_fp16")]; + tensor var_1751 = const()[name = tensor("op_1751"), val = tensor([1, 1280, 1, -1])]; + tensor x_187_cast_fp16 = reshape(shape = var_1751, x = attn_21_cast_fp16)[name = tensor("x_187_cast_fp16")]; + tensor layers_10_self_attn_o_proj_input_shift_to_fp16 = const()[name = tensor("layers_10_self_attn_o_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(109335744)))]; + tensor input_147_cast_fp16 = sub(x = x_187_cast_fp16, y = layers_10_self_attn_o_proj_input_shift_to_fp16)[name = tensor("input_147_cast_fp16")]; + tensor var_1759 = const()[name = tensor("op_1759"), val = tensor([1, 1])]; + tensor var_1761 = const()[name = tensor("op_1761"), val = tensor([1, 1])]; + tensor x_189_pad_type_0 = const()[name = tensor("x_189_pad_type_0"), val = tensor("custom")]; + tensor x_189_pad_0 = const()[name = tensor("x_189_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_10_self_attn_o_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(109338368))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(110157632))), name = tensor("layers_10_self_attn_o_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_10_self_attn_o_proj_module_bias_to_fp16 = const()[name = tensor("layers_10_self_attn_o_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(110157760)))]; + tensor x_189_cast_fp16 = conv(bias = layers_10_self_attn_o_proj_module_bias_to_fp16, dilations = var_1761, groups = var_1670, pad = x_189_pad_0, pad_type = x_189_pad_type_0, strides = var_1759, weight = layers_10_self_attn_o_proj_module_weight_to_fp16_palettized, x = input_147_cast_fp16)[name = tensor("x_189_cast_fp16")]; + tensor layers_10_self_attn_o_proj_output_scale_to_fp16 = const()[name = tensor("layers_10_self_attn_o_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(110160384)))]; + tensor obj_43_cast_fp16 = mul(x = x_189_cast_fp16, y = layers_10_self_attn_o_proj_output_scale_to_fp16)[name = tensor("obj_43_cast_fp16")]; + tensor inputs_43_cast_fp16 = add(x = inputs_41_cast_fp16, y = obj_43_cast_fp16)[name = tensor("inputs_43_cast_fp16")]; + tensor var_1768 = const()[name = tensor("op_1768"), val = tensor([1])]; + tensor channels_mean_43_cast_fp16 = reduce_mean(axes = var_1768, keep_dims = var_1671, x = inputs_43_cast_fp16)[name = tensor("channels_mean_43_cast_fp16")]; + tensor zero_mean_43_cast_fp16 = sub(x = inputs_43_cast_fp16, y = channels_mean_43_cast_fp16)[name = tensor("zero_mean_43_cast_fp16")]; + tensor zero_mean_sq_43_cast_fp16 = mul(x = zero_mean_43_cast_fp16, y = zero_mean_43_cast_fp16)[name = tensor("zero_mean_sq_43_cast_fp16")]; + tensor var_1772 = const()[name = tensor("op_1772"), val = tensor([1])]; + tensor var_1773_cast_fp16 = reduce_mean(axes = var_1772, keep_dims = var_1671, x = zero_mean_sq_43_cast_fp16)[name = tensor("op_1773_cast_fp16")]; + tensor var_1774_to_fp16 = const()[name = tensor("op_1774_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_1775_cast_fp16 = add(x = var_1773_cast_fp16, y = var_1774_to_fp16)[name = tensor("op_1775_cast_fp16")]; + tensor denom_43_epsilon_0_to_fp16 = const()[name = tensor("denom_43_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_43_cast_fp16 = rsqrt(epsilon = denom_43_epsilon_0_to_fp16, x = var_1775_cast_fp16)[name = tensor("denom_43_cast_fp16")]; + tensor out_43_cast_fp16 = mul(x = zero_mean_43_cast_fp16, y = denom_43_cast_fp16)[name = tensor("out_43_cast_fp16")]; + tensor x_191_gamma_0_to_fp16 = const()[name = tensor("x_191_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(110163008)))]; + tensor x_191_beta_0_to_fp16 = const()[name = tensor("x_191_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(110165632)))]; + tensor x_191_epsilon_0_to_fp16 = const()[name = tensor("x_191_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_191_cast_fp16 = batch_norm(beta = x_191_beta_0_to_fp16, epsilon = x_191_epsilon_0_to_fp16, gamma = x_191_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_43_cast_fp16)[name = tensor("x_191_cast_fp16")]; + tensor layers_10_fc1_input_shift_to_fp16 = const()[name = tensor("layers_10_fc1_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(110168256)))]; + tensor input_149_cast_fp16 = sub(x = x_191_cast_fp16, y = layers_10_fc1_input_shift_to_fp16)[name = tensor("input_149_cast_fp16")]; + tensor var_1790 = const()[name = tensor("op_1790"), val = tensor([1, 1])]; + tensor var_1792 = const()[name = tensor("op_1792"), val = tensor([1, 1])]; + tensor x_193_pad_type_0 = const()[name = tensor("x_193_pad_type_0"), val = tensor("custom")]; + tensor x_193_pad_0 = const()[name = tensor("x_193_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_10_fc1_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(110170880))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(113447744))), name = tensor("layers_10_fc1_module_weight_to_fp16_palettized"), shape = tensor([5120, 1280, 1, 1])]; + tensor layers_10_fc1_module_bias_to_fp16 = const()[name = tensor("layers_10_fc1_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(113447872)))]; + tensor x_193_cast_fp16 = conv(bias = layers_10_fc1_module_bias_to_fp16, dilations = var_1792, groups = var_1670, pad = x_193_pad_0, pad_type = x_193_pad_type_0, strides = var_1790, weight = layers_10_fc1_module_weight_to_fp16_palettized, x = input_149_cast_fp16)[name = tensor("x_193_cast_fp16")]; + tensor layers_10_fc1_output_scale_to_fp16 = const()[name = tensor("layers_10_fc1_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(113458176)))]; + tensor input_151_cast_fp16 = mul(x = x_193_cast_fp16, y = layers_10_fc1_output_scale_to_fp16)[name = tensor("input_151_cast_fp16")]; + tensor x_195_mode_0 = const()[name = tensor("x_195_mode_0"), val = tensor("EXACT")]; + tensor x_195_cast_fp16 = gelu(mode = x_195_mode_0, x = input_151_cast_fp16)[name = tensor("x_195_cast_fp16")]; + tensor layers_10_fc2_input_shift_to_fp16 = const()[name = tensor("layers_10_fc2_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(113468480)))]; + tensor input_153_cast_fp16 = sub(x = x_195_cast_fp16, y = layers_10_fc2_input_shift_to_fp16)[name = tensor("input_153_cast_fp16")]; + tensor var_1803 = const()[name = tensor("op_1803"), val = tensor([1, 1])]; + tensor var_1805 = const()[name = tensor("op_1805"), val = tensor([1, 1])]; + tensor x_197_pad_type_0 = const()[name = tensor("x_197_pad_type_0"), val = tensor("custom")]; + tensor x_197_pad_0 = const()[name = tensor("x_197_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_10_fc2_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(113478784))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(116755648))), name = tensor("layers_10_fc2_module_weight_to_fp16_palettized"), shape = tensor([1280, 5120, 1, 1])]; + tensor layers_10_fc2_module_bias_to_fp16 = const()[name = tensor("layers_10_fc2_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(116755776)))]; + tensor x_197_cast_fp16 = conv(bias = layers_10_fc2_module_bias_to_fp16, dilations = var_1805, groups = var_1670, pad = x_197_pad_0, pad_type = x_197_pad_type_0, strides = var_1803, weight = layers_10_fc2_module_weight_to_fp16_palettized, x = input_153_cast_fp16)[name = tensor("x_197_cast_fp16")]; + tensor layers_10_fc2_output_scale_to_fp16 = const()[name = tensor("layers_10_fc2_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(116758400)))]; + tensor hidden_states_25_cast_fp16 = mul(x = x_197_cast_fp16, y = layers_10_fc2_output_scale_to_fp16)[name = tensor("hidden_states_25_cast_fp16")]; + tensor inputs_45_cast_fp16 = add(x = inputs_43_cast_fp16, y = hidden_states_25_cast_fp16)[name = tensor("inputs_45_cast_fp16")]; + tensor var_1817 = const()[name = tensor("op_1817"), val = tensor(3)]; + tensor var_1819 = const()[name = tensor("op_1819"), val = tensor(1)]; + tensor var_1820 = const()[name = tensor("op_1820"), val = tensor(true)]; + tensor var_1830 = const()[name = tensor("op_1830"), val = tensor([1])]; + tensor channels_mean_45_cast_fp16 = reduce_mean(axes = var_1830, keep_dims = var_1820, x = inputs_45_cast_fp16)[name = tensor("channels_mean_45_cast_fp16")]; + tensor zero_mean_45_cast_fp16 = sub(x = inputs_45_cast_fp16, y = channels_mean_45_cast_fp16)[name = tensor("zero_mean_45_cast_fp16")]; + tensor zero_mean_sq_45_cast_fp16 = mul(x = zero_mean_45_cast_fp16, y = zero_mean_45_cast_fp16)[name = tensor("zero_mean_sq_45_cast_fp16")]; + tensor var_1834 = const()[name = tensor("op_1834"), val = tensor([1])]; + tensor var_1835_cast_fp16 = reduce_mean(axes = var_1834, keep_dims = var_1820, x = zero_mean_sq_45_cast_fp16)[name = tensor("op_1835_cast_fp16")]; + tensor var_1836_to_fp16 = const()[name = tensor("op_1836_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_1837_cast_fp16 = add(x = var_1835_cast_fp16, y = var_1836_to_fp16)[name = tensor("op_1837_cast_fp16")]; + tensor denom_45_epsilon_0_to_fp16 = const()[name = tensor("denom_45_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_45_cast_fp16 = rsqrt(epsilon = denom_45_epsilon_0_to_fp16, x = var_1837_cast_fp16)[name = tensor("denom_45_cast_fp16")]; + tensor out_45_cast_fp16 = mul(x = zero_mean_45_cast_fp16, y = denom_45_cast_fp16)[name = tensor("out_45_cast_fp16")]; + tensor obj_45_gamma_0_to_fp16 = const()[name = tensor("obj_45_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(116761024)))]; + tensor obj_45_beta_0_to_fp16 = const()[name = tensor("obj_45_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(116763648)))]; + tensor obj_45_epsilon_0_to_fp16 = const()[name = tensor("obj_45_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_45_cast_fp16 = batch_norm(beta = obj_45_beta_0_to_fp16, epsilon = obj_45_epsilon_0_to_fp16, gamma = obj_45_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_45_cast_fp16)[name = tensor("obj_45_cast_fp16")]; + tensor layers_11_self_attn_q_proj_input_shift_to_fp16 = const()[name = tensor("layers_11_self_attn_q_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(116766272)))]; + tensor input_155_cast_fp16 = sub(x = obj_45_cast_fp16, y = layers_11_self_attn_q_proj_input_shift_to_fp16)[name = tensor("input_155_cast_fp16")]; + tensor var_1856 = const()[name = tensor("op_1856"), val = tensor([1, 1])]; + tensor var_1858 = const()[name = tensor("op_1858"), val = tensor([1, 1])]; + tensor x_199_pad_type_0 = const()[name = tensor("x_199_pad_type_0"), val = tensor("custom")]; + tensor x_199_pad_0 = const()[name = tensor("x_199_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_11_self_attn_q_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(116768896))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(117588160))), name = tensor("layers_11_self_attn_q_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_11_self_attn_q_proj_module_bias_to_fp16 = const()[name = tensor("layers_11_self_attn_q_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(117588288)))]; + tensor x_199_cast_fp16 = conv(bias = layers_11_self_attn_q_proj_module_bias_to_fp16, dilations = var_1858, groups = var_1819, pad = x_199_pad_0, pad_type = x_199_pad_type_0, strides = var_1856, weight = layers_11_self_attn_q_proj_module_weight_to_fp16_palettized, x = input_155_cast_fp16)[name = tensor("x_199_cast_fp16")]; + tensor layers_11_self_attn_q_proj_output_scale_to_fp16 = const()[name = tensor("layers_11_self_attn_q_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(117590912)))]; + tensor query_23_cast_fp16 = mul(x = x_199_cast_fp16, y = layers_11_self_attn_q_proj_output_scale_to_fp16)[name = tensor("query_23_cast_fp16")]; + tensor var_1868 = const()[name = tensor("op_1868"), val = tensor([1, 1])]; + tensor var_1870 = const()[name = tensor("op_1870"), val = tensor([1, 1])]; + tensor x_201_pad_type_0 = const()[name = tensor("x_201_pad_type_0"), val = tensor("custom")]; + tensor x_201_pad_0 = const()[name = tensor("x_201_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_11_self_attn_k_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(117593536))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(118412800))), name = tensor("layers_11_self_attn_k_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_11_self_attn_k_proj_module_bias_to_fp16 = const()[name = tensor("layers_11_self_attn_k_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(118412928)))]; + tensor x_201_cast_fp16 = conv(bias = layers_11_self_attn_k_proj_module_bias_to_fp16, dilations = var_1870, groups = var_1819, pad = x_201_pad_0, pad_type = x_201_pad_type_0, strides = var_1868, weight = layers_11_self_attn_k_proj_module_weight_to_fp16_palettized, x = input_155_cast_fp16)[name = tensor("x_201_cast_fp16")]; + tensor layers_11_self_attn_k_proj_output_scale_to_fp16 = const()[name = tensor("layers_11_self_attn_k_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(118415552)))]; + tensor key_23_cast_fp16 = mul(x = x_201_cast_fp16, y = layers_11_self_attn_k_proj_output_scale_to_fp16)[name = tensor("key_23_cast_fp16")]; + tensor var_1880 = const()[name = tensor("op_1880"), val = tensor([1, 1])]; + tensor var_1882 = const()[name = tensor("op_1882"), val = tensor([1, 1])]; + tensor x_203_pad_type_0 = const()[name = tensor("x_203_pad_type_0"), val = tensor("custom")]; + tensor x_203_pad_0 = const()[name = tensor("x_203_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_11_self_attn_v_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(118418176))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(119237440))), name = tensor("layers_11_self_attn_v_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_11_self_attn_v_proj_module_bias_to_fp16 = const()[name = tensor("layers_11_self_attn_v_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(119237568)))]; + tensor x_203_cast_fp16 = conv(bias = layers_11_self_attn_v_proj_module_bias_to_fp16, dilations = var_1882, groups = var_1819, pad = x_203_pad_0, pad_type = x_203_pad_type_0, strides = var_1880, weight = layers_11_self_attn_v_proj_module_weight_to_fp16_palettized, x = input_155_cast_fp16)[name = tensor("x_203_cast_fp16")]; + tensor layers_11_self_attn_v_proj_output_scale_to_fp16 = const()[name = tensor("layers_11_self_attn_v_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(119240192)))]; + tensor value_23_cast_fp16 = mul(x = x_203_cast_fp16, y = layers_11_self_attn_v_proj_output_scale_to_fp16)[name = tensor("value_23_cast_fp16")]; + tensor var_1887 = const()[name = tensor("op_1887"), val = tensor([1, 20, 64, -1])]; + tensor var_1888_cast_fp16 = reshape(shape = var_1887, x = query_23_cast_fp16)[name = tensor("op_1888_cast_fp16")]; + tensor var_1889_to_fp16 = const()[name = tensor("op_1889_to_fp16"), val = tensor(0x1p-3)]; + tensor var_1890_cast_fp16 = mul(x = var_1888_cast_fp16, y = var_1889_to_fp16)[name = tensor("op_1890_cast_fp16")]; + tensor var_1891 = const()[name = tensor("op_1891"), val = tensor([1, 20, 64, -1])]; + tensor var_1892_cast_fp16 = reshape(shape = var_1891, x = key_23_cast_fp16)[name = tensor("op_1892_cast_fp16")]; + tensor mh_w_23_transpose_x_0 = const()[name = tensor("mh_w_23_transpose_x_0"), val = tensor(true)]; + tensor mh_w_23_transpose_y_0 = const()[name = tensor("mh_w_23_transpose_y_0"), val = tensor(false)]; + tensor mh_w_23_cast_fp16 = matmul(transpose_x = mh_w_23_transpose_x_0, transpose_y = mh_w_23_transpose_y_0, x = var_1890_cast_fp16, y = var_1892_cast_fp16)[name = tensor("mh_w_23_cast_fp16")]; + tensor var_1895_cast_fp16 = softmax(axis = var_1817, x = mh_w_23_cast_fp16)[name = tensor("op_1895_cast_fp16")]; + tensor var_1896 = const()[name = tensor("op_1896"), val = tensor([1, 20, 64, -1])]; + tensor var_1897_cast_fp16 = reshape(shape = var_1896, x = value_23_cast_fp16)[name = tensor("op_1897_cast_fp16")]; + tensor attn_23_transpose_x_0 = const()[name = tensor("attn_23_transpose_x_0"), val = tensor(false)]; + tensor attn_23_transpose_y_0 = const()[name = tensor("attn_23_transpose_y_0"), val = tensor(true)]; + tensor attn_23_cast_fp16 = matmul(transpose_x = attn_23_transpose_x_0, transpose_y = attn_23_transpose_y_0, x = var_1897_cast_fp16, y = var_1895_cast_fp16)[name = tensor("attn_23_cast_fp16")]; + tensor var_1900 = const()[name = tensor("op_1900"), val = tensor([1, 1280, 1, -1])]; + tensor x_205_cast_fp16 = reshape(shape = var_1900, x = attn_23_cast_fp16)[name = tensor("x_205_cast_fp16")]; + tensor layers_11_self_attn_o_proj_input_shift_to_fp16 = const()[name = tensor("layers_11_self_attn_o_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(119242816)))]; + tensor input_161_cast_fp16 = sub(x = x_205_cast_fp16, y = layers_11_self_attn_o_proj_input_shift_to_fp16)[name = tensor("input_161_cast_fp16")]; + tensor var_1908 = const()[name = tensor("op_1908"), val = tensor([1, 1])]; + tensor var_1910 = const()[name = tensor("op_1910"), val = tensor([1, 1])]; + tensor x_207_pad_type_0 = const()[name = tensor("x_207_pad_type_0"), val = tensor("custom")]; + tensor x_207_pad_0 = const()[name = tensor("x_207_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_11_self_attn_o_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(119245440))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(120064704))), name = tensor("layers_11_self_attn_o_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_11_self_attn_o_proj_module_bias_to_fp16 = const()[name = tensor("layers_11_self_attn_o_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(120064832)))]; + tensor x_207_cast_fp16 = conv(bias = layers_11_self_attn_o_proj_module_bias_to_fp16, dilations = var_1910, groups = var_1819, pad = x_207_pad_0, pad_type = x_207_pad_type_0, strides = var_1908, weight = layers_11_self_attn_o_proj_module_weight_to_fp16_palettized, x = input_161_cast_fp16)[name = tensor("x_207_cast_fp16")]; + tensor layers_11_self_attn_o_proj_output_scale_to_fp16 = const()[name = tensor("layers_11_self_attn_o_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(120067456)))]; + tensor obj_47_cast_fp16 = mul(x = x_207_cast_fp16, y = layers_11_self_attn_o_proj_output_scale_to_fp16)[name = tensor("obj_47_cast_fp16")]; + tensor inputs_47_cast_fp16 = add(x = inputs_45_cast_fp16, y = obj_47_cast_fp16)[name = tensor("inputs_47_cast_fp16")]; + tensor var_1917 = const()[name = tensor("op_1917"), val = tensor([1])]; + tensor channels_mean_47_cast_fp16 = reduce_mean(axes = var_1917, keep_dims = var_1820, x = inputs_47_cast_fp16)[name = tensor("channels_mean_47_cast_fp16")]; + tensor zero_mean_47_cast_fp16 = sub(x = inputs_47_cast_fp16, y = channels_mean_47_cast_fp16)[name = tensor("zero_mean_47_cast_fp16")]; + tensor zero_mean_sq_47_cast_fp16 = mul(x = zero_mean_47_cast_fp16, y = zero_mean_47_cast_fp16)[name = tensor("zero_mean_sq_47_cast_fp16")]; + tensor var_1921 = const()[name = tensor("op_1921"), val = tensor([1])]; + tensor var_1922_cast_fp16 = reduce_mean(axes = var_1921, keep_dims = var_1820, x = zero_mean_sq_47_cast_fp16)[name = tensor("op_1922_cast_fp16")]; + tensor var_1923_to_fp16 = const()[name = tensor("op_1923_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_1924_cast_fp16 = add(x = var_1922_cast_fp16, y = var_1923_to_fp16)[name = tensor("op_1924_cast_fp16")]; + tensor denom_47_epsilon_0_to_fp16 = const()[name = tensor("denom_47_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_47_cast_fp16 = rsqrt(epsilon = denom_47_epsilon_0_to_fp16, x = var_1924_cast_fp16)[name = tensor("denom_47_cast_fp16")]; + tensor out_47_cast_fp16 = mul(x = zero_mean_47_cast_fp16, y = denom_47_cast_fp16)[name = tensor("out_47_cast_fp16")]; + tensor x_209_gamma_0_to_fp16 = const()[name = tensor("x_209_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(120070080)))]; + tensor x_209_beta_0_to_fp16 = const()[name = tensor("x_209_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(120072704)))]; + tensor x_209_epsilon_0_to_fp16 = const()[name = tensor("x_209_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_209_cast_fp16 = batch_norm(beta = x_209_beta_0_to_fp16, epsilon = x_209_epsilon_0_to_fp16, gamma = x_209_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_47_cast_fp16)[name = tensor("x_209_cast_fp16")]; + tensor layers_11_fc1_input_shift_to_fp16 = const()[name = tensor("layers_11_fc1_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(120075328)))]; + tensor input_163_cast_fp16 = sub(x = x_209_cast_fp16, y = layers_11_fc1_input_shift_to_fp16)[name = tensor("input_163_cast_fp16")]; + tensor var_1939 = const()[name = tensor("op_1939"), val = tensor([1, 1])]; + tensor var_1941 = const()[name = tensor("op_1941"), val = tensor([1, 1])]; + tensor x_211_pad_type_0 = const()[name = tensor("x_211_pad_type_0"), val = tensor("custom")]; + tensor x_211_pad_0 = const()[name = tensor("x_211_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_11_fc1_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(120077952))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(123354816))), name = tensor("layers_11_fc1_module_weight_to_fp16_palettized"), shape = tensor([5120, 1280, 1, 1])]; + tensor layers_11_fc1_module_bias_to_fp16 = const()[name = tensor("layers_11_fc1_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(123354944)))]; + tensor x_211_cast_fp16 = conv(bias = layers_11_fc1_module_bias_to_fp16, dilations = var_1941, groups = var_1819, pad = x_211_pad_0, pad_type = x_211_pad_type_0, strides = var_1939, weight = layers_11_fc1_module_weight_to_fp16_palettized, x = input_163_cast_fp16)[name = tensor("x_211_cast_fp16")]; + tensor layers_11_fc1_output_scale_to_fp16 = const()[name = tensor("layers_11_fc1_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(123365248)))]; + tensor input_165_cast_fp16 = mul(x = x_211_cast_fp16, y = layers_11_fc1_output_scale_to_fp16)[name = tensor("input_165_cast_fp16")]; + tensor x_213_mode_0 = const()[name = tensor("x_213_mode_0"), val = tensor("EXACT")]; + tensor x_213_cast_fp16 = gelu(mode = x_213_mode_0, x = input_165_cast_fp16)[name = tensor("x_213_cast_fp16")]; + tensor layers_11_fc2_input_shift_to_fp16 = const()[name = tensor("layers_11_fc2_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(123375552)))]; + tensor input_167_cast_fp16 = sub(x = x_213_cast_fp16, y = layers_11_fc2_input_shift_to_fp16)[name = tensor("input_167_cast_fp16")]; + tensor var_1952 = const()[name = tensor("op_1952"), val = tensor([1, 1])]; + tensor var_1954 = const()[name = tensor("op_1954"), val = tensor([1, 1])]; + tensor x_215_pad_type_0 = const()[name = tensor("x_215_pad_type_0"), val = tensor("custom")]; + tensor x_215_pad_0 = const()[name = tensor("x_215_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_11_fc2_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(123385856))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(126662720))), name = tensor("layers_11_fc2_module_weight_to_fp16_palettized"), shape = tensor([1280, 5120, 1, 1])]; + tensor layers_11_fc2_module_bias_to_fp16 = const()[name = tensor("layers_11_fc2_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(126662848)))]; + tensor x_215_cast_fp16 = conv(bias = layers_11_fc2_module_bias_to_fp16, dilations = var_1954, groups = var_1819, pad = x_215_pad_0, pad_type = x_215_pad_type_0, strides = var_1952, weight = layers_11_fc2_module_weight_to_fp16_palettized, x = input_167_cast_fp16)[name = tensor("x_215_cast_fp16")]; + tensor layers_11_fc2_output_scale_to_fp16 = const()[name = tensor("layers_11_fc2_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(126665472)))]; + tensor hidden_states_27_cast_fp16 = mul(x = x_215_cast_fp16, y = layers_11_fc2_output_scale_to_fp16)[name = tensor("hidden_states_27_cast_fp16")]; + tensor inputs_49_cast_fp16 = add(x = inputs_47_cast_fp16, y = hidden_states_27_cast_fp16)[name = tensor("inputs_49_cast_fp16")]; + tensor var_1966 = const()[name = tensor("op_1966"), val = tensor(3)]; + tensor var_1968 = const()[name = tensor("op_1968"), val = tensor(1)]; + tensor var_1969 = const()[name = tensor("op_1969"), val = tensor(true)]; + tensor var_1979 = const()[name = tensor("op_1979"), val = tensor([1])]; + tensor channels_mean_49_cast_fp16 = reduce_mean(axes = var_1979, keep_dims = var_1969, x = inputs_49_cast_fp16)[name = tensor("channels_mean_49_cast_fp16")]; + tensor zero_mean_49_cast_fp16 = sub(x = inputs_49_cast_fp16, y = channels_mean_49_cast_fp16)[name = tensor("zero_mean_49_cast_fp16")]; + tensor zero_mean_sq_49_cast_fp16 = mul(x = zero_mean_49_cast_fp16, y = zero_mean_49_cast_fp16)[name = tensor("zero_mean_sq_49_cast_fp16")]; + tensor var_1983 = const()[name = tensor("op_1983"), val = tensor([1])]; + tensor var_1984_cast_fp16 = reduce_mean(axes = var_1983, keep_dims = var_1969, x = zero_mean_sq_49_cast_fp16)[name = tensor("op_1984_cast_fp16")]; + tensor var_1985_to_fp16 = const()[name = tensor("op_1985_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_1986_cast_fp16 = add(x = var_1984_cast_fp16, y = var_1985_to_fp16)[name = tensor("op_1986_cast_fp16")]; + tensor denom_49_epsilon_0_to_fp16 = const()[name = tensor("denom_49_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_49_cast_fp16 = rsqrt(epsilon = denom_49_epsilon_0_to_fp16, x = var_1986_cast_fp16)[name = tensor("denom_49_cast_fp16")]; + tensor out_49_cast_fp16 = mul(x = zero_mean_49_cast_fp16, y = denom_49_cast_fp16)[name = tensor("out_49_cast_fp16")]; + tensor obj_49_gamma_0_to_fp16 = const()[name = tensor("obj_49_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(126668096)))]; + tensor obj_49_beta_0_to_fp16 = const()[name = tensor("obj_49_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(126670720)))]; + tensor obj_49_epsilon_0_to_fp16 = const()[name = tensor("obj_49_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_49_cast_fp16 = batch_norm(beta = obj_49_beta_0_to_fp16, epsilon = obj_49_epsilon_0_to_fp16, gamma = obj_49_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_49_cast_fp16)[name = tensor("obj_49_cast_fp16")]; + tensor layers_12_self_attn_q_proj_input_shift_to_fp16 = const()[name = tensor("layers_12_self_attn_q_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(126673344)))]; + tensor input_169_cast_fp16 = sub(x = obj_49_cast_fp16, y = layers_12_self_attn_q_proj_input_shift_to_fp16)[name = tensor("input_169_cast_fp16")]; + tensor var_2005 = const()[name = tensor("op_2005"), val = tensor([1, 1])]; + tensor var_2007 = const()[name = tensor("op_2007"), val = tensor([1, 1])]; + tensor x_217_pad_type_0 = const()[name = tensor("x_217_pad_type_0"), val = tensor("custom")]; + tensor x_217_pad_0 = const()[name = tensor("x_217_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_12_self_attn_q_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(126675968))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(127495232))), name = tensor("layers_12_self_attn_q_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_12_self_attn_q_proj_module_bias_to_fp16 = const()[name = tensor("layers_12_self_attn_q_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(127495360)))]; + tensor x_217_cast_fp16 = conv(bias = layers_12_self_attn_q_proj_module_bias_to_fp16, dilations = var_2007, groups = var_1968, pad = x_217_pad_0, pad_type = x_217_pad_type_0, strides = var_2005, weight = layers_12_self_attn_q_proj_module_weight_to_fp16_palettized, x = input_169_cast_fp16)[name = tensor("x_217_cast_fp16")]; + tensor layers_12_self_attn_q_proj_output_scale_to_fp16 = const()[name = tensor("layers_12_self_attn_q_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(127497984)))]; + tensor query_25_cast_fp16 = mul(x = x_217_cast_fp16, y = layers_12_self_attn_q_proj_output_scale_to_fp16)[name = tensor("query_25_cast_fp16")]; + tensor var_2017 = const()[name = tensor("op_2017"), val = tensor([1, 1])]; + tensor var_2019 = const()[name = tensor("op_2019"), val = tensor([1, 1])]; + tensor x_219_pad_type_0 = const()[name = tensor("x_219_pad_type_0"), val = tensor("custom")]; + tensor x_219_pad_0 = const()[name = tensor("x_219_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_12_self_attn_k_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(127500608))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(128319872))), name = tensor("layers_12_self_attn_k_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_12_self_attn_k_proj_module_bias_to_fp16 = const()[name = tensor("layers_12_self_attn_k_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(128320000)))]; + tensor x_219_cast_fp16 = conv(bias = layers_12_self_attn_k_proj_module_bias_to_fp16, dilations = var_2019, groups = var_1968, pad = x_219_pad_0, pad_type = x_219_pad_type_0, strides = var_2017, weight = layers_12_self_attn_k_proj_module_weight_to_fp16_palettized, x = input_169_cast_fp16)[name = tensor("x_219_cast_fp16")]; + tensor layers_12_self_attn_k_proj_output_scale_to_fp16 = const()[name = tensor("layers_12_self_attn_k_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(128322624)))]; + tensor key_25_cast_fp16 = mul(x = x_219_cast_fp16, y = layers_12_self_attn_k_proj_output_scale_to_fp16)[name = tensor("key_25_cast_fp16")]; + tensor var_2029 = const()[name = tensor("op_2029"), val = tensor([1, 1])]; + tensor var_2031 = const()[name = tensor("op_2031"), val = tensor([1, 1])]; + tensor x_221_pad_type_0 = const()[name = tensor("x_221_pad_type_0"), val = tensor("custom")]; + tensor x_221_pad_0 = const()[name = tensor("x_221_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_12_self_attn_v_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(128325248))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(129144512))), name = tensor("layers_12_self_attn_v_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_12_self_attn_v_proj_module_bias_to_fp16 = const()[name = tensor("layers_12_self_attn_v_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(129144640)))]; + tensor x_221_cast_fp16 = conv(bias = layers_12_self_attn_v_proj_module_bias_to_fp16, dilations = var_2031, groups = var_1968, pad = x_221_pad_0, pad_type = x_221_pad_type_0, strides = var_2029, weight = layers_12_self_attn_v_proj_module_weight_to_fp16_palettized, x = input_169_cast_fp16)[name = tensor("x_221_cast_fp16")]; + tensor layers_12_self_attn_v_proj_output_scale_to_fp16 = const()[name = tensor("layers_12_self_attn_v_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(129147264)))]; + tensor value_25_cast_fp16 = mul(x = x_221_cast_fp16, y = layers_12_self_attn_v_proj_output_scale_to_fp16)[name = tensor("value_25_cast_fp16")]; + tensor var_2036 = const()[name = tensor("op_2036"), val = tensor([1, 20, 64, -1])]; + tensor var_2037_cast_fp16 = reshape(shape = var_2036, x = query_25_cast_fp16)[name = tensor("op_2037_cast_fp16")]; + tensor var_2038_to_fp16 = const()[name = tensor("op_2038_to_fp16"), val = tensor(0x1p-3)]; + tensor var_2039_cast_fp16 = mul(x = var_2037_cast_fp16, y = var_2038_to_fp16)[name = tensor("op_2039_cast_fp16")]; + tensor var_2040 = const()[name = tensor("op_2040"), val = tensor([1, 20, 64, -1])]; + tensor var_2041_cast_fp16 = reshape(shape = var_2040, x = key_25_cast_fp16)[name = tensor("op_2041_cast_fp16")]; + tensor mh_w_25_transpose_x_0 = const()[name = tensor("mh_w_25_transpose_x_0"), val = tensor(true)]; + tensor mh_w_25_transpose_y_0 = const()[name = tensor("mh_w_25_transpose_y_0"), val = tensor(false)]; + tensor mh_w_25_cast_fp16 = matmul(transpose_x = mh_w_25_transpose_x_0, transpose_y = mh_w_25_transpose_y_0, x = var_2039_cast_fp16, y = var_2041_cast_fp16)[name = tensor("mh_w_25_cast_fp16")]; + tensor var_2044_cast_fp16 = softmax(axis = var_1966, x = mh_w_25_cast_fp16)[name = tensor("op_2044_cast_fp16")]; + tensor var_2045 = const()[name = tensor("op_2045"), val = tensor([1, 20, 64, -1])]; + tensor var_2046_cast_fp16 = reshape(shape = var_2045, x = value_25_cast_fp16)[name = tensor("op_2046_cast_fp16")]; + tensor attn_25_transpose_x_0 = const()[name = tensor("attn_25_transpose_x_0"), val = tensor(false)]; + tensor attn_25_transpose_y_0 = const()[name = tensor("attn_25_transpose_y_0"), val = tensor(true)]; + tensor attn_25_cast_fp16 = matmul(transpose_x = attn_25_transpose_x_0, transpose_y = attn_25_transpose_y_0, x = var_2046_cast_fp16, y = var_2044_cast_fp16)[name = tensor("attn_25_cast_fp16")]; + tensor var_2049 = const()[name = tensor("op_2049"), val = tensor([1, 1280, 1, -1])]; + tensor x_223_cast_fp16 = reshape(shape = var_2049, x = attn_25_cast_fp16)[name = tensor("x_223_cast_fp16")]; + tensor layers_12_self_attn_o_proj_input_shift_to_fp16 = const()[name = tensor("layers_12_self_attn_o_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(129149888)))]; + tensor input_175_cast_fp16 = sub(x = x_223_cast_fp16, y = layers_12_self_attn_o_proj_input_shift_to_fp16)[name = tensor("input_175_cast_fp16")]; + tensor var_2057 = const()[name = tensor("op_2057"), val = tensor([1, 1])]; + tensor var_2059 = const()[name = tensor("op_2059"), val = tensor([1, 1])]; + tensor x_225_pad_type_0 = const()[name = tensor("x_225_pad_type_0"), val = tensor("custom")]; + tensor x_225_pad_0 = const()[name = tensor("x_225_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_12_self_attn_o_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(129152512))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(129971776))), name = tensor("layers_12_self_attn_o_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_12_self_attn_o_proj_module_bias_to_fp16 = const()[name = tensor("layers_12_self_attn_o_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(129971904)))]; + tensor x_225_cast_fp16 = conv(bias = layers_12_self_attn_o_proj_module_bias_to_fp16, dilations = var_2059, groups = var_1968, pad = x_225_pad_0, pad_type = x_225_pad_type_0, strides = var_2057, weight = layers_12_self_attn_o_proj_module_weight_to_fp16_palettized, x = input_175_cast_fp16)[name = tensor("x_225_cast_fp16")]; + tensor layers_12_self_attn_o_proj_output_scale_to_fp16 = const()[name = tensor("layers_12_self_attn_o_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(129974528)))]; + tensor obj_51_cast_fp16 = mul(x = x_225_cast_fp16, y = layers_12_self_attn_o_proj_output_scale_to_fp16)[name = tensor("obj_51_cast_fp16")]; + tensor inputs_51_cast_fp16 = add(x = inputs_49_cast_fp16, y = obj_51_cast_fp16)[name = tensor("inputs_51_cast_fp16")]; + tensor var_2066 = const()[name = tensor("op_2066"), val = tensor([1])]; + tensor channels_mean_51_cast_fp16 = reduce_mean(axes = var_2066, keep_dims = var_1969, x = inputs_51_cast_fp16)[name = tensor("channels_mean_51_cast_fp16")]; + tensor zero_mean_51_cast_fp16 = sub(x = inputs_51_cast_fp16, y = channels_mean_51_cast_fp16)[name = tensor("zero_mean_51_cast_fp16")]; + tensor zero_mean_sq_51_cast_fp16 = mul(x = zero_mean_51_cast_fp16, y = zero_mean_51_cast_fp16)[name = tensor("zero_mean_sq_51_cast_fp16")]; + tensor var_2070 = const()[name = tensor("op_2070"), val = tensor([1])]; + tensor var_2071_cast_fp16 = reduce_mean(axes = var_2070, keep_dims = var_1969, x = zero_mean_sq_51_cast_fp16)[name = tensor("op_2071_cast_fp16")]; + tensor var_2072_to_fp16 = const()[name = tensor("op_2072_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_2073_cast_fp16 = add(x = var_2071_cast_fp16, y = var_2072_to_fp16)[name = tensor("op_2073_cast_fp16")]; + tensor denom_51_epsilon_0_to_fp16 = const()[name = tensor("denom_51_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_51_cast_fp16 = rsqrt(epsilon = denom_51_epsilon_0_to_fp16, x = var_2073_cast_fp16)[name = tensor("denom_51_cast_fp16")]; + tensor out_51_cast_fp16 = mul(x = zero_mean_51_cast_fp16, y = denom_51_cast_fp16)[name = tensor("out_51_cast_fp16")]; + tensor x_227_gamma_0_to_fp16 = const()[name = tensor("x_227_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(129977152)))]; + tensor x_227_beta_0_to_fp16 = const()[name = tensor("x_227_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(129979776)))]; + tensor x_227_epsilon_0_to_fp16 = const()[name = tensor("x_227_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_227_cast_fp16 = batch_norm(beta = x_227_beta_0_to_fp16, epsilon = x_227_epsilon_0_to_fp16, gamma = x_227_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_51_cast_fp16)[name = tensor("x_227_cast_fp16")]; + tensor layers_12_fc1_input_shift_to_fp16 = const()[name = tensor("layers_12_fc1_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(129982400)))]; + tensor input_177_cast_fp16 = sub(x = x_227_cast_fp16, y = layers_12_fc1_input_shift_to_fp16)[name = tensor("input_177_cast_fp16")]; + tensor var_2088 = const()[name = tensor("op_2088"), val = tensor([1, 1])]; + tensor var_2090 = const()[name = tensor("op_2090"), val = tensor([1, 1])]; + tensor x_229_pad_type_0 = const()[name = tensor("x_229_pad_type_0"), val = tensor("custom")]; + tensor x_229_pad_0 = const()[name = tensor("x_229_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_12_fc1_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(129985024))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(133261888))), name = tensor("layers_12_fc1_module_weight_to_fp16_palettized"), shape = tensor([5120, 1280, 1, 1])]; + tensor layers_12_fc1_module_bias_to_fp16 = const()[name = tensor("layers_12_fc1_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(133262016)))]; + tensor x_229_cast_fp16 = conv(bias = layers_12_fc1_module_bias_to_fp16, dilations = var_2090, groups = var_1968, pad = x_229_pad_0, pad_type = x_229_pad_type_0, strides = var_2088, weight = layers_12_fc1_module_weight_to_fp16_palettized, x = input_177_cast_fp16)[name = tensor("x_229_cast_fp16")]; + tensor layers_12_fc1_output_scale_to_fp16 = const()[name = tensor("layers_12_fc1_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(133272320)))]; + tensor input_179_cast_fp16 = mul(x = x_229_cast_fp16, y = layers_12_fc1_output_scale_to_fp16)[name = tensor("input_179_cast_fp16")]; + tensor x_231_mode_0 = const()[name = tensor("x_231_mode_0"), val = tensor("EXACT")]; + tensor x_231_cast_fp16 = gelu(mode = x_231_mode_0, x = input_179_cast_fp16)[name = tensor("x_231_cast_fp16")]; + tensor layers_12_fc2_input_shift_to_fp16 = const()[name = tensor("layers_12_fc2_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(133282624)))]; + tensor input_181_cast_fp16 = sub(x = x_231_cast_fp16, y = layers_12_fc2_input_shift_to_fp16)[name = tensor("input_181_cast_fp16")]; + tensor var_2101 = const()[name = tensor("op_2101"), val = tensor([1, 1])]; + tensor var_2103 = const()[name = tensor("op_2103"), val = tensor([1, 1])]; + tensor x_233_pad_type_0 = const()[name = tensor("x_233_pad_type_0"), val = tensor("custom")]; + tensor x_233_pad_0 = const()[name = tensor("x_233_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_12_fc2_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(133292928))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(136569792))), name = tensor("layers_12_fc2_module_weight_to_fp16_palettized"), shape = tensor([1280, 5120, 1, 1])]; + tensor layers_12_fc2_module_bias_to_fp16 = const()[name = tensor("layers_12_fc2_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(136569920)))]; + tensor x_233_cast_fp16 = conv(bias = layers_12_fc2_module_bias_to_fp16, dilations = var_2103, groups = var_1968, pad = x_233_pad_0, pad_type = x_233_pad_type_0, strides = var_2101, weight = layers_12_fc2_module_weight_to_fp16_palettized, x = input_181_cast_fp16)[name = tensor("x_233_cast_fp16")]; + tensor layers_12_fc2_output_scale_to_fp16 = const()[name = tensor("layers_12_fc2_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(136572544)))]; + tensor hidden_states_29_cast_fp16 = mul(x = x_233_cast_fp16, y = layers_12_fc2_output_scale_to_fp16)[name = tensor("hidden_states_29_cast_fp16")]; + tensor inputs_53_cast_fp16 = add(x = inputs_51_cast_fp16, y = hidden_states_29_cast_fp16)[name = tensor("inputs_53_cast_fp16")]; + tensor var_2115 = const()[name = tensor("op_2115"), val = tensor(3)]; + tensor var_2117 = const()[name = tensor("op_2117"), val = tensor(1)]; + tensor var_2118 = const()[name = tensor("op_2118"), val = tensor(true)]; + tensor var_2128 = const()[name = tensor("op_2128"), val = tensor([1])]; + tensor channels_mean_53_cast_fp16 = reduce_mean(axes = var_2128, keep_dims = var_2118, x = inputs_53_cast_fp16)[name = tensor("channels_mean_53_cast_fp16")]; + tensor zero_mean_53_cast_fp16 = sub(x = inputs_53_cast_fp16, y = channels_mean_53_cast_fp16)[name = tensor("zero_mean_53_cast_fp16")]; + tensor zero_mean_sq_53_cast_fp16 = mul(x = zero_mean_53_cast_fp16, y = zero_mean_53_cast_fp16)[name = tensor("zero_mean_sq_53_cast_fp16")]; + tensor var_2132 = const()[name = tensor("op_2132"), val = tensor([1])]; + tensor var_2133_cast_fp16 = reduce_mean(axes = var_2132, keep_dims = var_2118, x = zero_mean_sq_53_cast_fp16)[name = tensor("op_2133_cast_fp16")]; + tensor var_2134_to_fp16 = const()[name = tensor("op_2134_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_2135_cast_fp16 = add(x = var_2133_cast_fp16, y = var_2134_to_fp16)[name = tensor("op_2135_cast_fp16")]; + tensor denom_53_epsilon_0_to_fp16 = const()[name = tensor("denom_53_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_53_cast_fp16 = rsqrt(epsilon = denom_53_epsilon_0_to_fp16, x = var_2135_cast_fp16)[name = tensor("denom_53_cast_fp16")]; + tensor out_53_cast_fp16 = mul(x = zero_mean_53_cast_fp16, y = denom_53_cast_fp16)[name = tensor("out_53_cast_fp16")]; + tensor obj_53_gamma_0_to_fp16 = const()[name = tensor("obj_53_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(136575168)))]; + tensor obj_53_beta_0_to_fp16 = const()[name = tensor("obj_53_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(136577792)))]; + tensor obj_53_epsilon_0_to_fp16 = const()[name = tensor("obj_53_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_53_cast_fp16 = batch_norm(beta = obj_53_beta_0_to_fp16, epsilon = obj_53_epsilon_0_to_fp16, gamma = obj_53_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_53_cast_fp16)[name = tensor("obj_53_cast_fp16")]; + tensor layers_13_self_attn_q_proj_input_shift_to_fp16 = const()[name = tensor("layers_13_self_attn_q_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(136580416)))]; + tensor input_183_cast_fp16 = sub(x = obj_53_cast_fp16, y = layers_13_self_attn_q_proj_input_shift_to_fp16)[name = tensor("input_183_cast_fp16")]; + tensor var_2154 = const()[name = tensor("op_2154"), val = tensor([1, 1])]; + tensor var_2156 = const()[name = tensor("op_2156"), val = tensor([1, 1])]; + tensor x_235_pad_type_0 = const()[name = tensor("x_235_pad_type_0"), val = tensor("custom")]; + tensor x_235_pad_0 = const()[name = tensor("x_235_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_13_self_attn_q_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(136583040))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(137402304))), name = tensor("layers_13_self_attn_q_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_13_self_attn_q_proj_module_bias_to_fp16 = const()[name = tensor("layers_13_self_attn_q_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(137402432)))]; + tensor x_235_cast_fp16 = conv(bias = layers_13_self_attn_q_proj_module_bias_to_fp16, dilations = var_2156, groups = var_2117, pad = x_235_pad_0, pad_type = x_235_pad_type_0, strides = var_2154, weight = layers_13_self_attn_q_proj_module_weight_to_fp16_palettized, x = input_183_cast_fp16)[name = tensor("x_235_cast_fp16")]; + tensor layers_13_self_attn_q_proj_output_scale_to_fp16 = const()[name = tensor("layers_13_self_attn_q_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(137405056)))]; + tensor query_27_cast_fp16 = mul(x = x_235_cast_fp16, y = layers_13_self_attn_q_proj_output_scale_to_fp16)[name = tensor("query_27_cast_fp16")]; + tensor var_2166 = const()[name = tensor("op_2166"), val = tensor([1, 1])]; + tensor var_2168 = const()[name = tensor("op_2168"), val = tensor([1, 1])]; + tensor x_237_pad_type_0 = const()[name = tensor("x_237_pad_type_0"), val = tensor("custom")]; + tensor x_237_pad_0 = const()[name = tensor("x_237_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_13_self_attn_k_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(137407680))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(138226944))), name = tensor("layers_13_self_attn_k_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_13_self_attn_k_proj_module_bias_to_fp16 = const()[name = tensor("layers_13_self_attn_k_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(138227072)))]; + tensor x_237_cast_fp16 = conv(bias = layers_13_self_attn_k_proj_module_bias_to_fp16, dilations = var_2168, groups = var_2117, pad = x_237_pad_0, pad_type = x_237_pad_type_0, strides = var_2166, weight = layers_13_self_attn_k_proj_module_weight_to_fp16_palettized, x = input_183_cast_fp16)[name = tensor("x_237_cast_fp16")]; + tensor layers_13_self_attn_k_proj_output_scale_to_fp16 = const()[name = tensor("layers_13_self_attn_k_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(138229696)))]; + tensor key_27_cast_fp16 = mul(x = x_237_cast_fp16, y = layers_13_self_attn_k_proj_output_scale_to_fp16)[name = tensor("key_27_cast_fp16")]; + tensor var_2178 = const()[name = tensor("op_2178"), val = tensor([1, 1])]; + tensor var_2180 = const()[name = tensor("op_2180"), val = tensor([1, 1])]; + tensor x_239_pad_type_0 = const()[name = tensor("x_239_pad_type_0"), val = tensor("custom")]; + tensor x_239_pad_0 = const()[name = tensor("x_239_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_13_self_attn_v_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(138232320))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(139051584))), name = tensor("layers_13_self_attn_v_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_13_self_attn_v_proj_module_bias_to_fp16 = const()[name = tensor("layers_13_self_attn_v_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(139051712)))]; + tensor x_239_cast_fp16 = conv(bias = layers_13_self_attn_v_proj_module_bias_to_fp16, dilations = var_2180, groups = var_2117, pad = x_239_pad_0, pad_type = x_239_pad_type_0, strides = var_2178, weight = layers_13_self_attn_v_proj_module_weight_to_fp16_palettized, x = input_183_cast_fp16)[name = tensor("x_239_cast_fp16")]; + tensor layers_13_self_attn_v_proj_output_scale_to_fp16 = const()[name = tensor("layers_13_self_attn_v_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(139054336)))]; + tensor value_27_cast_fp16 = mul(x = x_239_cast_fp16, y = layers_13_self_attn_v_proj_output_scale_to_fp16)[name = tensor("value_27_cast_fp16")]; + tensor var_2185 = const()[name = tensor("op_2185"), val = tensor([1, 20, 64, -1])]; + tensor var_2186_cast_fp16 = reshape(shape = var_2185, x = query_27_cast_fp16)[name = tensor("op_2186_cast_fp16")]; + tensor var_2187_to_fp16 = const()[name = tensor("op_2187_to_fp16"), val = tensor(0x1p-3)]; + tensor var_2188_cast_fp16 = mul(x = var_2186_cast_fp16, y = var_2187_to_fp16)[name = tensor("op_2188_cast_fp16")]; + tensor var_2189 = const()[name = tensor("op_2189"), val = tensor([1, 20, 64, -1])]; + tensor var_2190_cast_fp16 = reshape(shape = var_2189, x = key_27_cast_fp16)[name = tensor("op_2190_cast_fp16")]; + tensor mh_w_27_transpose_x_0 = const()[name = tensor("mh_w_27_transpose_x_0"), val = tensor(true)]; + tensor mh_w_27_transpose_y_0 = const()[name = tensor("mh_w_27_transpose_y_0"), val = tensor(false)]; + tensor mh_w_27_cast_fp16 = matmul(transpose_x = mh_w_27_transpose_x_0, transpose_y = mh_w_27_transpose_y_0, x = var_2188_cast_fp16, y = var_2190_cast_fp16)[name = tensor("mh_w_27_cast_fp16")]; + tensor var_2193_cast_fp16 = softmax(axis = var_2115, x = mh_w_27_cast_fp16)[name = tensor("op_2193_cast_fp16")]; + tensor var_2194 = const()[name = tensor("op_2194"), val = tensor([1, 20, 64, -1])]; + tensor var_2195_cast_fp16 = reshape(shape = var_2194, x = value_27_cast_fp16)[name = tensor("op_2195_cast_fp16")]; + tensor attn_27_transpose_x_0 = const()[name = tensor("attn_27_transpose_x_0"), val = tensor(false)]; + tensor attn_27_transpose_y_0 = const()[name = tensor("attn_27_transpose_y_0"), val = tensor(true)]; + tensor attn_27_cast_fp16 = matmul(transpose_x = attn_27_transpose_x_0, transpose_y = attn_27_transpose_y_0, x = var_2195_cast_fp16, y = var_2193_cast_fp16)[name = tensor("attn_27_cast_fp16")]; + tensor var_2198 = const()[name = tensor("op_2198"), val = tensor([1, 1280, 1, -1])]; + tensor x_241_cast_fp16 = reshape(shape = var_2198, x = attn_27_cast_fp16)[name = tensor("x_241_cast_fp16")]; + tensor layers_13_self_attn_o_proj_input_shift_to_fp16 = const()[name = tensor("layers_13_self_attn_o_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(139056960)))]; + tensor input_189_cast_fp16 = sub(x = x_241_cast_fp16, y = layers_13_self_attn_o_proj_input_shift_to_fp16)[name = tensor("input_189_cast_fp16")]; + tensor var_2206 = const()[name = tensor("op_2206"), val = tensor([1, 1])]; + tensor var_2208 = const()[name = tensor("op_2208"), val = tensor([1, 1])]; + tensor x_243_pad_type_0 = const()[name = tensor("x_243_pad_type_0"), val = tensor("custom")]; + tensor x_243_pad_0 = const()[name = tensor("x_243_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_13_self_attn_o_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(139059584))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(139878848))), name = tensor("layers_13_self_attn_o_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_13_self_attn_o_proj_module_bias_to_fp16 = const()[name = tensor("layers_13_self_attn_o_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(139878976)))]; + tensor x_243_cast_fp16 = conv(bias = layers_13_self_attn_o_proj_module_bias_to_fp16, dilations = var_2208, groups = var_2117, pad = x_243_pad_0, pad_type = x_243_pad_type_0, strides = var_2206, weight = layers_13_self_attn_o_proj_module_weight_to_fp16_palettized, x = input_189_cast_fp16)[name = tensor("x_243_cast_fp16")]; + tensor layers_13_self_attn_o_proj_output_scale_to_fp16 = const()[name = tensor("layers_13_self_attn_o_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(139881600)))]; + tensor obj_55_cast_fp16 = mul(x = x_243_cast_fp16, y = layers_13_self_attn_o_proj_output_scale_to_fp16)[name = tensor("obj_55_cast_fp16")]; + tensor inputs_55_cast_fp16 = add(x = inputs_53_cast_fp16, y = obj_55_cast_fp16)[name = tensor("inputs_55_cast_fp16")]; + tensor var_2215 = const()[name = tensor("op_2215"), val = tensor([1])]; + tensor channels_mean_55_cast_fp16 = reduce_mean(axes = var_2215, keep_dims = var_2118, x = inputs_55_cast_fp16)[name = tensor("channels_mean_55_cast_fp16")]; + tensor zero_mean_55_cast_fp16 = sub(x = inputs_55_cast_fp16, y = channels_mean_55_cast_fp16)[name = tensor("zero_mean_55_cast_fp16")]; + tensor zero_mean_sq_55_cast_fp16 = mul(x = zero_mean_55_cast_fp16, y = zero_mean_55_cast_fp16)[name = tensor("zero_mean_sq_55_cast_fp16")]; + tensor var_2219 = const()[name = tensor("op_2219"), val = tensor([1])]; + tensor var_2220_cast_fp16 = reduce_mean(axes = var_2219, keep_dims = var_2118, x = zero_mean_sq_55_cast_fp16)[name = tensor("op_2220_cast_fp16")]; + tensor var_2221_to_fp16 = const()[name = tensor("op_2221_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_2222_cast_fp16 = add(x = var_2220_cast_fp16, y = var_2221_to_fp16)[name = tensor("op_2222_cast_fp16")]; + tensor denom_55_epsilon_0_to_fp16 = const()[name = tensor("denom_55_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_55_cast_fp16 = rsqrt(epsilon = denom_55_epsilon_0_to_fp16, x = var_2222_cast_fp16)[name = tensor("denom_55_cast_fp16")]; + tensor out_55_cast_fp16 = mul(x = zero_mean_55_cast_fp16, y = denom_55_cast_fp16)[name = tensor("out_55_cast_fp16")]; + tensor x_245_gamma_0_to_fp16 = const()[name = tensor("x_245_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(139884224)))]; + tensor x_245_beta_0_to_fp16 = const()[name = tensor("x_245_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(139886848)))]; + tensor x_245_epsilon_0_to_fp16 = const()[name = tensor("x_245_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_245_cast_fp16 = batch_norm(beta = x_245_beta_0_to_fp16, epsilon = x_245_epsilon_0_to_fp16, gamma = x_245_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_55_cast_fp16)[name = tensor("x_245_cast_fp16")]; + tensor layers_13_fc1_input_shift_to_fp16 = const()[name = tensor("layers_13_fc1_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(139889472)))]; + tensor input_191_cast_fp16 = sub(x = x_245_cast_fp16, y = layers_13_fc1_input_shift_to_fp16)[name = tensor("input_191_cast_fp16")]; + tensor var_2237 = const()[name = tensor("op_2237"), val = tensor([1, 1])]; + tensor var_2239 = const()[name = tensor("op_2239"), val = tensor([1, 1])]; + tensor x_247_pad_type_0 = const()[name = tensor("x_247_pad_type_0"), val = tensor("custom")]; + tensor x_247_pad_0 = const()[name = tensor("x_247_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_13_fc1_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(139892096))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(143168960))), name = tensor("layers_13_fc1_module_weight_to_fp16_palettized"), shape = tensor([5120, 1280, 1, 1])]; + tensor layers_13_fc1_module_bias_to_fp16 = const()[name = tensor("layers_13_fc1_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(143169088)))]; + tensor x_247_cast_fp16 = conv(bias = layers_13_fc1_module_bias_to_fp16, dilations = var_2239, groups = var_2117, pad = x_247_pad_0, pad_type = x_247_pad_type_0, strides = var_2237, weight = layers_13_fc1_module_weight_to_fp16_palettized, x = input_191_cast_fp16)[name = tensor("x_247_cast_fp16")]; + tensor layers_13_fc1_output_scale_to_fp16 = const()[name = tensor("layers_13_fc1_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(143179392)))]; + tensor input_193_cast_fp16 = mul(x = x_247_cast_fp16, y = layers_13_fc1_output_scale_to_fp16)[name = tensor("input_193_cast_fp16")]; + tensor x_249_mode_0 = const()[name = tensor("x_249_mode_0"), val = tensor("EXACT")]; + tensor x_249_cast_fp16 = gelu(mode = x_249_mode_0, x = input_193_cast_fp16)[name = tensor("x_249_cast_fp16")]; + tensor layers_13_fc2_input_shift_to_fp16 = const()[name = tensor("layers_13_fc2_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(143189696)))]; + tensor input_195_cast_fp16 = sub(x = x_249_cast_fp16, y = layers_13_fc2_input_shift_to_fp16)[name = tensor("input_195_cast_fp16")]; + tensor var_2250 = const()[name = tensor("op_2250"), val = tensor([1, 1])]; + tensor var_2252 = const()[name = tensor("op_2252"), val = tensor([1, 1])]; + tensor x_251_pad_type_0 = const()[name = tensor("x_251_pad_type_0"), val = tensor("custom")]; + tensor x_251_pad_0 = const()[name = tensor("x_251_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_13_fc2_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(143200000))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(146476864))), name = tensor("layers_13_fc2_module_weight_to_fp16_palettized"), shape = tensor([1280, 5120, 1, 1])]; + tensor layers_13_fc2_module_bias_to_fp16 = const()[name = tensor("layers_13_fc2_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(146476992)))]; + tensor x_251_cast_fp16 = conv(bias = layers_13_fc2_module_bias_to_fp16, dilations = var_2252, groups = var_2117, pad = x_251_pad_0, pad_type = x_251_pad_type_0, strides = var_2250, weight = layers_13_fc2_module_weight_to_fp16_palettized, x = input_195_cast_fp16)[name = tensor("x_251_cast_fp16")]; + tensor layers_13_fc2_output_scale_to_fp16 = const()[name = tensor("layers_13_fc2_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(146479616)))]; + tensor hidden_states_31_cast_fp16 = mul(x = x_251_cast_fp16, y = layers_13_fc2_output_scale_to_fp16)[name = tensor("hidden_states_31_cast_fp16")]; + tensor inputs_57_cast_fp16 = add(x = inputs_55_cast_fp16, y = hidden_states_31_cast_fp16)[name = tensor("inputs_57_cast_fp16")]; + tensor var_2264 = const()[name = tensor("op_2264"), val = tensor(3)]; + tensor var_2266 = const()[name = tensor("op_2266"), val = tensor(1)]; + tensor var_2267 = const()[name = tensor("op_2267"), val = tensor(true)]; + tensor var_2277 = const()[name = tensor("op_2277"), val = tensor([1])]; + tensor channels_mean_57_cast_fp16 = reduce_mean(axes = var_2277, keep_dims = var_2267, x = inputs_57_cast_fp16)[name = tensor("channels_mean_57_cast_fp16")]; + tensor zero_mean_57_cast_fp16 = sub(x = inputs_57_cast_fp16, y = channels_mean_57_cast_fp16)[name = tensor("zero_mean_57_cast_fp16")]; + tensor zero_mean_sq_57_cast_fp16 = mul(x = zero_mean_57_cast_fp16, y = zero_mean_57_cast_fp16)[name = tensor("zero_mean_sq_57_cast_fp16")]; + tensor var_2281 = const()[name = tensor("op_2281"), val = tensor([1])]; + tensor var_2282_cast_fp16 = reduce_mean(axes = var_2281, keep_dims = var_2267, x = zero_mean_sq_57_cast_fp16)[name = tensor("op_2282_cast_fp16")]; + tensor var_2283_to_fp16 = const()[name = tensor("op_2283_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_2284_cast_fp16 = add(x = var_2282_cast_fp16, y = var_2283_to_fp16)[name = tensor("op_2284_cast_fp16")]; + tensor denom_57_epsilon_0_to_fp16 = const()[name = tensor("denom_57_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_57_cast_fp16 = rsqrt(epsilon = denom_57_epsilon_0_to_fp16, x = var_2284_cast_fp16)[name = tensor("denom_57_cast_fp16")]; + tensor out_57_cast_fp16 = mul(x = zero_mean_57_cast_fp16, y = denom_57_cast_fp16)[name = tensor("out_57_cast_fp16")]; + tensor obj_57_gamma_0_to_fp16 = const()[name = tensor("obj_57_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(146482240)))]; + tensor obj_57_beta_0_to_fp16 = const()[name = tensor("obj_57_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(146484864)))]; + tensor obj_57_epsilon_0_to_fp16 = const()[name = tensor("obj_57_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_57_cast_fp16 = batch_norm(beta = obj_57_beta_0_to_fp16, epsilon = obj_57_epsilon_0_to_fp16, gamma = obj_57_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_57_cast_fp16)[name = tensor("obj_57_cast_fp16")]; + tensor layers_14_self_attn_q_proj_input_shift_to_fp16 = const()[name = tensor("layers_14_self_attn_q_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(146487488)))]; + tensor input_197_cast_fp16 = sub(x = obj_57_cast_fp16, y = layers_14_self_attn_q_proj_input_shift_to_fp16)[name = tensor("input_197_cast_fp16")]; + tensor var_2303 = const()[name = tensor("op_2303"), val = tensor([1, 1])]; + tensor var_2305 = const()[name = tensor("op_2305"), val = tensor([1, 1])]; + tensor x_253_pad_type_0 = const()[name = tensor("x_253_pad_type_0"), val = tensor("custom")]; + tensor x_253_pad_0 = const()[name = tensor("x_253_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_14_self_attn_q_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(146490112))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(147309376))), name = tensor("layers_14_self_attn_q_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_14_self_attn_q_proj_module_bias_to_fp16 = const()[name = tensor("layers_14_self_attn_q_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(147309504)))]; + tensor x_253_cast_fp16 = conv(bias = layers_14_self_attn_q_proj_module_bias_to_fp16, dilations = var_2305, groups = var_2266, pad = x_253_pad_0, pad_type = x_253_pad_type_0, strides = var_2303, weight = layers_14_self_attn_q_proj_module_weight_to_fp16_palettized, x = input_197_cast_fp16)[name = tensor("x_253_cast_fp16")]; + tensor layers_14_self_attn_q_proj_output_scale_to_fp16 = const()[name = tensor("layers_14_self_attn_q_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(147312128)))]; + tensor query_29_cast_fp16 = mul(x = x_253_cast_fp16, y = layers_14_self_attn_q_proj_output_scale_to_fp16)[name = tensor("query_29_cast_fp16")]; + tensor var_2315 = const()[name = tensor("op_2315"), val = tensor([1, 1])]; + tensor var_2317 = const()[name = tensor("op_2317"), val = tensor([1, 1])]; + tensor x_255_pad_type_0 = const()[name = tensor("x_255_pad_type_0"), val = tensor("custom")]; + tensor x_255_pad_0 = const()[name = tensor("x_255_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_14_self_attn_k_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(147314752))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(148134016))), name = tensor("layers_14_self_attn_k_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_14_self_attn_k_proj_module_bias_to_fp16 = const()[name = tensor("layers_14_self_attn_k_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(148134144)))]; + tensor x_255_cast_fp16 = conv(bias = layers_14_self_attn_k_proj_module_bias_to_fp16, dilations = var_2317, groups = var_2266, pad = x_255_pad_0, pad_type = x_255_pad_type_0, strides = var_2315, weight = layers_14_self_attn_k_proj_module_weight_to_fp16_palettized, x = input_197_cast_fp16)[name = tensor("x_255_cast_fp16")]; + tensor layers_14_self_attn_k_proj_output_scale_to_fp16 = const()[name = tensor("layers_14_self_attn_k_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(148136768)))]; + tensor key_29_cast_fp16 = mul(x = x_255_cast_fp16, y = layers_14_self_attn_k_proj_output_scale_to_fp16)[name = tensor("key_29_cast_fp16")]; + tensor var_2327 = const()[name = tensor("op_2327"), val = tensor([1, 1])]; + tensor var_2329 = const()[name = tensor("op_2329"), val = tensor([1, 1])]; + tensor x_257_pad_type_0 = const()[name = tensor("x_257_pad_type_0"), val = tensor("custom")]; + tensor x_257_pad_0 = const()[name = tensor("x_257_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_14_self_attn_v_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(148139392))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(148958656))), name = tensor("layers_14_self_attn_v_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_14_self_attn_v_proj_module_bias_to_fp16 = const()[name = tensor("layers_14_self_attn_v_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(148958784)))]; + tensor x_257_cast_fp16 = conv(bias = layers_14_self_attn_v_proj_module_bias_to_fp16, dilations = var_2329, groups = var_2266, pad = x_257_pad_0, pad_type = x_257_pad_type_0, strides = var_2327, weight = layers_14_self_attn_v_proj_module_weight_to_fp16_palettized, x = input_197_cast_fp16)[name = tensor("x_257_cast_fp16")]; + tensor layers_14_self_attn_v_proj_output_scale_to_fp16 = const()[name = tensor("layers_14_self_attn_v_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(148961408)))]; + tensor value_29_cast_fp16 = mul(x = x_257_cast_fp16, y = layers_14_self_attn_v_proj_output_scale_to_fp16)[name = tensor("value_29_cast_fp16")]; + tensor var_2334 = const()[name = tensor("op_2334"), val = tensor([1, 20, 64, -1])]; + tensor var_2335_cast_fp16 = reshape(shape = var_2334, x = query_29_cast_fp16)[name = tensor("op_2335_cast_fp16")]; + tensor var_2336_to_fp16 = const()[name = tensor("op_2336_to_fp16"), val = tensor(0x1p-3)]; + tensor var_2337_cast_fp16 = mul(x = var_2335_cast_fp16, y = var_2336_to_fp16)[name = tensor("op_2337_cast_fp16")]; + tensor var_2338 = const()[name = tensor("op_2338"), val = tensor([1, 20, 64, -1])]; + tensor var_2339_cast_fp16 = reshape(shape = var_2338, x = key_29_cast_fp16)[name = tensor("op_2339_cast_fp16")]; + tensor mh_w_29_transpose_x_0 = const()[name = tensor("mh_w_29_transpose_x_0"), val = tensor(true)]; + tensor mh_w_29_transpose_y_0 = const()[name = tensor("mh_w_29_transpose_y_0"), val = tensor(false)]; + tensor mh_w_29_cast_fp16 = matmul(transpose_x = mh_w_29_transpose_x_0, transpose_y = mh_w_29_transpose_y_0, x = var_2337_cast_fp16, y = var_2339_cast_fp16)[name = tensor("mh_w_29_cast_fp16")]; + tensor var_2342_cast_fp16 = softmax(axis = var_2264, x = mh_w_29_cast_fp16)[name = tensor("op_2342_cast_fp16")]; + tensor var_2343 = const()[name = tensor("op_2343"), val = tensor([1, 20, 64, -1])]; + tensor var_2344_cast_fp16 = reshape(shape = var_2343, x = value_29_cast_fp16)[name = tensor("op_2344_cast_fp16")]; + tensor attn_29_transpose_x_0 = const()[name = tensor("attn_29_transpose_x_0"), val = tensor(false)]; + tensor attn_29_transpose_y_0 = const()[name = tensor("attn_29_transpose_y_0"), val = tensor(true)]; + tensor attn_29_cast_fp16 = matmul(transpose_x = attn_29_transpose_x_0, transpose_y = attn_29_transpose_y_0, x = var_2344_cast_fp16, y = var_2342_cast_fp16)[name = tensor("attn_29_cast_fp16")]; + tensor var_2347 = const()[name = tensor("op_2347"), val = tensor([1, 1280, 1, -1])]; + tensor x_259_cast_fp16 = reshape(shape = var_2347, x = attn_29_cast_fp16)[name = tensor("x_259_cast_fp16")]; + tensor layers_14_self_attn_o_proj_input_shift_to_fp16 = const()[name = tensor("layers_14_self_attn_o_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(148964032)))]; + tensor input_203_cast_fp16 = sub(x = x_259_cast_fp16, y = layers_14_self_attn_o_proj_input_shift_to_fp16)[name = tensor("input_203_cast_fp16")]; + tensor var_2355 = const()[name = tensor("op_2355"), val = tensor([1, 1])]; + tensor var_2357 = const()[name = tensor("op_2357"), val = tensor([1, 1])]; + tensor x_261_pad_type_0 = const()[name = tensor("x_261_pad_type_0"), val = tensor("custom")]; + tensor x_261_pad_0 = const()[name = tensor("x_261_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_14_self_attn_o_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(148966656))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(149785920))), name = tensor("layers_14_self_attn_o_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_14_self_attn_o_proj_module_bias_to_fp16 = const()[name = tensor("layers_14_self_attn_o_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(149786048)))]; + tensor x_261_cast_fp16 = conv(bias = layers_14_self_attn_o_proj_module_bias_to_fp16, dilations = var_2357, groups = var_2266, pad = x_261_pad_0, pad_type = x_261_pad_type_0, strides = var_2355, weight = layers_14_self_attn_o_proj_module_weight_to_fp16_palettized, x = input_203_cast_fp16)[name = tensor("x_261_cast_fp16")]; + tensor layers_14_self_attn_o_proj_output_scale_to_fp16 = const()[name = tensor("layers_14_self_attn_o_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(149788672)))]; + tensor obj_59_cast_fp16 = mul(x = x_261_cast_fp16, y = layers_14_self_attn_o_proj_output_scale_to_fp16)[name = tensor("obj_59_cast_fp16")]; + tensor inputs_59_cast_fp16 = add(x = inputs_57_cast_fp16, y = obj_59_cast_fp16)[name = tensor("inputs_59_cast_fp16")]; + tensor var_2364 = const()[name = tensor("op_2364"), val = tensor([1])]; + tensor channels_mean_59_cast_fp16 = reduce_mean(axes = var_2364, keep_dims = var_2267, x = inputs_59_cast_fp16)[name = tensor("channels_mean_59_cast_fp16")]; + tensor zero_mean_59_cast_fp16 = sub(x = inputs_59_cast_fp16, y = channels_mean_59_cast_fp16)[name = tensor("zero_mean_59_cast_fp16")]; + tensor zero_mean_sq_59_cast_fp16 = mul(x = zero_mean_59_cast_fp16, y = zero_mean_59_cast_fp16)[name = tensor("zero_mean_sq_59_cast_fp16")]; + tensor var_2368 = const()[name = tensor("op_2368"), val = tensor([1])]; + tensor var_2369_cast_fp16 = reduce_mean(axes = var_2368, keep_dims = var_2267, x = zero_mean_sq_59_cast_fp16)[name = tensor("op_2369_cast_fp16")]; + tensor var_2370_to_fp16 = const()[name = tensor("op_2370_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_2371_cast_fp16 = add(x = var_2369_cast_fp16, y = var_2370_to_fp16)[name = tensor("op_2371_cast_fp16")]; + tensor denom_59_epsilon_0_to_fp16 = const()[name = tensor("denom_59_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_59_cast_fp16 = rsqrt(epsilon = denom_59_epsilon_0_to_fp16, x = var_2371_cast_fp16)[name = tensor("denom_59_cast_fp16")]; + tensor out_59_cast_fp16 = mul(x = zero_mean_59_cast_fp16, y = denom_59_cast_fp16)[name = tensor("out_59_cast_fp16")]; + tensor x_263_gamma_0_to_fp16 = const()[name = tensor("x_263_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(149791296)))]; + tensor x_263_beta_0_to_fp16 = const()[name = tensor("x_263_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(149793920)))]; + tensor x_263_epsilon_0_to_fp16 = const()[name = tensor("x_263_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_263_cast_fp16 = batch_norm(beta = x_263_beta_0_to_fp16, epsilon = x_263_epsilon_0_to_fp16, gamma = x_263_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_59_cast_fp16)[name = tensor("x_263_cast_fp16")]; + tensor layers_14_fc1_input_shift_to_fp16 = const()[name = tensor("layers_14_fc1_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(149796544)))]; + tensor input_205_cast_fp16 = sub(x = x_263_cast_fp16, y = layers_14_fc1_input_shift_to_fp16)[name = tensor("input_205_cast_fp16")]; + tensor var_2386 = const()[name = tensor("op_2386"), val = tensor([1, 1])]; + tensor var_2388 = const()[name = tensor("op_2388"), val = tensor([1, 1])]; + tensor x_265_pad_type_0 = const()[name = tensor("x_265_pad_type_0"), val = tensor("custom")]; + tensor x_265_pad_0 = const()[name = tensor("x_265_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_14_fc1_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(149799168))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(153076032))), name = tensor("layers_14_fc1_module_weight_to_fp16_palettized"), shape = tensor([5120, 1280, 1, 1])]; + tensor layers_14_fc1_module_bias_to_fp16 = const()[name = tensor("layers_14_fc1_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(153076160)))]; + tensor x_265_cast_fp16 = conv(bias = layers_14_fc1_module_bias_to_fp16, dilations = var_2388, groups = var_2266, pad = x_265_pad_0, pad_type = x_265_pad_type_0, strides = var_2386, weight = layers_14_fc1_module_weight_to_fp16_palettized, x = input_205_cast_fp16)[name = tensor("x_265_cast_fp16")]; + tensor layers_14_fc1_output_scale_to_fp16 = const()[name = tensor("layers_14_fc1_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(153086464)))]; + tensor input_207_cast_fp16 = mul(x = x_265_cast_fp16, y = layers_14_fc1_output_scale_to_fp16)[name = tensor("input_207_cast_fp16")]; + tensor x_267_mode_0 = const()[name = tensor("x_267_mode_0"), val = tensor("EXACT")]; + tensor x_267_cast_fp16 = gelu(mode = x_267_mode_0, x = input_207_cast_fp16)[name = tensor("x_267_cast_fp16")]; + tensor layers_14_fc2_input_shift_to_fp16 = const()[name = tensor("layers_14_fc2_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(153096768)))]; + tensor input_209_cast_fp16 = sub(x = x_267_cast_fp16, y = layers_14_fc2_input_shift_to_fp16)[name = tensor("input_209_cast_fp16")]; + tensor var_2399 = const()[name = tensor("op_2399"), val = tensor([1, 1])]; + tensor var_2401 = const()[name = tensor("op_2401"), val = tensor([1, 1])]; + tensor x_269_pad_type_0 = const()[name = tensor("x_269_pad_type_0"), val = tensor("custom")]; + tensor x_269_pad_0 = const()[name = tensor("x_269_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_14_fc2_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(153107072))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(156383936))), name = tensor("layers_14_fc2_module_weight_to_fp16_palettized"), shape = tensor([1280, 5120, 1, 1])]; + tensor layers_14_fc2_module_bias_to_fp16 = const()[name = tensor("layers_14_fc2_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(156384064)))]; + tensor x_269_cast_fp16 = conv(bias = layers_14_fc2_module_bias_to_fp16, dilations = var_2401, groups = var_2266, pad = x_269_pad_0, pad_type = x_269_pad_type_0, strides = var_2399, weight = layers_14_fc2_module_weight_to_fp16_palettized, x = input_209_cast_fp16)[name = tensor("x_269_cast_fp16")]; + tensor layers_14_fc2_output_scale_to_fp16 = const()[name = tensor("layers_14_fc2_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(156386688)))]; + tensor hidden_states_33_cast_fp16 = mul(x = x_269_cast_fp16, y = layers_14_fc2_output_scale_to_fp16)[name = tensor("hidden_states_33_cast_fp16")]; + tensor inputs_61_cast_fp16 = add(x = inputs_59_cast_fp16, y = hidden_states_33_cast_fp16)[name = tensor("inputs_61_cast_fp16")]; + tensor var_2413 = const()[name = tensor("op_2413"), val = tensor(3)]; + tensor var_2415 = const()[name = tensor("op_2415"), val = tensor(1)]; + tensor var_2416 = const()[name = tensor("op_2416"), val = tensor(true)]; + tensor var_2426 = const()[name = tensor("op_2426"), val = tensor([1])]; + tensor channels_mean_61_cast_fp16 = reduce_mean(axes = var_2426, keep_dims = var_2416, x = inputs_61_cast_fp16)[name = tensor("channels_mean_61_cast_fp16")]; + tensor zero_mean_61_cast_fp16 = sub(x = inputs_61_cast_fp16, y = channels_mean_61_cast_fp16)[name = tensor("zero_mean_61_cast_fp16")]; + tensor zero_mean_sq_61_cast_fp16 = mul(x = zero_mean_61_cast_fp16, y = zero_mean_61_cast_fp16)[name = tensor("zero_mean_sq_61_cast_fp16")]; + tensor var_2430 = const()[name = tensor("op_2430"), val = tensor([1])]; + tensor var_2431_cast_fp16 = reduce_mean(axes = var_2430, keep_dims = var_2416, x = zero_mean_sq_61_cast_fp16)[name = tensor("op_2431_cast_fp16")]; + tensor var_2432_to_fp16 = const()[name = tensor("op_2432_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_2433_cast_fp16 = add(x = var_2431_cast_fp16, y = var_2432_to_fp16)[name = tensor("op_2433_cast_fp16")]; + tensor denom_61_epsilon_0_to_fp16 = const()[name = tensor("denom_61_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_61_cast_fp16 = rsqrt(epsilon = denom_61_epsilon_0_to_fp16, x = var_2433_cast_fp16)[name = tensor("denom_61_cast_fp16")]; + tensor out_61_cast_fp16 = mul(x = zero_mean_61_cast_fp16, y = denom_61_cast_fp16)[name = tensor("out_61_cast_fp16")]; + tensor obj_61_gamma_0_to_fp16 = const()[name = tensor("obj_61_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(156389312)))]; + tensor obj_61_beta_0_to_fp16 = const()[name = tensor("obj_61_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(156391936)))]; + tensor obj_61_epsilon_0_to_fp16 = const()[name = tensor("obj_61_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_61_cast_fp16 = batch_norm(beta = obj_61_beta_0_to_fp16, epsilon = obj_61_epsilon_0_to_fp16, gamma = obj_61_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_61_cast_fp16)[name = tensor("obj_61_cast_fp16")]; + tensor layers_15_self_attn_q_proj_input_shift_to_fp16 = const()[name = tensor("layers_15_self_attn_q_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(156394560)))]; + tensor input_211_cast_fp16 = sub(x = obj_61_cast_fp16, y = layers_15_self_attn_q_proj_input_shift_to_fp16)[name = tensor("input_211_cast_fp16")]; + tensor var_2452 = const()[name = tensor("op_2452"), val = tensor([1, 1])]; + tensor var_2454 = const()[name = tensor("op_2454"), val = tensor([1, 1])]; + tensor x_271_pad_type_0 = const()[name = tensor("x_271_pad_type_0"), val = tensor("custom")]; + tensor x_271_pad_0 = const()[name = tensor("x_271_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_15_self_attn_q_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(156397184))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(157216448))), name = tensor("layers_15_self_attn_q_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_15_self_attn_q_proj_module_bias_to_fp16 = const()[name = tensor("layers_15_self_attn_q_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(157216576)))]; + tensor x_271_cast_fp16 = conv(bias = layers_15_self_attn_q_proj_module_bias_to_fp16, dilations = var_2454, groups = var_2415, pad = x_271_pad_0, pad_type = x_271_pad_type_0, strides = var_2452, weight = layers_15_self_attn_q_proj_module_weight_to_fp16_palettized, x = input_211_cast_fp16)[name = tensor("x_271_cast_fp16")]; + tensor layers_15_self_attn_q_proj_output_scale_to_fp16 = const()[name = tensor("layers_15_self_attn_q_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(157219200)))]; + tensor query_31_cast_fp16 = mul(x = x_271_cast_fp16, y = layers_15_self_attn_q_proj_output_scale_to_fp16)[name = tensor("query_31_cast_fp16")]; + tensor var_2464 = const()[name = tensor("op_2464"), val = tensor([1, 1])]; + tensor var_2466 = const()[name = tensor("op_2466"), val = tensor([1, 1])]; + tensor x_273_pad_type_0 = const()[name = tensor("x_273_pad_type_0"), val = tensor("custom")]; + tensor x_273_pad_0 = const()[name = tensor("x_273_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_15_self_attn_k_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(157221824))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(158041088))), name = tensor("layers_15_self_attn_k_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_15_self_attn_k_proj_module_bias_to_fp16 = const()[name = tensor("layers_15_self_attn_k_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(158041216)))]; + tensor x_273_cast_fp16 = conv(bias = layers_15_self_attn_k_proj_module_bias_to_fp16, dilations = var_2466, groups = var_2415, pad = x_273_pad_0, pad_type = x_273_pad_type_0, strides = var_2464, weight = layers_15_self_attn_k_proj_module_weight_to_fp16_palettized, x = input_211_cast_fp16)[name = tensor("x_273_cast_fp16")]; + tensor layers_15_self_attn_k_proj_output_scale_to_fp16 = const()[name = tensor("layers_15_self_attn_k_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(158043840)))]; + tensor key_31_cast_fp16 = mul(x = x_273_cast_fp16, y = layers_15_self_attn_k_proj_output_scale_to_fp16)[name = tensor("key_31_cast_fp16")]; + tensor var_2476 = const()[name = tensor("op_2476"), val = tensor([1, 1])]; + tensor var_2478 = const()[name = tensor("op_2478"), val = tensor([1, 1])]; + tensor x_275_pad_type_0 = const()[name = tensor("x_275_pad_type_0"), val = tensor("custom")]; + tensor x_275_pad_0 = const()[name = tensor("x_275_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_15_self_attn_v_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(158046464))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(158865728))), name = tensor("layers_15_self_attn_v_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_15_self_attn_v_proj_module_bias_to_fp16 = const()[name = tensor("layers_15_self_attn_v_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(158865856)))]; + tensor x_275_cast_fp16 = conv(bias = layers_15_self_attn_v_proj_module_bias_to_fp16, dilations = var_2478, groups = var_2415, pad = x_275_pad_0, pad_type = x_275_pad_type_0, strides = var_2476, weight = layers_15_self_attn_v_proj_module_weight_to_fp16_palettized, x = input_211_cast_fp16)[name = tensor("x_275_cast_fp16")]; + tensor layers_15_self_attn_v_proj_output_scale_to_fp16 = const()[name = tensor("layers_15_self_attn_v_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(158868480)))]; + tensor value_31_cast_fp16 = mul(x = x_275_cast_fp16, y = layers_15_self_attn_v_proj_output_scale_to_fp16)[name = tensor("value_31_cast_fp16")]; + tensor var_2483 = const()[name = tensor("op_2483"), val = tensor([1, 20, 64, -1])]; + tensor var_2484_cast_fp16 = reshape(shape = var_2483, x = query_31_cast_fp16)[name = tensor("op_2484_cast_fp16")]; + tensor var_2485_to_fp16 = const()[name = tensor("op_2485_to_fp16"), val = tensor(0x1p-3)]; + tensor var_2486_cast_fp16 = mul(x = var_2484_cast_fp16, y = var_2485_to_fp16)[name = tensor("op_2486_cast_fp16")]; + tensor var_2487 = const()[name = tensor("op_2487"), val = tensor([1, 20, 64, -1])]; + tensor var_2488_cast_fp16 = reshape(shape = var_2487, x = key_31_cast_fp16)[name = tensor("op_2488_cast_fp16")]; + tensor mh_w_31_transpose_x_0 = const()[name = tensor("mh_w_31_transpose_x_0"), val = tensor(true)]; + tensor mh_w_31_transpose_y_0 = const()[name = tensor("mh_w_31_transpose_y_0"), val = tensor(false)]; + tensor mh_w_31_cast_fp16 = matmul(transpose_x = mh_w_31_transpose_x_0, transpose_y = mh_w_31_transpose_y_0, x = var_2486_cast_fp16, y = var_2488_cast_fp16)[name = tensor("mh_w_31_cast_fp16")]; + tensor var_2491_cast_fp16 = softmax(axis = var_2413, x = mh_w_31_cast_fp16)[name = tensor("op_2491_cast_fp16")]; + tensor var_2492 = const()[name = tensor("op_2492"), val = tensor([1, 20, 64, -1])]; + tensor var_2493_cast_fp16 = reshape(shape = var_2492, x = value_31_cast_fp16)[name = tensor("op_2493_cast_fp16")]; + tensor attn_31_transpose_x_0 = const()[name = tensor("attn_31_transpose_x_0"), val = tensor(false)]; + tensor attn_31_transpose_y_0 = const()[name = tensor("attn_31_transpose_y_0"), val = tensor(true)]; + tensor attn_31_cast_fp16 = matmul(transpose_x = attn_31_transpose_x_0, transpose_y = attn_31_transpose_y_0, x = var_2493_cast_fp16, y = var_2491_cast_fp16)[name = tensor("attn_31_cast_fp16")]; + tensor var_2496 = const()[name = tensor("op_2496"), val = tensor([1, 1280, 1, -1])]; + tensor x_277_cast_fp16 = reshape(shape = var_2496, x = attn_31_cast_fp16)[name = tensor("x_277_cast_fp16")]; + tensor layers_15_self_attn_o_proj_input_shift_to_fp16 = const()[name = tensor("layers_15_self_attn_o_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(158871104)))]; + tensor input_217_cast_fp16 = sub(x = x_277_cast_fp16, y = layers_15_self_attn_o_proj_input_shift_to_fp16)[name = tensor("input_217_cast_fp16")]; + tensor var_2504 = const()[name = tensor("op_2504"), val = tensor([1, 1])]; + tensor var_2506 = const()[name = tensor("op_2506"), val = tensor([1, 1])]; + tensor x_279_pad_type_0 = const()[name = tensor("x_279_pad_type_0"), val = tensor("custom")]; + tensor x_279_pad_0 = const()[name = tensor("x_279_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_15_self_attn_o_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(158873728))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(159692992))), name = tensor("layers_15_self_attn_o_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_15_self_attn_o_proj_module_bias_to_fp16 = const()[name = tensor("layers_15_self_attn_o_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(159693120)))]; + tensor x_279_cast_fp16 = conv(bias = layers_15_self_attn_o_proj_module_bias_to_fp16, dilations = var_2506, groups = var_2415, pad = x_279_pad_0, pad_type = x_279_pad_type_0, strides = var_2504, weight = layers_15_self_attn_o_proj_module_weight_to_fp16_palettized, x = input_217_cast_fp16)[name = tensor("x_279_cast_fp16")]; + tensor layers_15_self_attn_o_proj_output_scale_to_fp16 = const()[name = tensor("layers_15_self_attn_o_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(159695744)))]; + tensor obj_63_cast_fp16 = mul(x = x_279_cast_fp16, y = layers_15_self_attn_o_proj_output_scale_to_fp16)[name = tensor("obj_63_cast_fp16")]; + tensor inputs_63_cast_fp16 = add(x = inputs_61_cast_fp16, y = obj_63_cast_fp16)[name = tensor("inputs_63_cast_fp16")]; + tensor var_2513 = const()[name = tensor("op_2513"), val = tensor([1])]; + tensor channels_mean_63_cast_fp16 = reduce_mean(axes = var_2513, keep_dims = var_2416, x = inputs_63_cast_fp16)[name = tensor("channels_mean_63_cast_fp16")]; + tensor zero_mean_63_cast_fp16 = sub(x = inputs_63_cast_fp16, y = channels_mean_63_cast_fp16)[name = tensor("zero_mean_63_cast_fp16")]; + tensor zero_mean_sq_63_cast_fp16 = mul(x = zero_mean_63_cast_fp16, y = zero_mean_63_cast_fp16)[name = tensor("zero_mean_sq_63_cast_fp16")]; + tensor var_2517 = const()[name = tensor("op_2517"), val = tensor([1])]; + tensor var_2518_cast_fp16 = reduce_mean(axes = var_2517, keep_dims = var_2416, x = zero_mean_sq_63_cast_fp16)[name = tensor("op_2518_cast_fp16")]; + tensor var_2519_to_fp16 = const()[name = tensor("op_2519_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_2520_cast_fp16 = add(x = var_2518_cast_fp16, y = var_2519_to_fp16)[name = tensor("op_2520_cast_fp16")]; + tensor denom_63_epsilon_0_to_fp16 = const()[name = tensor("denom_63_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_63_cast_fp16 = rsqrt(epsilon = denom_63_epsilon_0_to_fp16, x = var_2520_cast_fp16)[name = tensor("denom_63_cast_fp16")]; + tensor out_63_cast_fp16 = mul(x = zero_mean_63_cast_fp16, y = denom_63_cast_fp16)[name = tensor("out_63_cast_fp16")]; + tensor x_281_gamma_0_to_fp16 = const()[name = tensor("x_281_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(159698368)))]; + tensor x_281_beta_0_to_fp16 = const()[name = tensor("x_281_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(159700992)))]; + tensor x_281_epsilon_0_to_fp16 = const()[name = tensor("x_281_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_281_cast_fp16 = batch_norm(beta = x_281_beta_0_to_fp16, epsilon = x_281_epsilon_0_to_fp16, gamma = x_281_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_63_cast_fp16)[name = tensor("x_281_cast_fp16")]; + tensor layers_15_fc1_input_shift_to_fp16 = const()[name = tensor("layers_15_fc1_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(159703616)))]; + tensor input_219_cast_fp16 = sub(x = x_281_cast_fp16, y = layers_15_fc1_input_shift_to_fp16)[name = tensor("input_219_cast_fp16")]; + tensor var_2535 = const()[name = tensor("op_2535"), val = tensor([1, 1])]; + tensor var_2537 = const()[name = tensor("op_2537"), val = tensor([1, 1])]; + tensor x_283_pad_type_0 = const()[name = tensor("x_283_pad_type_0"), val = tensor("custom")]; + tensor x_283_pad_0 = const()[name = tensor("x_283_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_15_fc1_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(159706240))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(162983104))), name = tensor("layers_15_fc1_module_weight_to_fp16_palettized"), shape = tensor([5120, 1280, 1, 1])]; + tensor layers_15_fc1_module_bias_to_fp16 = const()[name = tensor("layers_15_fc1_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(162983232)))]; + tensor x_283_cast_fp16 = conv(bias = layers_15_fc1_module_bias_to_fp16, dilations = var_2537, groups = var_2415, pad = x_283_pad_0, pad_type = x_283_pad_type_0, strides = var_2535, weight = layers_15_fc1_module_weight_to_fp16_palettized, x = input_219_cast_fp16)[name = tensor("x_283_cast_fp16")]; + tensor layers_15_fc1_output_scale_to_fp16 = const()[name = tensor("layers_15_fc1_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(162993536)))]; + tensor input_221_cast_fp16 = mul(x = x_283_cast_fp16, y = layers_15_fc1_output_scale_to_fp16)[name = tensor("input_221_cast_fp16")]; + tensor x_285_mode_0 = const()[name = tensor("x_285_mode_0"), val = tensor("EXACT")]; + tensor x_285_cast_fp16 = gelu(mode = x_285_mode_0, x = input_221_cast_fp16)[name = tensor("x_285_cast_fp16")]; + tensor layers_15_fc2_input_shift_to_fp16 = const()[name = tensor("layers_15_fc2_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(163003840)))]; + tensor input_223_cast_fp16 = sub(x = x_285_cast_fp16, y = layers_15_fc2_input_shift_to_fp16)[name = tensor("input_223_cast_fp16")]; + tensor var_2548 = const()[name = tensor("op_2548"), val = tensor([1, 1])]; + tensor var_2550 = const()[name = tensor("op_2550"), val = tensor([1, 1])]; + tensor x_287_pad_type_0 = const()[name = tensor("x_287_pad_type_0"), val = tensor("custom")]; + tensor x_287_pad_0 = const()[name = tensor("x_287_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_15_fc2_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(163014144))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(166291008))), name = tensor("layers_15_fc2_module_weight_to_fp16_palettized"), shape = tensor([1280, 5120, 1, 1])]; + tensor layers_15_fc2_module_bias_to_fp16 = const()[name = tensor("layers_15_fc2_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(166291136)))]; + tensor x_287_cast_fp16 = conv(bias = layers_15_fc2_module_bias_to_fp16, dilations = var_2550, groups = var_2415, pad = x_287_pad_0, pad_type = x_287_pad_type_0, strides = var_2548, weight = layers_15_fc2_module_weight_to_fp16_palettized, x = input_223_cast_fp16)[name = tensor("x_287_cast_fp16")]; + tensor layers_15_fc2_output_scale_to_fp16 = const()[name = tensor("layers_15_fc2_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(166293760)))]; + tensor hidden_states_35_cast_fp16 = mul(x = x_287_cast_fp16, y = layers_15_fc2_output_scale_to_fp16)[name = tensor("hidden_states_35_cast_fp16")]; + tensor inputs_65_cast_fp16 = add(x = inputs_63_cast_fp16, y = hidden_states_35_cast_fp16)[name = tensor("inputs_65_cast_fp16")]; + tensor var_2562 = const()[name = tensor("op_2562"), val = tensor(3)]; + tensor var_2564 = const()[name = tensor("op_2564"), val = tensor(1)]; + tensor var_2565 = const()[name = tensor("op_2565"), val = tensor(true)]; + tensor var_2575 = const()[name = tensor("op_2575"), val = tensor([1])]; + tensor channels_mean_65_cast_fp16 = reduce_mean(axes = var_2575, keep_dims = var_2565, x = inputs_65_cast_fp16)[name = tensor("channels_mean_65_cast_fp16")]; + tensor zero_mean_65_cast_fp16 = sub(x = inputs_65_cast_fp16, y = channels_mean_65_cast_fp16)[name = tensor("zero_mean_65_cast_fp16")]; + tensor zero_mean_sq_65_cast_fp16 = mul(x = zero_mean_65_cast_fp16, y = zero_mean_65_cast_fp16)[name = tensor("zero_mean_sq_65_cast_fp16")]; + tensor var_2579 = const()[name = tensor("op_2579"), val = tensor([1])]; + tensor var_2580_cast_fp16 = reduce_mean(axes = var_2579, keep_dims = var_2565, x = zero_mean_sq_65_cast_fp16)[name = tensor("op_2580_cast_fp16")]; + tensor var_2581_to_fp16 = const()[name = tensor("op_2581_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_2582_cast_fp16 = add(x = var_2580_cast_fp16, y = var_2581_to_fp16)[name = tensor("op_2582_cast_fp16")]; + tensor denom_65_epsilon_0_to_fp16 = const()[name = tensor("denom_65_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_65_cast_fp16 = rsqrt(epsilon = denom_65_epsilon_0_to_fp16, x = var_2582_cast_fp16)[name = tensor("denom_65_cast_fp16")]; + tensor out_65_cast_fp16 = mul(x = zero_mean_65_cast_fp16, y = denom_65_cast_fp16)[name = tensor("out_65_cast_fp16")]; + tensor obj_65_gamma_0_to_fp16 = const()[name = tensor("obj_65_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(166296384)))]; + tensor obj_65_beta_0_to_fp16 = const()[name = tensor("obj_65_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(166299008)))]; + tensor obj_65_epsilon_0_to_fp16 = const()[name = tensor("obj_65_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_65_cast_fp16 = batch_norm(beta = obj_65_beta_0_to_fp16, epsilon = obj_65_epsilon_0_to_fp16, gamma = obj_65_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_65_cast_fp16)[name = tensor("obj_65_cast_fp16")]; + tensor layers_16_self_attn_q_proj_input_shift_to_fp16 = const()[name = tensor("layers_16_self_attn_q_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(166301632)))]; + tensor input_225_cast_fp16 = sub(x = obj_65_cast_fp16, y = layers_16_self_attn_q_proj_input_shift_to_fp16)[name = tensor("input_225_cast_fp16")]; + tensor var_2601 = const()[name = tensor("op_2601"), val = tensor([1, 1])]; + tensor var_2603 = const()[name = tensor("op_2603"), val = tensor([1, 1])]; + tensor x_289_pad_type_0 = const()[name = tensor("x_289_pad_type_0"), val = tensor("custom")]; + tensor x_289_pad_0 = const()[name = tensor("x_289_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_16_self_attn_q_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(166304256))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(167123520))), name = tensor("layers_16_self_attn_q_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_16_self_attn_q_proj_module_bias_to_fp16 = const()[name = tensor("layers_16_self_attn_q_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(167123648)))]; + tensor x_289_cast_fp16 = conv(bias = layers_16_self_attn_q_proj_module_bias_to_fp16, dilations = var_2603, groups = var_2564, pad = x_289_pad_0, pad_type = x_289_pad_type_0, strides = var_2601, weight = layers_16_self_attn_q_proj_module_weight_to_fp16_palettized, x = input_225_cast_fp16)[name = tensor("x_289_cast_fp16")]; + tensor layers_16_self_attn_q_proj_output_scale_to_fp16 = const()[name = tensor("layers_16_self_attn_q_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(167126272)))]; + tensor query_33_cast_fp16 = mul(x = x_289_cast_fp16, y = layers_16_self_attn_q_proj_output_scale_to_fp16)[name = tensor("query_33_cast_fp16")]; + tensor var_2613 = const()[name = tensor("op_2613"), val = tensor([1, 1])]; + tensor var_2615 = const()[name = tensor("op_2615"), val = tensor([1, 1])]; + tensor x_291_pad_type_0 = const()[name = tensor("x_291_pad_type_0"), val = tensor("custom")]; + tensor x_291_pad_0 = const()[name = tensor("x_291_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_16_self_attn_k_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(167128896))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(167948160))), name = tensor("layers_16_self_attn_k_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_16_self_attn_k_proj_module_bias_to_fp16 = const()[name = tensor("layers_16_self_attn_k_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(167948288)))]; + tensor x_291_cast_fp16 = conv(bias = layers_16_self_attn_k_proj_module_bias_to_fp16, dilations = var_2615, groups = var_2564, pad = x_291_pad_0, pad_type = x_291_pad_type_0, strides = var_2613, weight = layers_16_self_attn_k_proj_module_weight_to_fp16_palettized, x = input_225_cast_fp16)[name = tensor("x_291_cast_fp16")]; + tensor layers_16_self_attn_k_proj_output_scale_to_fp16 = const()[name = tensor("layers_16_self_attn_k_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(167950912)))]; + tensor key_33_cast_fp16 = mul(x = x_291_cast_fp16, y = layers_16_self_attn_k_proj_output_scale_to_fp16)[name = tensor("key_33_cast_fp16")]; + tensor var_2625 = const()[name = tensor("op_2625"), val = tensor([1, 1])]; + tensor var_2627 = const()[name = tensor("op_2627"), val = tensor([1, 1])]; + tensor x_293_pad_type_0 = const()[name = tensor("x_293_pad_type_0"), val = tensor("custom")]; + tensor x_293_pad_0 = const()[name = tensor("x_293_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_16_self_attn_v_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(167953536))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(168772800))), name = tensor("layers_16_self_attn_v_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_16_self_attn_v_proj_module_bias_to_fp16 = const()[name = tensor("layers_16_self_attn_v_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(168772928)))]; + tensor x_293_cast_fp16 = conv(bias = layers_16_self_attn_v_proj_module_bias_to_fp16, dilations = var_2627, groups = var_2564, pad = x_293_pad_0, pad_type = x_293_pad_type_0, strides = var_2625, weight = layers_16_self_attn_v_proj_module_weight_to_fp16_palettized, x = input_225_cast_fp16)[name = tensor("x_293_cast_fp16")]; + tensor layers_16_self_attn_v_proj_output_scale_to_fp16 = const()[name = tensor("layers_16_self_attn_v_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(168775552)))]; + tensor value_33_cast_fp16 = mul(x = x_293_cast_fp16, y = layers_16_self_attn_v_proj_output_scale_to_fp16)[name = tensor("value_33_cast_fp16")]; + tensor var_2632 = const()[name = tensor("op_2632"), val = tensor([1, 20, 64, -1])]; + tensor var_2633_cast_fp16 = reshape(shape = var_2632, x = query_33_cast_fp16)[name = tensor("op_2633_cast_fp16")]; + tensor var_2634_to_fp16 = const()[name = tensor("op_2634_to_fp16"), val = tensor(0x1p-3)]; + tensor var_2635_cast_fp16 = mul(x = var_2633_cast_fp16, y = var_2634_to_fp16)[name = tensor("op_2635_cast_fp16")]; + tensor var_2636 = const()[name = tensor("op_2636"), val = tensor([1, 20, 64, -1])]; + tensor var_2637_cast_fp16 = reshape(shape = var_2636, x = key_33_cast_fp16)[name = tensor("op_2637_cast_fp16")]; + tensor mh_w_33_transpose_x_0 = const()[name = tensor("mh_w_33_transpose_x_0"), val = tensor(true)]; + tensor mh_w_33_transpose_y_0 = const()[name = tensor("mh_w_33_transpose_y_0"), val = tensor(false)]; + tensor mh_w_33_cast_fp16 = matmul(transpose_x = mh_w_33_transpose_x_0, transpose_y = mh_w_33_transpose_y_0, x = var_2635_cast_fp16, y = var_2637_cast_fp16)[name = tensor("mh_w_33_cast_fp16")]; + tensor var_2640_cast_fp16 = softmax(axis = var_2562, x = mh_w_33_cast_fp16)[name = tensor("op_2640_cast_fp16")]; + tensor var_2641 = const()[name = tensor("op_2641"), val = tensor([1, 20, 64, -1])]; + tensor var_2642_cast_fp16 = reshape(shape = var_2641, x = value_33_cast_fp16)[name = tensor("op_2642_cast_fp16")]; + tensor attn_33_transpose_x_0 = const()[name = tensor("attn_33_transpose_x_0"), val = tensor(false)]; + tensor attn_33_transpose_y_0 = const()[name = tensor("attn_33_transpose_y_0"), val = tensor(true)]; + tensor attn_33_cast_fp16 = matmul(transpose_x = attn_33_transpose_x_0, transpose_y = attn_33_transpose_y_0, x = var_2642_cast_fp16, y = var_2640_cast_fp16)[name = tensor("attn_33_cast_fp16")]; + tensor var_2645 = const()[name = tensor("op_2645"), val = tensor([1, 1280, 1, -1])]; + tensor x_295_cast_fp16 = reshape(shape = var_2645, x = attn_33_cast_fp16)[name = tensor("x_295_cast_fp16")]; + tensor layers_16_self_attn_o_proj_input_shift_to_fp16 = const()[name = tensor("layers_16_self_attn_o_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(168778176)))]; + tensor input_231_cast_fp16 = sub(x = x_295_cast_fp16, y = layers_16_self_attn_o_proj_input_shift_to_fp16)[name = tensor("input_231_cast_fp16")]; + tensor var_2653 = const()[name = tensor("op_2653"), val = tensor([1, 1])]; + tensor var_2655 = const()[name = tensor("op_2655"), val = tensor([1, 1])]; + tensor x_297_pad_type_0 = const()[name = tensor("x_297_pad_type_0"), val = tensor("custom")]; + tensor x_297_pad_0 = const()[name = tensor("x_297_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_16_self_attn_o_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(168780800))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(169600064))), name = tensor("layers_16_self_attn_o_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_16_self_attn_o_proj_module_bias_to_fp16 = const()[name = tensor("layers_16_self_attn_o_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(169600192)))]; + tensor x_297_cast_fp16 = conv(bias = layers_16_self_attn_o_proj_module_bias_to_fp16, dilations = var_2655, groups = var_2564, pad = x_297_pad_0, pad_type = x_297_pad_type_0, strides = var_2653, weight = layers_16_self_attn_o_proj_module_weight_to_fp16_palettized, x = input_231_cast_fp16)[name = tensor("x_297_cast_fp16")]; + tensor layers_16_self_attn_o_proj_output_scale_to_fp16 = const()[name = tensor("layers_16_self_attn_o_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(169602816)))]; + tensor obj_67_cast_fp16 = mul(x = x_297_cast_fp16, y = layers_16_self_attn_o_proj_output_scale_to_fp16)[name = tensor("obj_67_cast_fp16")]; + tensor inputs_67_cast_fp16 = add(x = inputs_65_cast_fp16, y = obj_67_cast_fp16)[name = tensor("inputs_67_cast_fp16")]; + tensor var_2662 = const()[name = tensor("op_2662"), val = tensor([1])]; + tensor channels_mean_67_cast_fp16 = reduce_mean(axes = var_2662, keep_dims = var_2565, x = inputs_67_cast_fp16)[name = tensor("channels_mean_67_cast_fp16")]; + tensor zero_mean_67_cast_fp16 = sub(x = inputs_67_cast_fp16, y = channels_mean_67_cast_fp16)[name = tensor("zero_mean_67_cast_fp16")]; + tensor zero_mean_sq_67_cast_fp16 = mul(x = zero_mean_67_cast_fp16, y = zero_mean_67_cast_fp16)[name = tensor("zero_mean_sq_67_cast_fp16")]; + tensor var_2666 = const()[name = tensor("op_2666"), val = tensor([1])]; + tensor var_2667_cast_fp16 = reduce_mean(axes = var_2666, keep_dims = var_2565, x = zero_mean_sq_67_cast_fp16)[name = tensor("op_2667_cast_fp16")]; + tensor var_2668_to_fp16 = const()[name = tensor("op_2668_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_2669_cast_fp16 = add(x = var_2667_cast_fp16, y = var_2668_to_fp16)[name = tensor("op_2669_cast_fp16")]; + tensor denom_67_epsilon_0_to_fp16 = const()[name = tensor("denom_67_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_67_cast_fp16 = rsqrt(epsilon = denom_67_epsilon_0_to_fp16, x = var_2669_cast_fp16)[name = tensor("denom_67_cast_fp16")]; + tensor out_67_cast_fp16 = mul(x = zero_mean_67_cast_fp16, y = denom_67_cast_fp16)[name = tensor("out_67_cast_fp16")]; + tensor x_299_gamma_0_to_fp16 = const()[name = tensor("x_299_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(169605440)))]; + tensor x_299_beta_0_to_fp16 = const()[name = tensor("x_299_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(169608064)))]; + tensor x_299_epsilon_0_to_fp16 = const()[name = tensor("x_299_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_299_cast_fp16 = batch_norm(beta = x_299_beta_0_to_fp16, epsilon = x_299_epsilon_0_to_fp16, gamma = x_299_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_67_cast_fp16)[name = tensor("x_299_cast_fp16")]; + tensor layers_16_fc1_input_shift_to_fp16 = const()[name = tensor("layers_16_fc1_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(169610688)))]; + tensor input_233_cast_fp16 = sub(x = x_299_cast_fp16, y = layers_16_fc1_input_shift_to_fp16)[name = tensor("input_233_cast_fp16")]; + tensor var_2684 = const()[name = tensor("op_2684"), val = tensor([1, 1])]; + tensor var_2686 = const()[name = tensor("op_2686"), val = tensor([1, 1])]; + tensor x_301_pad_type_0 = const()[name = tensor("x_301_pad_type_0"), val = tensor("custom")]; + tensor x_301_pad_0 = const()[name = tensor("x_301_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_16_fc1_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(169613312))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(172890176))), name = tensor("layers_16_fc1_module_weight_to_fp16_palettized"), shape = tensor([5120, 1280, 1, 1])]; + tensor layers_16_fc1_module_bias_to_fp16 = const()[name = tensor("layers_16_fc1_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(172890304)))]; + tensor x_301_cast_fp16 = conv(bias = layers_16_fc1_module_bias_to_fp16, dilations = var_2686, groups = var_2564, pad = x_301_pad_0, pad_type = x_301_pad_type_0, strides = var_2684, weight = layers_16_fc1_module_weight_to_fp16_palettized, x = input_233_cast_fp16)[name = tensor("x_301_cast_fp16")]; + tensor layers_16_fc1_output_scale_to_fp16 = const()[name = tensor("layers_16_fc1_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(172900608)))]; + tensor input_235_cast_fp16 = mul(x = x_301_cast_fp16, y = layers_16_fc1_output_scale_to_fp16)[name = tensor("input_235_cast_fp16")]; + tensor x_303_mode_0 = const()[name = tensor("x_303_mode_0"), val = tensor("EXACT")]; + tensor x_303_cast_fp16 = gelu(mode = x_303_mode_0, x = input_235_cast_fp16)[name = tensor("x_303_cast_fp16")]; + tensor layers_16_fc2_input_shift_to_fp16 = const()[name = tensor("layers_16_fc2_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(172910912)))]; + tensor input_237_cast_fp16 = sub(x = x_303_cast_fp16, y = layers_16_fc2_input_shift_to_fp16)[name = tensor("input_237_cast_fp16")]; + tensor var_2697 = const()[name = tensor("op_2697"), val = tensor([1, 1])]; + tensor var_2699 = const()[name = tensor("op_2699"), val = tensor([1, 1])]; + tensor x_305_pad_type_0 = const()[name = tensor("x_305_pad_type_0"), val = tensor("custom")]; + tensor x_305_pad_0 = const()[name = tensor("x_305_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_16_fc2_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(172921216))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(176198080))), name = tensor("layers_16_fc2_module_weight_to_fp16_palettized"), shape = tensor([1280, 5120, 1, 1])]; + tensor layers_16_fc2_module_bias_to_fp16 = const()[name = tensor("layers_16_fc2_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(176198208)))]; + tensor x_305_cast_fp16 = conv(bias = layers_16_fc2_module_bias_to_fp16, dilations = var_2699, groups = var_2564, pad = x_305_pad_0, pad_type = x_305_pad_type_0, strides = var_2697, weight = layers_16_fc2_module_weight_to_fp16_palettized, x = input_237_cast_fp16)[name = tensor("x_305_cast_fp16")]; + tensor layers_16_fc2_output_scale_to_fp16 = const()[name = tensor("layers_16_fc2_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(176200832)))]; + tensor hidden_states_37_cast_fp16 = mul(x = x_305_cast_fp16, y = layers_16_fc2_output_scale_to_fp16)[name = tensor("hidden_states_37_cast_fp16")]; + tensor inputs_69_cast_fp16 = add(x = inputs_67_cast_fp16, y = hidden_states_37_cast_fp16)[name = tensor("inputs_69_cast_fp16")]; + tensor var_2711 = const()[name = tensor("op_2711"), val = tensor(3)]; + tensor var_2713 = const()[name = tensor("op_2713"), val = tensor(1)]; + tensor var_2714 = const()[name = tensor("op_2714"), val = tensor(true)]; + tensor var_2724 = const()[name = tensor("op_2724"), val = tensor([1])]; + tensor channels_mean_69_cast_fp16 = reduce_mean(axes = var_2724, keep_dims = var_2714, x = inputs_69_cast_fp16)[name = tensor("channels_mean_69_cast_fp16")]; + tensor zero_mean_69_cast_fp16 = sub(x = inputs_69_cast_fp16, y = channels_mean_69_cast_fp16)[name = tensor("zero_mean_69_cast_fp16")]; + tensor zero_mean_sq_69_cast_fp16 = mul(x = zero_mean_69_cast_fp16, y = zero_mean_69_cast_fp16)[name = tensor("zero_mean_sq_69_cast_fp16")]; + tensor var_2728 = const()[name = tensor("op_2728"), val = tensor([1])]; + tensor var_2729_cast_fp16 = reduce_mean(axes = var_2728, keep_dims = var_2714, x = zero_mean_sq_69_cast_fp16)[name = tensor("op_2729_cast_fp16")]; + tensor var_2730_to_fp16 = const()[name = tensor("op_2730_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_2731_cast_fp16 = add(x = var_2729_cast_fp16, y = var_2730_to_fp16)[name = tensor("op_2731_cast_fp16")]; + tensor denom_69_epsilon_0_to_fp16 = const()[name = tensor("denom_69_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_69_cast_fp16 = rsqrt(epsilon = denom_69_epsilon_0_to_fp16, x = var_2731_cast_fp16)[name = tensor("denom_69_cast_fp16")]; + tensor out_69_cast_fp16 = mul(x = zero_mean_69_cast_fp16, y = denom_69_cast_fp16)[name = tensor("out_69_cast_fp16")]; + tensor obj_69_gamma_0_to_fp16 = const()[name = tensor("obj_69_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(176203456)))]; + tensor obj_69_beta_0_to_fp16 = const()[name = tensor("obj_69_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(176206080)))]; + tensor obj_69_epsilon_0_to_fp16 = const()[name = tensor("obj_69_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_69_cast_fp16 = batch_norm(beta = obj_69_beta_0_to_fp16, epsilon = obj_69_epsilon_0_to_fp16, gamma = obj_69_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_69_cast_fp16)[name = tensor("obj_69_cast_fp16")]; + tensor layers_17_self_attn_q_proj_input_shift_to_fp16 = const()[name = tensor("layers_17_self_attn_q_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(176208704)))]; + tensor input_239_cast_fp16 = sub(x = obj_69_cast_fp16, y = layers_17_self_attn_q_proj_input_shift_to_fp16)[name = tensor("input_239_cast_fp16")]; + tensor var_2750 = const()[name = tensor("op_2750"), val = tensor([1, 1])]; + tensor var_2752 = const()[name = tensor("op_2752"), val = tensor([1, 1])]; + tensor x_307_pad_type_0 = const()[name = tensor("x_307_pad_type_0"), val = tensor("custom")]; + tensor x_307_pad_0 = const()[name = tensor("x_307_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_17_self_attn_q_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(176211328))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(177030592))), name = tensor("layers_17_self_attn_q_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_17_self_attn_q_proj_module_bias_to_fp16 = const()[name = tensor("layers_17_self_attn_q_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(177030720)))]; + tensor x_307_cast_fp16 = conv(bias = layers_17_self_attn_q_proj_module_bias_to_fp16, dilations = var_2752, groups = var_2713, pad = x_307_pad_0, pad_type = x_307_pad_type_0, strides = var_2750, weight = layers_17_self_attn_q_proj_module_weight_to_fp16_palettized, x = input_239_cast_fp16)[name = tensor("x_307_cast_fp16")]; + tensor layers_17_self_attn_q_proj_output_scale_to_fp16 = const()[name = tensor("layers_17_self_attn_q_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(177033344)))]; + tensor query_35_cast_fp16 = mul(x = x_307_cast_fp16, y = layers_17_self_attn_q_proj_output_scale_to_fp16)[name = tensor("query_35_cast_fp16")]; + tensor var_2762 = const()[name = tensor("op_2762"), val = tensor([1, 1])]; + tensor var_2764 = const()[name = tensor("op_2764"), val = tensor([1, 1])]; + tensor x_309_pad_type_0 = const()[name = tensor("x_309_pad_type_0"), val = tensor("custom")]; + tensor x_309_pad_0 = const()[name = tensor("x_309_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_17_self_attn_k_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(177035968))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(177855232))), name = tensor("layers_17_self_attn_k_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_17_self_attn_k_proj_module_bias_to_fp16 = const()[name = tensor("layers_17_self_attn_k_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(177855360)))]; + tensor x_309_cast_fp16 = conv(bias = layers_17_self_attn_k_proj_module_bias_to_fp16, dilations = var_2764, groups = var_2713, pad = x_309_pad_0, pad_type = x_309_pad_type_0, strides = var_2762, weight = layers_17_self_attn_k_proj_module_weight_to_fp16_palettized, x = input_239_cast_fp16)[name = tensor("x_309_cast_fp16")]; + tensor layers_17_self_attn_k_proj_output_scale_to_fp16 = const()[name = tensor("layers_17_self_attn_k_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(177857984)))]; + tensor key_35_cast_fp16 = mul(x = x_309_cast_fp16, y = layers_17_self_attn_k_proj_output_scale_to_fp16)[name = tensor("key_35_cast_fp16")]; + tensor var_2774 = const()[name = tensor("op_2774"), val = tensor([1, 1])]; + tensor var_2776 = const()[name = tensor("op_2776"), val = tensor([1, 1])]; + tensor x_311_pad_type_0 = const()[name = tensor("x_311_pad_type_0"), val = tensor("custom")]; + tensor x_311_pad_0 = const()[name = tensor("x_311_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_17_self_attn_v_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(177860608))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(178679872))), name = tensor("layers_17_self_attn_v_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_17_self_attn_v_proj_module_bias_to_fp16 = const()[name = tensor("layers_17_self_attn_v_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(178680000)))]; + tensor x_311_cast_fp16 = conv(bias = layers_17_self_attn_v_proj_module_bias_to_fp16, dilations = var_2776, groups = var_2713, pad = x_311_pad_0, pad_type = x_311_pad_type_0, strides = var_2774, weight = layers_17_self_attn_v_proj_module_weight_to_fp16_palettized, x = input_239_cast_fp16)[name = tensor("x_311_cast_fp16")]; + tensor layers_17_self_attn_v_proj_output_scale_to_fp16 = const()[name = tensor("layers_17_self_attn_v_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(178682624)))]; + tensor value_35_cast_fp16 = mul(x = x_311_cast_fp16, y = layers_17_self_attn_v_proj_output_scale_to_fp16)[name = tensor("value_35_cast_fp16")]; + tensor var_2781 = const()[name = tensor("op_2781"), val = tensor([1, 20, 64, -1])]; + tensor var_2782_cast_fp16 = reshape(shape = var_2781, x = query_35_cast_fp16)[name = tensor("op_2782_cast_fp16")]; + tensor var_2783_to_fp16 = const()[name = tensor("op_2783_to_fp16"), val = tensor(0x1p-3)]; + tensor var_2784_cast_fp16 = mul(x = var_2782_cast_fp16, y = var_2783_to_fp16)[name = tensor("op_2784_cast_fp16")]; + tensor var_2785 = const()[name = tensor("op_2785"), val = tensor([1, 20, 64, -1])]; + tensor var_2786_cast_fp16 = reshape(shape = var_2785, x = key_35_cast_fp16)[name = tensor("op_2786_cast_fp16")]; + tensor mh_w_35_transpose_x_0 = const()[name = tensor("mh_w_35_transpose_x_0"), val = tensor(true)]; + tensor mh_w_35_transpose_y_0 = const()[name = tensor("mh_w_35_transpose_y_0"), val = tensor(false)]; + tensor mh_w_35_cast_fp16 = matmul(transpose_x = mh_w_35_transpose_x_0, transpose_y = mh_w_35_transpose_y_0, x = var_2784_cast_fp16, y = var_2786_cast_fp16)[name = tensor("mh_w_35_cast_fp16")]; + tensor var_2789_cast_fp16 = softmax(axis = var_2711, x = mh_w_35_cast_fp16)[name = tensor("op_2789_cast_fp16")]; + tensor var_2790 = const()[name = tensor("op_2790"), val = tensor([1, 20, 64, -1])]; + tensor var_2791_cast_fp16 = reshape(shape = var_2790, x = value_35_cast_fp16)[name = tensor("op_2791_cast_fp16")]; + tensor attn_35_transpose_x_0 = const()[name = tensor("attn_35_transpose_x_0"), val = tensor(false)]; + tensor attn_35_transpose_y_0 = const()[name = tensor("attn_35_transpose_y_0"), val = tensor(true)]; + tensor attn_35_cast_fp16 = matmul(transpose_x = attn_35_transpose_x_0, transpose_y = attn_35_transpose_y_0, x = var_2791_cast_fp16, y = var_2789_cast_fp16)[name = tensor("attn_35_cast_fp16")]; + tensor var_2794 = const()[name = tensor("op_2794"), val = tensor([1, 1280, 1, -1])]; + tensor x_313_cast_fp16 = reshape(shape = var_2794, x = attn_35_cast_fp16)[name = tensor("x_313_cast_fp16")]; + tensor layers_17_self_attn_o_proj_input_shift_to_fp16 = const()[name = tensor("layers_17_self_attn_o_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(178685248)))]; + tensor input_245_cast_fp16 = sub(x = x_313_cast_fp16, y = layers_17_self_attn_o_proj_input_shift_to_fp16)[name = tensor("input_245_cast_fp16")]; + tensor var_2802 = const()[name = tensor("op_2802"), val = tensor([1, 1])]; + tensor var_2804 = const()[name = tensor("op_2804"), val = tensor([1, 1])]; + tensor x_315_pad_type_0 = const()[name = tensor("x_315_pad_type_0"), val = tensor("custom")]; + tensor x_315_pad_0 = const()[name = tensor("x_315_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_17_self_attn_o_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(178687872))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(179507136))), name = tensor("layers_17_self_attn_o_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_17_self_attn_o_proj_module_bias_to_fp16 = const()[name = tensor("layers_17_self_attn_o_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(179507264)))]; + tensor x_315_cast_fp16 = conv(bias = layers_17_self_attn_o_proj_module_bias_to_fp16, dilations = var_2804, groups = var_2713, pad = x_315_pad_0, pad_type = x_315_pad_type_0, strides = var_2802, weight = layers_17_self_attn_o_proj_module_weight_to_fp16_palettized, x = input_245_cast_fp16)[name = tensor("x_315_cast_fp16")]; + tensor layers_17_self_attn_o_proj_output_scale_to_fp16 = const()[name = tensor("layers_17_self_attn_o_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(179509888)))]; + tensor obj_71_cast_fp16 = mul(x = x_315_cast_fp16, y = layers_17_self_attn_o_proj_output_scale_to_fp16)[name = tensor("obj_71_cast_fp16")]; + tensor inputs_71_cast_fp16 = add(x = inputs_69_cast_fp16, y = obj_71_cast_fp16)[name = tensor("inputs_71_cast_fp16")]; + tensor var_2811 = const()[name = tensor("op_2811"), val = tensor([1])]; + tensor channels_mean_71_cast_fp16 = reduce_mean(axes = var_2811, keep_dims = var_2714, x = inputs_71_cast_fp16)[name = tensor("channels_mean_71_cast_fp16")]; + tensor zero_mean_71_cast_fp16 = sub(x = inputs_71_cast_fp16, y = channels_mean_71_cast_fp16)[name = tensor("zero_mean_71_cast_fp16")]; + tensor zero_mean_sq_71_cast_fp16 = mul(x = zero_mean_71_cast_fp16, y = zero_mean_71_cast_fp16)[name = tensor("zero_mean_sq_71_cast_fp16")]; + tensor var_2815 = const()[name = tensor("op_2815"), val = tensor([1])]; + tensor var_2816_cast_fp16 = reduce_mean(axes = var_2815, keep_dims = var_2714, x = zero_mean_sq_71_cast_fp16)[name = tensor("op_2816_cast_fp16")]; + tensor var_2817_to_fp16 = const()[name = tensor("op_2817_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_2818_cast_fp16 = add(x = var_2816_cast_fp16, y = var_2817_to_fp16)[name = tensor("op_2818_cast_fp16")]; + tensor denom_71_epsilon_0_to_fp16 = const()[name = tensor("denom_71_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_71_cast_fp16 = rsqrt(epsilon = denom_71_epsilon_0_to_fp16, x = var_2818_cast_fp16)[name = tensor("denom_71_cast_fp16")]; + tensor out_71_cast_fp16 = mul(x = zero_mean_71_cast_fp16, y = denom_71_cast_fp16)[name = tensor("out_71_cast_fp16")]; + tensor x_317_gamma_0_to_fp16 = const()[name = tensor("x_317_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(179512512)))]; + tensor x_317_beta_0_to_fp16 = const()[name = tensor("x_317_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(179515136)))]; + tensor x_317_epsilon_0_to_fp16 = const()[name = tensor("x_317_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_317_cast_fp16 = batch_norm(beta = x_317_beta_0_to_fp16, epsilon = x_317_epsilon_0_to_fp16, gamma = x_317_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_71_cast_fp16)[name = tensor("x_317_cast_fp16")]; + tensor layers_17_fc1_input_shift_to_fp16 = const()[name = tensor("layers_17_fc1_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(179517760)))]; + tensor input_247_cast_fp16 = sub(x = x_317_cast_fp16, y = layers_17_fc1_input_shift_to_fp16)[name = tensor("input_247_cast_fp16")]; + tensor var_2833 = const()[name = tensor("op_2833"), val = tensor([1, 1])]; + tensor var_2835 = const()[name = tensor("op_2835"), val = tensor([1, 1])]; + tensor x_319_pad_type_0 = const()[name = tensor("x_319_pad_type_0"), val = tensor("custom")]; + tensor x_319_pad_0 = const()[name = tensor("x_319_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_17_fc1_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(179520384))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(182797248))), name = tensor("layers_17_fc1_module_weight_to_fp16_palettized"), shape = tensor([5120, 1280, 1, 1])]; + tensor layers_17_fc1_module_bias_to_fp16 = const()[name = tensor("layers_17_fc1_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(182797376)))]; + tensor x_319_cast_fp16 = conv(bias = layers_17_fc1_module_bias_to_fp16, dilations = var_2835, groups = var_2713, pad = x_319_pad_0, pad_type = x_319_pad_type_0, strides = var_2833, weight = layers_17_fc1_module_weight_to_fp16_palettized, x = input_247_cast_fp16)[name = tensor("x_319_cast_fp16")]; + tensor layers_17_fc1_output_scale_to_fp16 = const()[name = tensor("layers_17_fc1_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(182807680)))]; + tensor input_249_cast_fp16 = mul(x = x_319_cast_fp16, y = layers_17_fc1_output_scale_to_fp16)[name = tensor("input_249_cast_fp16")]; + tensor x_321_mode_0 = const()[name = tensor("x_321_mode_0"), val = tensor("EXACT")]; + tensor x_321_cast_fp16 = gelu(mode = x_321_mode_0, x = input_249_cast_fp16)[name = tensor("x_321_cast_fp16")]; + tensor layers_17_fc2_input_shift_to_fp16 = const()[name = tensor("layers_17_fc2_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(182817984)))]; + tensor input_251_cast_fp16 = sub(x = x_321_cast_fp16, y = layers_17_fc2_input_shift_to_fp16)[name = tensor("input_251_cast_fp16")]; + tensor var_2846 = const()[name = tensor("op_2846"), val = tensor([1, 1])]; + tensor var_2848 = const()[name = tensor("op_2848"), val = tensor([1, 1])]; + tensor x_323_pad_type_0 = const()[name = tensor("x_323_pad_type_0"), val = tensor("custom")]; + tensor x_323_pad_0 = const()[name = tensor("x_323_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_17_fc2_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(182828288))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(186105152))), name = tensor("layers_17_fc2_module_weight_to_fp16_palettized"), shape = tensor([1280, 5120, 1, 1])]; + tensor layers_17_fc2_module_bias_to_fp16 = const()[name = tensor("layers_17_fc2_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(186105280)))]; + tensor x_323_cast_fp16 = conv(bias = layers_17_fc2_module_bias_to_fp16, dilations = var_2848, groups = var_2713, pad = x_323_pad_0, pad_type = x_323_pad_type_0, strides = var_2846, weight = layers_17_fc2_module_weight_to_fp16_palettized, x = input_251_cast_fp16)[name = tensor("x_323_cast_fp16")]; + tensor layers_17_fc2_output_scale_to_fp16 = const()[name = tensor("layers_17_fc2_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(186107904)))]; + tensor hidden_states_39_cast_fp16 = mul(x = x_323_cast_fp16, y = layers_17_fc2_output_scale_to_fp16)[name = tensor("hidden_states_39_cast_fp16")]; + tensor inputs_73_cast_fp16 = add(x = inputs_71_cast_fp16, y = hidden_states_39_cast_fp16)[name = tensor("inputs_73_cast_fp16")]; + tensor var_2860 = const()[name = tensor("op_2860"), val = tensor(3)]; + tensor var_2862 = const()[name = tensor("op_2862"), val = tensor(1)]; + tensor var_2863 = const()[name = tensor("op_2863"), val = tensor(true)]; + tensor var_2873 = const()[name = tensor("op_2873"), val = tensor([1])]; + tensor channels_mean_73_cast_fp16 = reduce_mean(axes = var_2873, keep_dims = var_2863, x = inputs_73_cast_fp16)[name = tensor("channels_mean_73_cast_fp16")]; + tensor zero_mean_73_cast_fp16 = sub(x = inputs_73_cast_fp16, y = channels_mean_73_cast_fp16)[name = tensor("zero_mean_73_cast_fp16")]; + tensor zero_mean_sq_73_cast_fp16 = mul(x = zero_mean_73_cast_fp16, y = zero_mean_73_cast_fp16)[name = tensor("zero_mean_sq_73_cast_fp16")]; + tensor var_2877 = const()[name = tensor("op_2877"), val = tensor([1])]; + tensor var_2878_cast_fp16 = reduce_mean(axes = var_2877, keep_dims = var_2863, x = zero_mean_sq_73_cast_fp16)[name = tensor("op_2878_cast_fp16")]; + tensor var_2879_to_fp16 = const()[name = tensor("op_2879_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_2880_cast_fp16 = add(x = var_2878_cast_fp16, y = var_2879_to_fp16)[name = tensor("op_2880_cast_fp16")]; + tensor denom_73_epsilon_0_to_fp16 = const()[name = tensor("denom_73_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_73_cast_fp16 = rsqrt(epsilon = denom_73_epsilon_0_to_fp16, x = var_2880_cast_fp16)[name = tensor("denom_73_cast_fp16")]; + tensor out_73_cast_fp16 = mul(x = zero_mean_73_cast_fp16, y = denom_73_cast_fp16)[name = tensor("out_73_cast_fp16")]; + tensor obj_73_gamma_0_to_fp16 = const()[name = tensor("obj_73_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(186110528)))]; + tensor obj_73_beta_0_to_fp16 = const()[name = tensor("obj_73_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(186113152)))]; + tensor obj_73_epsilon_0_to_fp16 = const()[name = tensor("obj_73_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_73_cast_fp16 = batch_norm(beta = obj_73_beta_0_to_fp16, epsilon = obj_73_epsilon_0_to_fp16, gamma = obj_73_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_73_cast_fp16)[name = tensor("obj_73_cast_fp16")]; + tensor layers_18_self_attn_q_proj_input_shift_to_fp16 = const()[name = tensor("layers_18_self_attn_q_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(186115776)))]; + tensor input_253_cast_fp16 = sub(x = obj_73_cast_fp16, y = layers_18_self_attn_q_proj_input_shift_to_fp16)[name = tensor("input_253_cast_fp16")]; + tensor var_2899 = const()[name = tensor("op_2899"), val = tensor([1, 1])]; + tensor var_2901 = const()[name = tensor("op_2901"), val = tensor([1, 1])]; + tensor x_325_pad_type_0 = const()[name = tensor("x_325_pad_type_0"), val = tensor("custom")]; + tensor x_325_pad_0 = const()[name = tensor("x_325_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_18_self_attn_q_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(186118400))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(186937664))), name = tensor("layers_18_self_attn_q_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_18_self_attn_q_proj_module_bias_to_fp16 = const()[name = tensor("layers_18_self_attn_q_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(186937792)))]; + tensor x_325_cast_fp16 = conv(bias = layers_18_self_attn_q_proj_module_bias_to_fp16, dilations = var_2901, groups = var_2862, pad = x_325_pad_0, pad_type = x_325_pad_type_0, strides = var_2899, weight = layers_18_self_attn_q_proj_module_weight_to_fp16_palettized, x = input_253_cast_fp16)[name = tensor("x_325_cast_fp16")]; + tensor layers_18_self_attn_q_proj_output_scale_to_fp16 = const()[name = tensor("layers_18_self_attn_q_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(186940416)))]; + tensor query_37_cast_fp16 = mul(x = x_325_cast_fp16, y = layers_18_self_attn_q_proj_output_scale_to_fp16)[name = tensor("query_37_cast_fp16")]; + tensor var_2911 = const()[name = tensor("op_2911"), val = tensor([1, 1])]; + tensor var_2913 = const()[name = tensor("op_2913"), val = tensor([1, 1])]; + tensor x_327_pad_type_0 = const()[name = tensor("x_327_pad_type_0"), val = tensor("custom")]; + tensor x_327_pad_0 = const()[name = tensor("x_327_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_18_self_attn_k_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(186943040))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(187762304))), name = tensor("layers_18_self_attn_k_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_18_self_attn_k_proj_module_bias_to_fp16 = const()[name = tensor("layers_18_self_attn_k_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(187762432)))]; + tensor x_327_cast_fp16 = conv(bias = layers_18_self_attn_k_proj_module_bias_to_fp16, dilations = var_2913, groups = var_2862, pad = x_327_pad_0, pad_type = x_327_pad_type_0, strides = var_2911, weight = layers_18_self_attn_k_proj_module_weight_to_fp16_palettized, x = input_253_cast_fp16)[name = tensor("x_327_cast_fp16")]; + tensor layers_18_self_attn_k_proj_output_scale_to_fp16 = const()[name = tensor("layers_18_self_attn_k_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(187765056)))]; + tensor key_37_cast_fp16 = mul(x = x_327_cast_fp16, y = layers_18_self_attn_k_proj_output_scale_to_fp16)[name = tensor("key_37_cast_fp16")]; + tensor var_2923 = const()[name = tensor("op_2923"), val = tensor([1, 1])]; + tensor var_2925 = const()[name = tensor("op_2925"), val = tensor([1, 1])]; + tensor x_329_pad_type_0 = const()[name = tensor("x_329_pad_type_0"), val = tensor("custom")]; + tensor x_329_pad_0 = const()[name = tensor("x_329_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_18_self_attn_v_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(187767680))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(188586944))), name = tensor("layers_18_self_attn_v_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_18_self_attn_v_proj_module_bias_to_fp16 = const()[name = tensor("layers_18_self_attn_v_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(188587072)))]; + tensor x_329_cast_fp16 = conv(bias = layers_18_self_attn_v_proj_module_bias_to_fp16, dilations = var_2925, groups = var_2862, pad = x_329_pad_0, pad_type = x_329_pad_type_0, strides = var_2923, weight = layers_18_self_attn_v_proj_module_weight_to_fp16_palettized, x = input_253_cast_fp16)[name = tensor("x_329_cast_fp16")]; + tensor layers_18_self_attn_v_proj_output_scale_to_fp16 = const()[name = tensor("layers_18_self_attn_v_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(188589696)))]; + tensor value_37_cast_fp16 = mul(x = x_329_cast_fp16, y = layers_18_self_attn_v_proj_output_scale_to_fp16)[name = tensor("value_37_cast_fp16")]; + tensor var_2930 = const()[name = tensor("op_2930"), val = tensor([1, 20, 64, -1])]; + tensor var_2931_cast_fp16 = reshape(shape = var_2930, x = query_37_cast_fp16)[name = tensor("op_2931_cast_fp16")]; + tensor var_2932_to_fp16 = const()[name = tensor("op_2932_to_fp16"), val = tensor(0x1p-3)]; + tensor var_2933_cast_fp16 = mul(x = var_2931_cast_fp16, y = var_2932_to_fp16)[name = tensor("op_2933_cast_fp16")]; + tensor var_2934 = const()[name = tensor("op_2934"), val = tensor([1, 20, 64, -1])]; + tensor var_2935_cast_fp16 = reshape(shape = var_2934, x = key_37_cast_fp16)[name = tensor("op_2935_cast_fp16")]; + tensor mh_w_37_transpose_x_0 = const()[name = tensor("mh_w_37_transpose_x_0"), val = tensor(true)]; + tensor mh_w_37_transpose_y_0 = const()[name = tensor("mh_w_37_transpose_y_0"), val = tensor(false)]; + tensor mh_w_37_cast_fp16 = matmul(transpose_x = mh_w_37_transpose_x_0, transpose_y = mh_w_37_transpose_y_0, x = var_2933_cast_fp16, y = var_2935_cast_fp16)[name = tensor("mh_w_37_cast_fp16")]; + tensor var_2938_cast_fp16 = softmax(axis = var_2860, x = mh_w_37_cast_fp16)[name = tensor("op_2938_cast_fp16")]; + tensor var_2939 = const()[name = tensor("op_2939"), val = tensor([1, 20, 64, -1])]; + tensor var_2940_cast_fp16 = reshape(shape = var_2939, x = value_37_cast_fp16)[name = tensor("op_2940_cast_fp16")]; + tensor attn_37_transpose_x_0 = const()[name = tensor("attn_37_transpose_x_0"), val = tensor(false)]; + tensor attn_37_transpose_y_0 = const()[name = tensor("attn_37_transpose_y_0"), val = tensor(true)]; + tensor attn_37_cast_fp16 = matmul(transpose_x = attn_37_transpose_x_0, transpose_y = attn_37_transpose_y_0, x = var_2940_cast_fp16, y = var_2938_cast_fp16)[name = tensor("attn_37_cast_fp16")]; + tensor var_2943 = const()[name = tensor("op_2943"), val = tensor([1, 1280, 1, -1])]; + tensor x_331_cast_fp16 = reshape(shape = var_2943, x = attn_37_cast_fp16)[name = tensor("x_331_cast_fp16")]; + tensor layers_18_self_attn_o_proj_input_shift_to_fp16 = const()[name = tensor("layers_18_self_attn_o_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(188592320)))]; + tensor input_259_cast_fp16 = sub(x = x_331_cast_fp16, y = layers_18_self_attn_o_proj_input_shift_to_fp16)[name = tensor("input_259_cast_fp16")]; + tensor var_2951 = const()[name = tensor("op_2951"), val = tensor([1, 1])]; + tensor var_2953 = const()[name = tensor("op_2953"), val = tensor([1, 1])]; + tensor x_333_pad_type_0 = const()[name = tensor("x_333_pad_type_0"), val = tensor("custom")]; + tensor x_333_pad_0 = const()[name = tensor("x_333_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_18_self_attn_o_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(188594944))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(189414208))), name = tensor("layers_18_self_attn_o_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_18_self_attn_o_proj_module_bias_to_fp16 = const()[name = tensor("layers_18_self_attn_o_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(189414336)))]; + tensor x_333_cast_fp16 = conv(bias = layers_18_self_attn_o_proj_module_bias_to_fp16, dilations = var_2953, groups = var_2862, pad = x_333_pad_0, pad_type = x_333_pad_type_0, strides = var_2951, weight = layers_18_self_attn_o_proj_module_weight_to_fp16_palettized, x = input_259_cast_fp16)[name = tensor("x_333_cast_fp16")]; + tensor layers_18_self_attn_o_proj_output_scale_to_fp16 = const()[name = tensor("layers_18_self_attn_o_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(189416960)))]; + tensor obj_75_cast_fp16 = mul(x = x_333_cast_fp16, y = layers_18_self_attn_o_proj_output_scale_to_fp16)[name = tensor("obj_75_cast_fp16")]; + tensor inputs_75_cast_fp16 = add(x = inputs_73_cast_fp16, y = obj_75_cast_fp16)[name = tensor("inputs_75_cast_fp16")]; + tensor var_2960 = const()[name = tensor("op_2960"), val = tensor([1])]; + tensor channels_mean_75_cast_fp16 = reduce_mean(axes = var_2960, keep_dims = var_2863, x = inputs_75_cast_fp16)[name = tensor("channels_mean_75_cast_fp16")]; + tensor zero_mean_75_cast_fp16 = sub(x = inputs_75_cast_fp16, y = channels_mean_75_cast_fp16)[name = tensor("zero_mean_75_cast_fp16")]; + tensor zero_mean_sq_75_cast_fp16 = mul(x = zero_mean_75_cast_fp16, y = zero_mean_75_cast_fp16)[name = tensor("zero_mean_sq_75_cast_fp16")]; + tensor var_2964 = const()[name = tensor("op_2964"), val = tensor([1])]; + tensor var_2965_cast_fp16 = reduce_mean(axes = var_2964, keep_dims = var_2863, x = zero_mean_sq_75_cast_fp16)[name = tensor("op_2965_cast_fp16")]; + tensor var_2966_to_fp16 = const()[name = tensor("op_2966_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_2967_cast_fp16 = add(x = var_2965_cast_fp16, y = var_2966_to_fp16)[name = tensor("op_2967_cast_fp16")]; + tensor denom_75_epsilon_0_to_fp16 = const()[name = tensor("denom_75_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_75_cast_fp16 = rsqrt(epsilon = denom_75_epsilon_0_to_fp16, x = var_2967_cast_fp16)[name = tensor("denom_75_cast_fp16")]; + tensor out_75_cast_fp16 = mul(x = zero_mean_75_cast_fp16, y = denom_75_cast_fp16)[name = tensor("out_75_cast_fp16")]; + tensor x_335_gamma_0_to_fp16 = const()[name = tensor("x_335_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(189419584)))]; + tensor x_335_beta_0_to_fp16 = const()[name = tensor("x_335_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(189422208)))]; + tensor x_335_epsilon_0_to_fp16 = const()[name = tensor("x_335_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_335_cast_fp16 = batch_norm(beta = x_335_beta_0_to_fp16, epsilon = x_335_epsilon_0_to_fp16, gamma = x_335_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_75_cast_fp16)[name = tensor("x_335_cast_fp16")]; + tensor layers_18_fc1_input_shift_to_fp16 = const()[name = tensor("layers_18_fc1_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(189424832)))]; + tensor input_261_cast_fp16 = sub(x = x_335_cast_fp16, y = layers_18_fc1_input_shift_to_fp16)[name = tensor("input_261_cast_fp16")]; + tensor var_2982 = const()[name = tensor("op_2982"), val = tensor([1, 1])]; + tensor var_2984 = const()[name = tensor("op_2984"), val = tensor([1, 1])]; + tensor x_337_pad_type_0 = const()[name = tensor("x_337_pad_type_0"), val = tensor("custom")]; + tensor x_337_pad_0 = const()[name = tensor("x_337_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_18_fc1_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(189427456))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(192704320))), name = tensor("layers_18_fc1_module_weight_to_fp16_palettized"), shape = tensor([5120, 1280, 1, 1])]; + tensor layers_18_fc1_module_bias_to_fp16 = const()[name = tensor("layers_18_fc1_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(192704448)))]; + tensor x_337_cast_fp16 = conv(bias = layers_18_fc1_module_bias_to_fp16, dilations = var_2984, groups = var_2862, pad = x_337_pad_0, pad_type = x_337_pad_type_0, strides = var_2982, weight = layers_18_fc1_module_weight_to_fp16_palettized, x = input_261_cast_fp16)[name = tensor("x_337_cast_fp16")]; + tensor layers_18_fc1_output_scale_to_fp16 = const()[name = tensor("layers_18_fc1_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(192714752)))]; + tensor input_263_cast_fp16 = mul(x = x_337_cast_fp16, y = layers_18_fc1_output_scale_to_fp16)[name = tensor("input_263_cast_fp16")]; + tensor x_339_mode_0 = const()[name = tensor("x_339_mode_0"), val = tensor("EXACT")]; + tensor x_339_cast_fp16 = gelu(mode = x_339_mode_0, x = input_263_cast_fp16)[name = tensor("x_339_cast_fp16")]; + tensor layers_18_fc2_input_shift_to_fp16 = const()[name = tensor("layers_18_fc2_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(192725056)))]; + tensor input_265_cast_fp16 = sub(x = x_339_cast_fp16, y = layers_18_fc2_input_shift_to_fp16)[name = tensor("input_265_cast_fp16")]; + tensor var_2995 = const()[name = tensor("op_2995"), val = tensor([1, 1])]; + tensor var_2997 = const()[name = tensor("op_2997"), val = tensor([1, 1])]; + tensor x_341_pad_type_0 = const()[name = tensor("x_341_pad_type_0"), val = tensor("custom")]; + tensor x_341_pad_0 = const()[name = tensor("x_341_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_18_fc2_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(192735360))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(196012224))), name = tensor("layers_18_fc2_module_weight_to_fp16_palettized"), shape = tensor([1280, 5120, 1, 1])]; + tensor layers_18_fc2_module_bias_to_fp16 = const()[name = tensor("layers_18_fc2_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(196012352)))]; + tensor x_341_cast_fp16 = conv(bias = layers_18_fc2_module_bias_to_fp16, dilations = var_2997, groups = var_2862, pad = x_341_pad_0, pad_type = x_341_pad_type_0, strides = var_2995, weight = layers_18_fc2_module_weight_to_fp16_palettized, x = input_265_cast_fp16)[name = tensor("x_341_cast_fp16")]; + tensor layers_18_fc2_output_scale_to_fp16 = const()[name = tensor("layers_18_fc2_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(196014976)))]; + tensor hidden_states_41_cast_fp16 = mul(x = x_341_cast_fp16, y = layers_18_fc2_output_scale_to_fp16)[name = tensor("hidden_states_41_cast_fp16")]; + tensor inputs_77_cast_fp16 = add(x = inputs_75_cast_fp16, y = hidden_states_41_cast_fp16)[name = tensor("inputs_77_cast_fp16")]; + tensor var_3009 = const()[name = tensor("op_3009"), val = tensor(3)]; + tensor var_3011 = const()[name = tensor("op_3011"), val = tensor(1)]; + tensor var_3012 = const()[name = tensor("op_3012"), val = tensor(true)]; + tensor var_3022 = const()[name = tensor("op_3022"), val = tensor([1])]; + tensor channels_mean_77_cast_fp16 = reduce_mean(axes = var_3022, keep_dims = var_3012, x = inputs_77_cast_fp16)[name = tensor("channels_mean_77_cast_fp16")]; + tensor zero_mean_77_cast_fp16 = sub(x = inputs_77_cast_fp16, y = channels_mean_77_cast_fp16)[name = tensor("zero_mean_77_cast_fp16")]; + tensor zero_mean_sq_77_cast_fp16 = mul(x = zero_mean_77_cast_fp16, y = zero_mean_77_cast_fp16)[name = tensor("zero_mean_sq_77_cast_fp16")]; + tensor var_3026 = const()[name = tensor("op_3026"), val = tensor([1])]; + tensor var_3027_cast_fp16 = reduce_mean(axes = var_3026, keep_dims = var_3012, x = zero_mean_sq_77_cast_fp16)[name = tensor("op_3027_cast_fp16")]; + tensor var_3028_to_fp16 = const()[name = tensor("op_3028_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_3029_cast_fp16 = add(x = var_3027_cast_fp16, y = var_3028_to_fp16)[name = tensor("op_3029_cast_fp16")]; + tensor denom_77_epsilon_0_to_fp16 = const()[name = tensor("denom_77_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_77_cast_fp16 = rsqrt(epsilon = denom_77_epsilon_0_to_fp16, x = var_3029_cast_fp16)[name = tensor("denom_77_cast_fp16")]; + tensor out_77_cast_fp16 = mul(x = zero_mean_77_cast_fp16, y = denom_77_cast_fp16)[name = tensor("out_77_cast_fp16")]; + tensor obj_77_gamma_0_to_fp16 = const()[name = tensor("obj_77_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(196017600)))]; + tensor obj_77_beta_0_to_fp16 = const()[name = tensor("obj_77_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(196020224)))]; + tensor obj_77_epsilon_0_to_fp16 = const()[name = tensor("obj_77_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_77_cast_fp16 = batch_norm(beta = obj_77_beta_0_to_fp16, epsilon = obj_77_epsilon_0_to_fp16, gamma = obj_77_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_77_cast_fp16)[name = tensor("obj_77_cast_fp16")]; + tensor layers_19_self_attn_q_proj_input_shift_to_fp16 = const()[name = tensor("layers_19_self_attn_q_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(196022848)))]; + tensor input_267_cast_fp16 = sub(x = obj_77_cast_fp16, y = layers_19_self_attn_q_proj_input_shift_to_fp16)[name = tensor("input_267_cast_fp16")]; + tensor var_3048 = const()[name = tensor("op_3048"), val = tensor([1, 1])]; + tensor var_3050 = const()[name = tensor("op_3050"), val = tensor([1, 1])]; + tensor x_343_pad_type_0 = const()[name = tensor("x_343_pad_type_0"), val = tensor("custom")]; + tensor x_343_pad_0 = const()[name = tensor("x_343_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_19_self_attn_q_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(196025472))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(196844736))), name = tensor("layers_19_self_attn_q_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_19_self_attn_q_proj_module_bias_to_fp16 = const()[name = tensor("layers_19_self_attn_q_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(196844864)))]; + tensor x_343_cast_fp16 = conv(bias = layers_19_self_attn_q_proj_module_bias_to_fp16, dilations = var_3050, groups = var_3011, pad = x_343_pad_0, pad_type = x_343_pad_type_0, strides = var_3048, weight = layers_19_self_attn_q_proj_module_weight_to_fp16_palettized, x = input_267_cast_fp16)[name = tensor("x_343_cast_fp16")]; + tensor layers_19_self_attn_q_proj_output_scale_to_fp16 = const()[name = tensor("layers_19_self_attn_q_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(196847488)))]; + tensor query_39_cast_fp16 = mul(x = x_343_cast_fp16, y = layers_19_self_attn_q_proj_output_scale_to_fp16)[name = tensor("query_39_cast_fp16")]; + tensor var_3060 = const()[name = tensor("op_3060"), val = tensor([1, 1])]; + tensor var_3062 = const()[name = tensor("op_3062"), val = tensor([1, 1])]; + tensor x_345_pad_type_0 = const()[name = tensor("x_345_pad_type_0"), val = tensor("custom")]; + tensor x_345_pad_0 = const()[name = tensor("x_345_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_19_self_attn_k_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(196850112))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(197669376))), name = tensor("layers_19_self_attn_k_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_19_self_attn_k_proj_module_bias_to_fp16 = const()[name = tensor("layers_19_self_attn_k_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(197669504)))]; + tensor x_345_cast_fp16 = conv(bias = layers_19_self_attn_k_proj_module_bias_to_fp16, dilations = var_3062, groups = var_3011, pad = x_345_pad_0, pad_type = x_345_pad_type_0, strides = var_3060, weight = layers_19_self_attn_k_proj_module_weight_to_fp16_palettized, x = input_267_cast_fp16)[name = tensor("x_345_cast_fp16")]; + tensor layers_19_self_attn_k_proj_output_scale_to_fp16 = const()[name = tensor("layers_19_self_attn_k_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(197672128)))]; + tensor key_39_cast_fp16 = mul(x = x_345_cast_fp16, y = layers_19_self_attn_k_proj_output_scale_to_fp16)[name = tensor("key_39_cast_fp16")]; + tensor var_3072 = const()[name = tensor("op_3072"), val = tensor([1, 1])]; + tensor var_3074 = const()[name = tensor("op_3074"), val = tensor([1, 1])]; + tensor x_347_pad_type_0 = const()[name = tensor("x_347_pad_type_0"), val = tensor("custom")]; + tensor x_347_pad_0 = const()[name = tensor("x_347_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_19_self_attn_v_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(197674752))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(198494016))), name = tensor("layers_19_self_attn_v_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_19_self_attn_v_proj_module_bias_to_fp16 = const()[name = tensor("layers_19_self_attn_v_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(198494144)))]; + tensor x_347_cast_fp16 = conv(bias = layers_19_self_attn_v_proj_module_bias_to_fp16, dilations = var_3074, groups = var_3011, pad = x_347_pad_0, pad_type = x_347_pad_type_0, strides = var_3072, weight = layers_19_self_attn_v_proj_module_weight_to_fp16_palettized, x = input_267_cast_fp16)[name = tensor("x_347_cast_fp16")]; + tensor layers_19_self_attn_v_proj_output_scale_to_fp16 = const()[name = tensor("layers_19_self_attn_v_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(198496768)))]; + tensor value_39_cast_fp16 = mul(x = x_347_cast_fp16, y = layers_19_self_attn_v_proj_output_scale_to_fp16)[name = tensor("value_39_cast_fp16")]; + tensor var_3079 = const()[name = tensor("op_3079"), val = tensor([1, 20, 64, -1])]; + tensor var_3080_cast_fp16 = reshape(shape = var_3079, x = query_39_cast_fp16)[name = tensor("op_3080_cast_fp16")]; + tensor var_3081_to_fp16 = const()[name = tensor("op_3081_to_fp16"), val = tensor(0x1p-3)]; + tensor var_3082_cast_fp16 = mul(x = var_3080_cast_fp16, y = var_3081_to_fp16)[name = tensor("op_3082_cast_fp16")]; + tensor var_3083 = const()[name = tensor("op_3083"), val = tensor([1, 20, 64, -1])]; + tensor var_3084_cast_fp16 = reshape(shape = var_3083, x = key_39_cast_fp16)[name = tensor("op_3084_cast_fp16")]; + tensor mh_w_39_transpose_x_0 = const()[name = tensor("mh_w_39_transpose_x_0"), val = tensor(true)]; + tensor mh_w_39_transpose_y_0 = const()[name = tensor("mh_w_39_transpose_y_0"), val = tensor(false)]; + tensor mh_w_39_cast_fp16 = matmul(transpose_x = mh_w_39_transpose_x_0, transpose_y = mh_w_39_transpose_y_0, x = var_3082_cast_fp16, y = var_3084_cast_fp16)[name = tensor("mh_w_39_cast_fp16")]; + tensor var_3087_cast_fp16 = softmax(axis = var_3009, x = mh_w_39_cast_fp16)[name = tensor("op_3087_cast_fp16")]; + tensor var_3088 = const()[name = tensor("op_3088"), val = tensor([1, 20, 64, -1])]; + tensor var_3089_cast_fp16 = reshape(shape = var_3088, x = value_39_cast_fp16)[name = tensor("op_3089_cast_fp16")]; + tensor attn_39_transpose_x_0 = const()[name = tensor("attn_39_transpose_x_0"), val = tensor(false)]; + tensor attn_39_transpose_y_0 = const()[name = tensor("attn_39_transpose_y_0"), val = tensor(true)]; + tensor attn_39_cast_fp16 = matmul(transpose_x = attn_39_transpose_x_0, transpose_y = attn_39_transpose_y_0, x = var_3089_cast_fp16, y = var_3087_cast_fp16)[name = tensor("attn_39_cast_fp16")]; + tensor var_3092 = const()[name = tensor("op_3092"), val = tensor([1, 1280, 1, -1])]; + tensor x_349_cast_fp16 = reshape(shape = var_3092, x = attn_39_cast_fp16)[name = tensor("x_349_cast_fp16")]; + tensor layers_19_self_attn_o_proj_input_shift_to_fp16 = const()[name = tensor("layers_19_self_attn_o_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(198499392)))]; + tensor input_273_cast_fp16 = sub(x = x_349_cast_fp16, y = layers_19_self_attn_o_proj_input_shift_to_fp16)[name = tensor("input_273_cast_fp16")]; + tensor var_3100 = const()[name = tensor("op_3100"), val = tensor([1, 1])]; + tensor var_3102 = const()[name = tensor("op_3102"), val = tensor([1, 1])]; + tensor x_351_pad_type_0 = const()[name = tensor("x_351_pad_type_0"), val = tensor("custom")]; + tensor x_351_pad_0 = const()[name = tensor("x_351_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_19_self_attn_o_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(198502016))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(199321280))), name = tensor("layers_19_self_attn_o_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_19_self_attn_o_proj_module_bias_to_fp16 = const()[name = tensor("layers_19_self_attn_o_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(199321408)))]; + tensor x_351_cast_fp16 = conv(bias = layers_19_self_attn_o_proj_module_bias_to_fp16, dilations = var_3102, groups = var_3011, pad = x_351_pad_0, pad_type = x_351_pad_type_0, strides = var_3100, weight = layers_19_self_attn_o_proj_module_weight_to_fp16_palettized, x = input_273_cast_fp16)[name = tensor("x_351_cast_fp16")]; + tensor layers_19_self_attn_o_proj_output_scale_to_fp16 = const()[name = tensor("layers_19_self_attn_o_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(199324032)))]; + tensor obj_79_cast_fp16 = mul(x = x_351_cast_fp16, y = layers_19_self_attn_o_proj_output_scale_to_fp16)[name = tensor("obj_79_cast_fp16")]; + tensor inputs_79_cast_fp16 = add(x = inputs_77_cast_fp16, y = obj_79_cast_fp16)[name = tensor("inputs_79_cast_fp16")]; + tensor var_3109 = const()[name = tensor("op_3109"), val = tensor([1])]; + tensor channels_mean_79_cast_fp16 = reduce_mean(axes = var_3109, keep_dims = var_3012, x = inputs_79_cast_fp16)[name = tensor("channels_mean_79_cast_fp16")]; + tensor zero_mean_79_cast_fp16 = sub(x = inputs_79_cast_fp16, y = channels_mean_79_cast_fp16)[name = tensor("zero_mean_79_cast_fp16")]; + tensor zero_mean_sq_79_cast_fp16 = mul(x = zero_mean_79_cast_fp16, y = zero_mean_79_cast_fp16)[name = tensor("zero_mean_sq_79_cast_fp16")]; + tensor var_3113 = const()[name = tensor("op_3113"), val = tensor([1])]; + tensor var_3114_cast_fp16 = reduce_mean(axes = var_3113, keep_dims = var_3012, x = zero_mean_sq_79_cast_fp16)[name = tensor("op_3114_cast_fp16")]; + tensor var_3115_to_fp16 = const()[name = tensor("op_3115_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_3116_cast_fp16 = add(x = var_3114_cast_fp16, y = var_3115_to_fp16)[name = tensor("op_3116_cast_fp16")]; + tensor denom_79_epsilon_0_to_fp16 = const()[name = tensor("denom_79_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_79_cast_fp16 = rsqrt(epsilon = denom_79_epsilon_0_to_fp16, x = var_3116_cast_fp16)[name = tensor("denom_79_cast_fp16")]; + tensor out_79_cast_fp16 = mul(x = zero_mean_79_cast_fp16, y = denom_79_cast_fp16)[name = tensor("out_79_cast_fp16")]; + tensor x_353_gamma_0_to_fp16 = const()[name = tensor("x_353_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(199326656)))]; + tensor x_353_beta_0_to_fp16 = const()[name = tensor("x_353_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(199329280)))]; + tensor x_353_epsilon_0_to_fp16 = const()[name = tensor("x_353_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_353_cast_fp16 = batch_norm(beta = x_353_beta_0_to_fp16, epsilon = x_353_epsilon_0_to_fp16, gamma = x_353_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_79_cast_fp16)[name = tensor("x_353_cast_fp16")]; + tensor layers_19_fc1_input_shift_to_fp16 = const()[name = tensor("layers_19_fc1_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(199331904)))]; + tensor input_275_cast_fp16 = sub(x = x_353_cast_fp16, y = layers_19_fc1_input_shift_to_fp16)[name = tensor("input_275_cast_fp16")]; + tensor var_3131 = const()[name = tensor("op_3131"), val = tensor([1, 1])]; + tensor var_3133 = const()[name = tensor("op_3133"), val = tensor([1, 1])]; + tensor x_355_pad_type_0 = const()[name = tensor("x_355_pad_type_0"), val = tensor("custom")]; + tensor x_355_pad_0 = const()[name = tensor("x_355_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_19_fc1_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(199334528))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(202611392))), name = tensor("layers_19_fc1_module_weight_to_fp16_palettized"), shape = tensor([5120, 1280, 1, 1])]; + tensor layers_19_fc1_module_bias_to_fp16 = const()[name = tensor("layers_19_fc1_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(202611520)))]; + tensor x_355_cast_fp16 = conv(bias = layers_19_fc1_module_bias_to_fp16, dilations = var_3133, groups = var_3011, pad = x_355_pad_0, pad_type = x_355_pad_type_0, strides = var_3131, weight = layers_19_fc1_module_weight_to_fp16_palettized, x = input_275_cast_fp16)[name = tensor("x_355_cast_fp16")]; + tensor layers_19_fc1_output_scale_to_fp16 = const()[name = tensor("layers_19_fc1_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(202621824)))]; + tensor input_277_cast_fp16 = mul(x = x_355_cast_fp16, y = layers_19_fc1_output_scale_to_fp16)[name = tensor("input_277_cast_fp16")]; + tensor x_357_mode_0 = const()[name = tensor("x_357_mode_0"), val = tensor("EXACT")]; + tensor x_357_cast_fp16 = gelu(mode = x_357_mode_0, x = input_277_cast_fp16)[name = tensor("x_357_cast_fp16")]; + tensor layers_19_fc2_input_shift_to_fp16 = const()[name = tensor("layers_19_fc2_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(202632128)))]; + tensor input_279_cast_fp16 = sub(x = x_357_cast_fp16, y = layers_19_fc2_input_shift_to_fp16)[name = tensor("input_279_cast_fp16")]; + tensor var_3144 = const()[name = tensor("op_3144"), val = tensor([1, 1])]; + tensor var_3146 = const()[name = tensor("op_3146"), val = tensor([1, 1])]; + tensor x_359_pad_type_0 = const()[name = tensor("x_359_pad_type_0"), val = tensor("custom")]; + tensor x_359_pad_0 = const()[name = tensor("x_359_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_19_fc2_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(202642432))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(205919296))), name = tensor("layers_19_fc2_module_weight_to_fp16_palettized"), shape = tensor([1280, 5120, 1, 1])]; + tensor layers_19_fc2_module_bias_to_fp16 = const()[name = tensor("layers_19_fc2_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(205919424)))]; + tensor x_359_cast_fp16 = conv(bias = layers_19_fc2_module_bias_to_fp16, dilations = var_3146, groups = var_3011, pad = x_359_pad_0, pad_type = x_359_pad_type_0, strides = var_3144, weight = layers_19_fc2_module_weight_to_fp16_palettized, x = input_279_cast_fp16)[name = tensor("x_359_cast_fp16")]; + tensor layers_19_fc2_output_scale_to_fp16 = const()[name = tensor("layers_19_fc2_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(205922048)))]; + tensor hidden_states_43_cast_fp16 = mul(x = x_359_cast_fp16, y = layers_19_fc2_output_scale_to_fp16)[name = tensor("hidden_states_43_cast_fp16")]; + tensor inputs_81_cast_fp16 = add(x = inputs_79_cast_fp16, y = hidden_states_43_cast_fp16)[name = tensor("inputs_81_cast_fp16")]; + tensor var_3158 = const()[name = tensor("op_3158"), val = tensor(3)]; + tensor var_3160 = const()[name = tensor("op_3160"), val = tensor(1)]; + tensor var_3161 = const()[name = tensor("op_3161"), val = tensor(true)]; + tensor var_3171 = const()[name = tensor("op_3171"), val = tensor([1])]; + tensor channels_mean_81_cast_fp16 = reduce_mean(axes = var_3171, keep_dims = var_3161, x = inputs_81_cast_fp16)[name = tensor("channels_mean_81_cast_fp16")]; + tensor zero_mean_81_cast_fp16 = sub(x = inputs_81_cast_fp16, y = channels_mean_81_cast_fp16)[name = tensor("zero_mean_81_cast_fp16")]; + tensor zero_mean_sq_81_cast_fp16 = mul(x = zero_mean_81_cast_fp16, y = zero_mean_81_cast_fp16)[name = tensor("zero_mean_sq_81_cast_fp16")]; + tensor var_3175 = const()[name = tensor("op_3175"), val = tensor([1])]; + tensor var_3176_cast_fp16 = reduce_mean(axes = var_3175, keep_dims = var_3161, x = zero_mean_sq_81_cast_fp16)[name = tensor("op_3176_cast_fp16")]; + tensor var_3177_to_fp16 = const()[name = tensor("op_3177_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_3178_cast_fp16 = add(x = var_3176_cast_fp16, y = var_3177_to_fp16)[name = tensor("op_3178_cast_fp16")]; + tensor denom_81_epsilon_0_to_fp16 = const()[name = tensor("denom_81_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_81_cast_fp16 = rsqrt(epsilon = denom_81_epsilon_0_to_fp16, x = var_3178_cast_fp16)[name = tensor("denom_81_cast_fp16")]; + tensor out_81_cast_fp16 = mul(x = zero_mean_81_cast_fp16, y = denom_81_cast_fp16)[name = tensor("out_81_cast_fp16")]; + tensor obj_81_gamma_0_to_fp16 = const()[name = tensor("obj_81_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(205924672)))]; + tensor obj_81_beta_0_to_fp16 = const()[name = tensor("obj_81_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(205927296)))]; + tensor obj_81_epsilon_0_to_fp16 = const()[name = tensor("obj_81_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_81_cast_fp16 = batch_norm(beta = obj_81_beta_0_to_fp16, epsilon = obj_81_epsilon_0_to_fp16, gamma = obj_81_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_81_cast_fp16)[name = tensor("obj_81_cast_fp16")]; + tensor layers_20_self_attn_q_proj_input_shift_to_fp16 = const()[name = tensor("layers_20_self_attn_q_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(205929920)))]; + tensor input_281_cast_fp16 = sub(x = obj_81_cast_fp16, y = layers_20_self_attn_q_proj_input_shift_to_fp16)[name = tensor("input_281_cast_fp16")]; + tensor var_3197 = const()[name = tensor("op_3197"), val = tensor([1, 1])]; + tensor var_3199 = const()[name = tensor("op_3199"), val = tensor([1, 1])]; + tensor x_361_pad_type_0 = const()[name = tensor("x_361_pad_type_0"), val = tensor("custom")]; + tensor x_361_pad_0 = const()[name = tensor("x_361_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_20_self_attn_q_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(205932544))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(206751808))), name = tensor("layers_20_self_attn_q_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_20_self_attn_q_proj_module_bias_to_fp16 = const()[name = tensor("layers_20_self_attn_q_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(206751936)))]; + tensor x_361_cast_fp16 = conv(bias = layers_20_self_attn_q_proj_module_bias_to_fp16, dilations = var_3199, groups = var_3160, pad = x_361_pad_0, pad_type = x_361_pad_type_0, strides = var_3197, weight = layers_20_self_attn_q_proj_module_weight_to_fp16_palettized, x = input_281_cast_fp16)[name = tensor("x_361_cast_fp16")]; + tensor layers_20_self_attn_q_proj_output_scale_to_fp16 = const()[name = tensor("layers_20_self_attn_q_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(206754560)))]; + tensor query_41_cast_fp16 = mul(x = x_361_cast_fp16, y = layers_20_self_attn_q_proj_output_scale_to_fp16)[name = tensor("query_41_cast_fp16")]; + tensor var_3209 = const()[name = tensor("op_3209"), val = tensor([1, 1])]; + tensor var_3211 = const()[name = tensor("op_3211"), val = tensor([1, 1])]; + tensor x_363_pad_type_0 = const()[name = tensor("x_363_pad_type_0"), val = tensor("custom")]; + tensor x_363_pad_0 = const()[name = tensor("x_363_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_20_self_attn_k_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(206757184))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(207576448))), name = tensor("layers_20_self_attn_k_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_20_self_attn_k_proj_module_bias_to_fp16 = const()[name = tensor("layers_20_self_attn_k_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(207576576)))]; + tensor x_363_cast_fp16 = conv(bias = layers_20_self_attn_k_proj_module_bias_to_fp16, dilations = var_3211, groups = var_3160, pad = x_363_pad_0, pad_type = x_363_pad_type_0, strides = var_3209, weight = layers_20_self_attn_k_proj_module_weight_to_fp16_palettized, x = input_281_cast_fp16)[name = tensor("x_363_cast_fp16")]; + tensor layers_20_self_attn_k_proj_output_scale_to_fp16 = const()[name = tensor("layers_20_self_attn_k_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(207579200)))]; + tensor key_41_cast_fp16 = mul(x = x_363_cast_fp16, y = layers_20_self_attn_k_proj_output_scale_to_fp16)[name = tensor("key_41_cast_fp16")]; + tensor var_3221 = const()[name = tensor("op_3221"), val = tensor([1, 1])]; + tensor var_3223 = const()[name = tensor("op_3223"), val = tensor([1, 1])]; + tensor x_365_pad_type_0 = const()[name = tensor("x_365_pad_type_0"), val = tensor("custom")]; + tensor x_365_pad_0 = const()[name = tensor("x_365_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_20_self_attn_v_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(207581824))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(208401088))), name = tensor("layers_20_self_attn_v_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_20_self_attn_v_proj_module_bias_to_fp16 = const()[name = tensor("layers_20_self_attn_v_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(208401216)))]; + tensor x_365_cast_fp16 = conv(bias = layers_20_self_attn_v_proj_module_bias_to_fp16, dilations = var_3223, groups = var_3160, pad = x_365_pad_0, pad_type = x_365_pad_type_0, strides = var_3221, weight = layers_20_self_attn_v_proj_module_weight_to_fp16_palettized, x = input_281_cast_fp16)[name = tensor("x_365_cast_fp16")]; + tensor layers_20_self_attn_v_proj_output_scale_to_fp16 = const()[name = tensor("layers_20_self_attn_v_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(208403840)))]; + tensor value_41_cast_fp16 = mul(x = x_365_cast_fp16, y = layers_20_self_attn_v_proj_output_scale_to_fp16)[name = tensor("value_41_cast_fp16")]; + tensor var_3228 = const()[name = tensor("op_3228"), val = tensor([1, 20, 64, -1])]; + tensor var_3229_cast_fp16 = reshape(shape = var_3228, x = query_41_cast_fp16)[name = tensor("op_3229_cast_fp16")]; + tensor var_3230_to_fp16 = const()[name = tensor("op_3230_to_fp16"), val = tensor(0x1p-3)]; + tensor var_3231_cast_fp16 = mul(x = var_3229_cast_fp16, y = var_3230_to_fp16)[name = tensor("op_3231_cast_fp16")]; + tensor var_3232 = const()[name = tensor("op_3232"), val = tensor([1, 20, 64, -1])]; + tensor var_3233_cast_fp16 = reshape(shape = var_3232, x = key_41_cast_fp16)[name = tensor("op_3233_cast_fp16")]; + tensor mh_w_41_transpose_x_0 = const()[name = tensor("mh_w_41_transpose_x_0"), val = tensor(true)]; + tensor mh_w_41_transpose_y_0 = const()[name = tensor("mh_w_41_transpose_y_0"), val = tensor(false)]; + tensor mh_w_41_cast_fp16 = matmul(transpose_x = mh_w_41_transpose_x_0, transpose_y = mh_w_41_transpose_y_0, x = var_3231_cast_fp16, y = var_3233_cast_fp16)[name = tensor("mh_w_41_cast_fp16")]; + tensor var_3236_cast_fp16 = softmax(axis = var_3158, x = mh_w_41_cast_fp16)[name = tensor("op_3236_cast_fp16")]; + tensor var_3237 = const()[name = tensor("op_3237"), val = tensor([1, 20, 64, -1])]; + tensor var_3238_cast_fp16 = reshape(shape = var_3237, x = value_41_cast_fp16)[name = tensor("op_3238_cast_fp16")]; + tensor attn_41_transpose_x_0 = const()[name = tensor("attn_41_transpose_x_0"), val = tensor(false)]; + tensor attn_41_transpose_y_0 = const()[name = tensor("attn_41_transpose_y_0"), val = tensor(true)]; + tensor attn_41_cast_fp16 = matmul(transpose_x = attn_41_transpose_x_0, transpose_y = attn_41_transpose_y_0, x = var_3238_cast_fp16, y = var_3236_cast_fp16)[name = tensor("attn_41_cast_fp16")]; + tensor var_3241 = const()[name = tensor("op_3241"), val = tensor([1, 1280, 1, -1])]; + tensor x_367_cast_fp16 = reshape(shape = var_3241, x = attn_41_cast_fp16)[name = tensor("x_367_cast_fp16")]; + tensor layers_20_self_attn_o_proj_input_shift_to_fp16 = const()[name = tensor("layers_20_self_attn_o_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(208406464)))]; + tensor input_287_cast_fp16 = sub(x = x_367_cast_fp16, y = layers_20_self_attn_o_proj_input_shift_to_fp16)[name = tensor("input_287_cast_fp16")]; + tensor var_3249 = const()[name = tensor("op_3249"), val = tensor([1, 1])]; + tensor var_3251 = const()[name = tensor("op_3251"), val = tensor([1, 1])]; + tensor x_369_pad_type_0 = const()[name = tensor("x_369_pad_type_0"), val = tensor("custom")]; + tensor x_369_pad_0 = const()[name = tensor("x_369_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_20_self_attn_o_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(208409088))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(209228352))), name = tensor("layers_20_self_attn_o_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_20_self_attn_o_proj_module_bias_to_fp16 = const()[name = tensor("layers_20_self_attn_o_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(209228480)))]; + tensor x_369_cast_fp16 = conv(bias = layers_20_self_attn_o_proj_module_bias_to_fp16, dilations = var_3251, groups = var_3160, pad = x_369_pad_0, pad_type = x_369_pad_type_0, strides = var_3249, weight = layers_20_self_attn_o_proj_module_weight_to_fp16_palettized, x = input_287_cast_fp16)[name = tensor("x_369_cast_fp16")]; + tensor layers_20_self_attn_o_proj_output_scale_to_fp16 = const()[name = tensor("layers_20_self_attn_o_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(209231104)))]; + tensor obj_83_cast_fp16 = mul(x = x_369_cast_fp16, y = layers_20_self_attn_o_proj_output_scale_to_fp16)[name = tensor("obj_83_cast_fp16")]; + tensor inputs_83_cast_fp16 = add(x = inputs_81_cast_fp16, y = obj_83_cast_fp16)[name = tensor("inputs_83_cast_fp16")]; + tensor var_3258 = const()[name = tensor("op_3258"), val = tensor([1])]; + tensor channels_mean_83_cast_fp16 = reduce_mean(axes = var_3258, keep_dims = var_3161, x = inputs_83_cast_fp16)[name = tensor("channels_mean_83_cast_fp16")]; + tensor zero_mean_83_cast_fp16 = sub(x = inputs_83_cast_fp16, y = channels_mean_83_cast_fp16)[name = tensor("zero_mean_83_cast_fp16")]; + tensor zero_mean_sq_83_cast_fp16 = mul(x = zero_mean_83_cast_fp16, y = zero_mean_83_cast_fp16)[name = tensor("zero_mean_sq_83_cast_fp16")]; + tensor var_3262 = const()[name = tensor("op_3262"), val = tensor([1])]; + tensor var_3263_cast_fp16 = reduce_mean(axes = var_3262, keep_dims = var_3161, x = zero_mean_sq_83_cast_fp16)[name = tensor("op_3263_cast_fp16")]; + tensor var_3264_to_fp16 = const()[name = tensor("op_3264_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_3265_cast_fp16 = add(x = var_3263_cast_fp16, y = var_3264_to_fp16)[name = tensor("op_3265_cast_fp16")]; + tensor denom_83_epsilon_0_to_fp16 = const()[name = tensor("denom_83_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_83_cast_fp16 = rsqrt(epsilon = denom_83_epsilon_0_to_fp16, x = var_3265_cast_fp16)[name = tensor("denom_83_cast_fp16")]; + tensor out_83_cast_fp16 = mul(x = zero_mean_83_cast_fp16, y = denom_83_cast_fp16)[name = tensor("out_83_cast_fp16")]; + tensor x_371_gamma_0_to_fp16 = const()[name = tensor("x_371_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(209233728)))]; + tensor x_371_beta_0_to_fp16 = const()[name = tensor("x_371_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(209236352)))]; + tensor x_371_epsilon_0_to_fp16 = const()[name = tensor("x_371_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_371_cast_fp16 = batch_norm(beta = x_371_beta_0_to_fp16, epsilon = x_371_epsilon_0_to_fp16, gamma = x_371_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_83_cast_fp16)[name = tensor("x_371_cast_fp16")]; + tensor layers_20_fc1_input_shift_to_fp16 = const()[name = tensor("layers_20_fc1_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(209238976)))]; + tensor input_289_cast_fp16 = sub(x = x_371_cast_fp16, y = layers_20_fc1_input_shift_to_fp16)[name = tensor("input_289_cast_fp16")]; + tensor var_3280 = const()[name = tensor("op_3280"), val = tensor([1, 1])]; + tensor var_3282 = const()[name = tensor("op_3282"), val = tensor([1, 1])]; + tensor x_373_pad_type_0 = const()[name = tensor("x_373_pad_type_0"), val = tensor("custom")]; + tensor x_373_pad_0 = const()[name = tensor("x_373_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_20_fc1_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(209241600))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(212518464))), name = tensor("layers_20_fc1_module_weight_to_fp16_palettized"), shape = tensor([5120, 1280, 1, 1])]; + tensor layers_20_fc1_module_bias_to_fp16 = const()[name = tensor("layers_20_fc1_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(212518592)))]; + tensor x_373_cast_fp16 = conv(bias = layers_20_fc1_module_bias_to_fp16, dilations = var_3282, groups = var_3160, pad = x_373_pad_0, pad_type = x_373_pad_type_0, strides = var_3280, weight = layers_20_fc1_module_weight_to_fp16_palettized, x = input_289_cast_fp16)[name = tensor("x_373_cast_fp16")]; + tensor layers_20_fc1_output_scale_to_fp16 = const()[name = tensor("layers_20_fc1_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(212528896)))]; + tensor input_291_cast_fp16 = mul(x = x_373_cast_fp16, y = layers_20_fc1_output_scale_to_fp16)[name = tensor("input_291_cast_fp16")]; + tensor x_375_mode_0 = const()[name = tensor("x_375_mode_0"), val = tensor("EXACT")]; + tensor x_375_cast_fp16 = gelu(mode = x_375_mode_0, x = input_291_cast_fp16)[name = tensor("x_375_cast_fp16")]; + tensor layers_20_fc2_input_shift_to_fp16 = const()[name = tensor("layers_20_fc2_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(212539200)))]; + tensor input_293_cast_fp16 = sub(x = x_375_cast_fp16, y = layers_20_fc2_input_shift_to_fp16)[name = tensor("input_293_cast_fp16")]; + tensor var_3293 = const()[name = tensor("op_3293"), val = tensor([1, 1])]; + tensor var_3295 = const()[name = tensor("op_3295"), val = tensor([1, 1])]; + tensor x_377_pad_type_0 = const()[name = tensor("x_377_pad_type_0"), val = tensor("custom")]; + tensor x_377_pad_0 = const()[name = tensor("x_377_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_20_fc2_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(212549504))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(217464768))), name = tensor("layers_20_fc2_module_weight_to_fp16_palettized"), shape = tensor([1280, 5120, 1, 1])]; + tensor layers_20_fc2_module_bias_to_fp16 = const()[name = tensor("layers_20_fc2_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(217464960)))]; + tensor x_377_cast_fp16 = conv(bias = layers_20_fc2_module_bias_to_fp16, dilations = var_3295, groups = var_3160, pad = x_377_pad_0, pad_type = x_377_pad_type_0, strides = var_3293, weight = layers_20_fc2_module_weight_to_fp16_palettized, x = input_293_cast_fp16)[name = tensor("x_377_cast_fp16")]; + tensor layers_20_fc2_output_scale_to_fp16 = const()[name = tensor("layers_20_fc2_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(217467584)))]; + tensor hidden_states_45_cast_fp16 = mul(x = x_377_cast_fp16, y = layers_20_fc2_output_scale_to_fp16)[name = tensor("hidden_states_45_cast_fp16")]; + tensor inputs_85_cast_fp16 = add(x = inputs_83_cast_fp16, y = hidden_states_45_cast_fp16)[name = tensor("inputs_85_cast_fp16")]; + tensor var_3307 = const()[name = tensor("op_3307"), val = tensor(3)]; + tensor var_3309 = const()[name = tensor("op_3309"), val = tensor(1)]; + tensor var_3310 = const()[name = tensor("op_3310"), val = tensor(true)]; + tensor var_3320 = const()[name = tensor("op_3320"), val = tensor([1])]; + tensor channels_mean_85_cast_fp16 = reduce_mean(axes = var_3320, keep_dims = var_3310, x = inputs_85_cast_fp16)[name = tensor("channels_mean_85_cast_fp16")]; + tensor zero_mean_85_cast_fp16 = sub(x = inputs_85_cast_fp16, y = channels_mean_85_cast_fp16)[name = tensor("zero_mean_85_cast_fp16")]; + tensor zero_mean_sq_85_cast_fp16 = mul(x = zero_mean_85_cast_fp16, y = zero_mean_85_cast_fp16)[name = tensor("zero_mean_sq_85_cast_fp16")]; + tensor var_3324 = const()[name = tensor("op_3324"), val = tensor([1])]; + tensor var_3325_cast_fp16 = reduce_mean(axes = var_3324, keep_dims = var_3310, x = zero_mean_sq_85_cast_fp16)[name = tensor("op_3325_cast_fp16")]; + tensor var_3326_to_fp16 = const()[name = tensor("op_3326_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_3327_cast_fp16 = add(x = var_3325_cast_fp16, y = var_3326_to_fp16)[name = tensor("op_3327_cast_fp16")]; + tensor denom_85_epsilon_0_to_fp16 = const()[name = tensor("denom_85_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_85_cast_fp16 = rsqrt(epsilon = denom_85_epsilon_0_to_fp16, x = var_3327_cast_fp16)[name = tensor("denom_85_cast_fp16")]; + tensor out_85_cast_fp16 = mul(x = zero_mean_85_cast_fp16, y = denom_85_cast_fp16)[name = tensor("out_85_cast_fp16")]; + tensor obj_85_gamma_0_to_fp16 = const()[name = tensor("obj_85_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(217470208)))]; + tensor obj_85_beta_0_to_fp16 = const()[name = tensor("obj_85_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(217472832)))]; + tensor obj_85_epsilon_0_to_fp16 = const()[name = tensor("obj_85_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_85_cast_fp16 = batch_norm(beta = obj_85_beta_0_to_fp16, epsilon = obj_85_epsilon_0_to_fp16, gamma = obj_85_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_85_cast_fp16)[name = tensor("obj_85_cast_fp16")]; + tensor layers_21_self_attn_q_proj_input_shift_to_fp16 = const()[name = tensor("layers_21_self_attn_q_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(217475456)))]; + tensor input_295_cast_fp16 = sub(x = obj_85_cast_fp16, y = layers_21_self_attn_q_proj_input_shift_to_fp16)[name = tensor("input_295_cast_fp16")]; + tensor var_3346 = const()[name = tensor("op_3346"), val = tensor([1, 1])]; + tensor var_3348 = const()[name = tensor("op_3348"), val = tensor([1, 1])]; + tensor x_379_pad_type_0 = const()[name = tensor("x_379_pad_type_0"), val = tensor("custom")]; + tensor x_379_pad_0 = const()[name = tensor("x_379_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_21_self_attn_q_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(217478080))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(218297344))), name = tensor("layers_21_self_attn_q_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_21_self_attn_q_proj_module_bias_to_fp16 = const()[name = tensor("layers_21_self_attn_q_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(218297472)))]; + tensor x_379_cast_fp16 = conv(bias = layers_21_self_attn_q_proj_module_bias_to_fp16, dilations = var_3348, groups = var_3309, pad = x_379_pad_0, pad_type = x_379_pad_type_0, strides = var_3346, weight = layers_21_self_attn_q_proj_module_weight_to_fp16_palettized, x = input_295_cast_fp16)[name = tensor("x_379_cast_fp16")]; + tensor layers_21_self_attn_q_proj_output_scale_to_fp16 = const()[name = tensor("layers_21_self_attn_q_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(218300096)))]; + tensor query_43_cast_fp16 = mul(x = x_379_cast_fp16, y = layers_21_self_attn_q_proj_output_scale_to_fp16)[name = tensor("query_43_cast_fp16")]; + tensor var_3358 = const()[name = tensor("op_3358"), val = tensor([1, 1])]; + tensor var_3360 = const()[name = tensor("op_3360"), val = tensor([1, 1])]; + tensor x_381_pad_type_0 = const()[name = tensor("x_381_pad_type_0"), val = tensor("custom")]; + tensor x_381_pad_0 = const()[name = tensor("x_381_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_21_self_attn_k_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(218302720))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(219121984))), name = tensor("layers_21_self_attn_k_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_21_self_attn_k_proj_module_bias_to_fp16 = const()[name = tensor("layers_21_self_attn_k_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(219122112)))]; + tensor x_381_cast_fp16 = conv(bias = layers_21_self_attn_k_proj_module_bias_to_fp16, dilations = var_3360, groups = var_3309, pad = x_381_pad_0, pad_type = x_381_pad_type_0, strides = var_3358, weight = layers_21_self_attn_k_proj_module_weight_to_fp16_palettized, x = input_295_cast_fp16)[name = tensor("x_381_cast_fp16")]; + tensor layers_21_self_attn_k_proj_output_scale_to_fp16 = const()[name = tensor("layers_21_self_attn_k_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(219124736)))]; + tensor key_43_cast_fp16 = mul(x = x_381_cast_fp16, y = layers_21_self_attn_k_proj_output_scale_to_fp16)[name = tensor("key_43_cast_fp16")]; + tensor var_3370 = const()[name = tensor("op_3370"), val = tensor([1, 1])]; + tensor var_3372 = const()[name = tensor("op_3372"), val = tensor([1, 1])]; + tensor x_383_pad_type_0 = const()[name = tensor("x_383_pad_type_0"), val = tensor("custom")]; + tensor x_383_pad_0 = const()[name = tensor("x_383_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_21_self_attn_v_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(219127360))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(219946624))), name = tensor("layers_21_self_attn_v_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_21_self_attn_v_proj_module_bias_to_fp16 = const()[name = tensor("layers_21_self_attn_v_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(219946752)))]; + tensor x_383_cast_fp16 = conv(bias = layers_21_self_attn_v_proj_module_bias_to_fp16, dilations = var_3372, groups = var_3309, pad = x_383_pad_0, pad_type = x_383_pad_type_0, strides = var_3370, weight = layers_21_self_attn_v_proj_module_weight_to_fp16_palettized, x = input_295_cast_fp16)[name = tensor("x_383_cast_fp16")]; + tensor layers_21_self_attn_v_proj_output_scale_to_fp16 = const()[name = tensor("layers_21_self_attn_v_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(219949376)))]; + tensor value_43_cast_fp16 = mul(x = x_383_cast_fp16, y = layers_21_self_attn_v_proj_output_scale_to_fp16)[name = tensor("value_43_cast_fp16")]; + tensor var_3377 = const()[name = tensor("op_3377"), val = tensor([1, 20, 64, -1])]; + tensor var_3378_cast_fp16 = reshape(shape = var_3377, x = query_43_cast_fp16)[name = tensor("op_3378_cast_fp16")]; + tensor var_3379_to_fp16 = const()[name = tensor("op_3379_to_fp16"), val = tensor(0x1p-3)]; + tensor var_3380_cast_fp16 = mul(x = var_3378_cast_fp16, y = var_3379_to_fp16)[name = tensor("op_3380_cast_fp16")]; + tensor var_3381 = const()[name = tensor("op_3381"), val = tensor([1, 20, 64, -1])]; + tensor var_3382_cast_fp16 = reshape(shape = var_3381, x = key_43_cast_fp16)[name = tensor("op_3382_cast_fp16")]; + tensor mh_w_43_transpose_x_0 = const()[name = tensor("mh_w_43_transpose_x_0"), val = tensor(true)]; + tensor mh_w_43_transpose_y_0 = const()[name = tensor("mh_w_43_transpose_y_0"), val = tensor(false)]; + tensor mh_w_43_cast_fp16 = matmul(transpose_x = mh_w_43_transpose_x_0, transpose_y = mh_w_43_transpose_y_0, x = var_3380_cast_fp16, y = var_3382_cast_fp16)[name = tensor("mh_w_43_cast_fp16")]; + tensor var_3385_cast_fp16 = softmax(axis = var_3307, x = mh_w_43_cast_fp16)[name = tensor("op_3385_cast_fp16")]; + tensor var_3386 = const()[name = tensor("op_3386"), val = tensor([1, 20, 64, -1])]; + tensor var_3387_cast_fp16 = reshape(shape = var_3386, x = value_43_cast_fp16)[name = tensor("op_3387_cast_fp16")]; + tensor attn_43_transpose_x_0 = const()[name = tensor("attn_43_transpose_x_0"), val = tensor(false)]; + tensor attn_43_transpose_y_0 = const()[name = tensor("attn_43_transpose_y_0"), val = tensor(true)]; + tensor attn_43_cast_fp16 = matmul(transpose_x = attn_43_transpose_x_0, transpose_y = attn_43_transpose_y_0, x = var_3387_cast_fp16, y = var_3385_cast_fp16)[name = tensor("attn_43_cast_fp16")]; + tensor var_3390 = const()[name = tensor("op_3390"), val = tensor([1, 1280, 1, -1])]; + tensor x_385_cast_fp16 = reshape(shape = var_3390, x = attn_43_cast_fp16)[name = tensor("x_385_cast_fp16")]; + tensor layers_21_self_attn_o_proj_input_shift_to_fp16 = const()[name = tensor("layers_21_self_attn_o_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(219952000)))]; + tensor input_301_cast_fp16 = sub(x = x_385_cast_fp16, y = layers_21_self_attn_o_proj_input_shift_to_fp16)[name = tensor("input_301_cast_fp16")]; + tensor var_3398 = const()[name = tensor("op_3398"), val = tensor([1, 1])]; + tensor var_3400 = const()[name = tensor("op_3400"), val = tensor([1, 1])]; + tensor x_387_pad_type_0 = const()[name = tensor("x_387_pad_type_0"), val = tensor("custom")]; + tensor x_387_pad_0 = const()[name = tensor("x_387_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_21_self_attn_o_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(219954624))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(220773888))), name = tensor("layers_21_self_attn_o_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_21_self_attn_o_proj_module_bias_to_fp16 = const()[name = tensor("layers_21_self_attn_o_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(220774016)))]; + tensor x_387_cast_fp16 = conv(bias = layers_21_self_attn_o_proj_module_bias_to_fp16, dilations = var_3400, groups = var_3309, pad = x_387_pad_0, pad_type = x_387_pad_type_0, strides = var_3398, weight = layers_21_self_attn_o_proj_module_weight_to_fp16_palettized, x = input_301_cast_fp16)[name = tensor("x_387_cast_fp16")]; + tensor layers_21_self_attn_o_proj_output_scale_to_fp16 = const()[name = tensor("layers_21_self_attn_o_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(220776640)))]; + tensor obj_87_cast_fp16 = mul(x = x_387_cast_fp16, y = layers_21_self_attn_o_proj_output_scale_to_fp16)[name = tensor("obj_87_cast_fp16")]; + tensor inputs_87_cast_fp16 = add(x = inputs_85_cast_fp16, y = obj_87_cast_fp16)[name = tensor("inputs_87_cast_fp16")]; + tensor var_3407 = const()[name = tensor("op_3407"), val = tensor([1])]; + tensor channels_mean_87_cast_fp16 = reduce_mean(axes = var_3407, keep_dims = var_3310, x = inputs_87_cast_fp16)[name = tensor("channels_mean_87_cast_fp16")]; + tensor zero_mean_87_cast_fp16 = sub(x = inputs_87_cast_fp16, y = channels_mean_87_cast_fp16)[name = tensor("zero_mean_87_cast_fp16")]; + tensor zero_mean_sq_87_cast_fp16 = mul(x = zero_mean_87_cast_fp16, y = zero_mean_87_cast_fp16)[name = tensor("zero_mean_sq_87_cast_fp16")]; + tensor var_3411 = const()[name = tensor("op_3411"), val = tensor([1])]; + tensor var_3412_cast_fp16 = reduce_mean(axes = var_3411, keep_dims = var_3310, x = zero_mean_sq_87_cast_fp16)[name = tensor("op_3412_cast_fp16")]; + tensor var_3413_to_fp16 = const()[name = tensor("op_3413_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_3414_cast_fp16 = add(x = var_3412_cast_fp16, y = var_3413_to_fp16)[name = tensor("op_3414_cast_fp16")]; + tensor denom_87_epsilon_0_to_fp16 = const()[name = tensor("denom_87_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_87_cast_fp16 = rsqrt(epsilon = denom_87_epsilon_0_to_fp16, x = var_3414_cast_fp16)[name = tensor("denom_87_cast_fp16")]; + tensor out_87_cast_fp16 = mul(x = zero_mean_87_cast_fp16, y = denom_87_cast_fp16)[name = tensor("out_87_cast_fp16")]; + tensor x_389_gamma_0_to_fp16 = const()[name = tensor("x_389_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(220779264)))]; + tensor x_389_beta_0_to_fp16 = const()[name = tensor("x_389_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(220781888)))]; + tensor x_389_epsilon_0_to_fp16 = const()[name = tensor("x_389_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_389_cast_fp16 = batch_norm(beta = x_389_beta_0_to_fp16, epsilon = x_389_epsilon_0_to_fp16, gamma = x_389_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_87_cast_fp16)[name = tensor("x_389_cast_fp16")]; + tensor layers_21_fc1_input_shift_to_fp16 = const()[name = tensor("layers_21_fc1_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(220784512)))]; + tensor input_303_cast_fp16 = sub(x = x_389_cast_fp16, y = layers_21_fc1_input_shift_to_fp16)[name = tensor("input_303_cast_fp16")]; + tensor var_3429 = const()[name = tensor("op_3429"), val = tensor([1, 1])]; + tensor var_3431 = const()[name = tensor("op_3431"), val = tensor([1, 1])]; + tensor x_391_pad_type_0 = const()[name = tensor("x_391_pad_type_0"), val = tensor("custom")]; + tensor x_391_pad_0 = const()[name = tensor("x_391_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_21_fc1_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(220787136))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(224064000))), name = tensor("layers_21_fc1_module_weight_to_fp16_palettized"), shape = tensor([5120, 1280, 1, 1])]; + tensor layers_21_fc1_module_bias_to_fp16 = const()[name = tensor("layers_21_fc1_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(224064128)))]; + tensor x_391_cast_fp16 = conv(bias = layers_21_fc1_module_bias_to_fp16, dilations = var_3431, groups = var_3309, pad = x_391_pad_0, pad_type = x_391_pad_type_0, strides = var_3429, weight = layers_21_fc1_module_weight_to_fp16_palettized, x = input_303_cast_fp16)[name = tensor("x_391_cast_fp16")]; + tensor layers_21_fc1_output_scale_to_fp16 = const()[name = tensor("layers_21_fc1_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(224074432)))]; + tensor input_305_cast_fp16 = mul(x = x_391_cast_fp16, y = layers_21_fc1_output_scale_to_fp16)[name = tensor("input_305_cast_fp16")]; + tensor x_393_mode_0 = const()[name = tensor("x_393_mode_0"), val = tensor("EXACT")]; + tensor x_393_cast_fp16 = gelu(mode = x_393_mode_0, x = input_305_cast_fp16)[name = tensor("x_393_cast_fp16")]; + tensor layers_21_fc2_input_shift_to_fp16 = const()[name = tensor("layers_21_fc2_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(224084736)))]; + tensor input_307_cast_fp16 = sub(x = x_393_cast_fp16, y = layers_21_fc2_input_shift_to_fp16)[name = tensor("input_307_cast_fp16")]; + tensor var_3442 = const()[name = tensor("op_3442"), val = tensor([1, 1])]; + tensor var_3444 = const()[name = tensor("op_3444"), val = tensor([1, 1])]; + tensor x_395_pad_type_0 = const()[name = tensor("x_395_pad_type_0"), val = tensor("custom")]; + tensor x_395_pad_0 = const()[name = tensor("x_395_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_21_fc2_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(224095040))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(227371904))), name = tensor("layers_21_fc2_module_weight_to_fp16_palettized"), shape = tensor([1280, 5120, 1, 1])]; + tensor layers_21_fc2_module_bias_to_fp16 = const()[name = tensor("layers_21_fc2_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(227372032)))]; + tensor x_395_cast_fp16 = conv(bias = layers_21_fc2_module_bias_to_fp16, dilations = var_3444, groups = var_3309, pad = x_395_pad_0, pad_type = x_395_pad_type_0, strides = var_3442, weight = layers_21_fc2_module_weight_to_fp16_palettized, x = input_307_cast_fp16)[name = tensor("x_395_cast_fp16")]; + tensor layers_21_fc2_output_scale_to_fp16 = const()[name = tensor("layers_21_fc2_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(227374656)))]; + tensor hidden_states_47_cast_fp16 = mul(x = x_395_cast_fp16, y = layers_21_fc2_output_scale_to_fp16)[name = tensor("hidden_states_47_cast_fp16")]; + tensor inputs_89_cast_fp16 = add(x = inputs_87_cast_fp16, y = hidden_states_47_cast_fp16)[name = tensor("inputs_89_cast_fp16")]; + tensor var_3456 = const()[name = tensor("op_3456"), val = tensor(3)]; + tensor var_3458 = const()[name = tensor("op_3458"), val = tensor(1)]; + tensor var_3459 = const()[name = tensor("op_3459"), val = tensor(true)]; + tensor var_3469 = const()[name = tensor("op_3469"), val = tensor([1])]; + tensor channels_mean_89_cast_fp16 = reduce_mean(axes = var_3469, keep_dims = var_3459, x = inputs_89_cast_fp16)[name = tensor("channels_mean_89_cast_fp16")]; + tensor zero_mean_89_cast_fp16 = sub(x = inputs_89_cast_fp16, y = channels_mean_89_cast_fp16)[name = tensor("zero_mean_89_cast_fp16")]; + tensor zero_mean_sq_89_cast_fp16 = mul(x = zero_mean_89_cast_fp16, y = zero_mean_89_cast_fp16)[name = tensor("zero_mean_sq_89_cast_fp16")]; + tensor var_3473 = const()[name = tensor("op_3473"), val = tensor([1])]; + tensor var_3474_cast_fp16 = reduce_mean(axes = var_3473, keep_dims = var_3459, x = zero_mean_sq_89_cast_fp16)[name = tensor("op_3474_cast_fp16")]; + tensor var_3475_to_fp16 = const()[name = tensor("op_3475_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_3476_cast_fp16 = add(x = var_3474_cast_fp16, y = var_3475_to_fp16)[name = tensor("op_3476_cast_fp16")]; + tensor denom_89_epsilon_0_to_fp16 = const()[name = tensor("denom_89_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_89_cast_fp16 = rsqrt(epsilon = denom_89_epsilon_0_to_fp16, x = var_3476_cast_fp16)[name = tensor("denom_89_cast_fp16")]; + tensor out_89_cast_fp16 = mul(x = zero_mean_89_cast_fp16, y = denom_89_cast_fp16)[name = tensor("out_89_cast_fp16")]; + tensor obj_89_gamma_0_to_fp16 = const()[name = tensor("obj_89_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(227377280)))]; + tensor obj_89_beta_0_to_fp16 = const()[name = tensor("obj_89_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(227379904)))]; + tensor obj_89_epsilon_0_to_fp16 = const()[name = tensor("obj_89_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_89_cast_fp16 = batch_norm(beta = obj_89_beta_0_to_fp16, epsilon = obj_89_epsilon_0_to_fp16, gamma = obj_89_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_89_cast_fp16)[name = tensor("obj_89_cast_fp16")]; + tensor layers_22_self_attn_q_proj_input_shift_to_fp16 = const()[name = tensor("layers_22_self_attn_q_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(227382528)))]; + tensor input_309_cast_fp16 = sub(x = obj_89_cast_fp16, y = layers_22_self_attn_q_proj_input_shift_to_fp16)[name = tensor("input_309_cast_fp16")]; + tensor var_3495 = const()[name = tensor("op_3495"), val = tensor([1, 1])]; + tensor var_3497 = const()[name = tensor("op_3497"), val = tensor([1, 1])]; + tensor x_397_pad_type_0 = const()[name = tensor("x_397_pad_type_0"), val = tensor("custom")]; + tensor x_397_pad_0 = const()[name = tensor("x_397_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_22_self_attn_q_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(227385152))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(228204416))), name = tensor("layers_22_self_attn_q_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_22_self_attn_q_proj_module_bias_to_fp16 = const()[name = tensor("layers_22_self_attn_q_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(228204544)))]; + tensor x_397_cast_fp16 = conv(bias = layers_22_self_attn_q_proj_module_bias_to_fp16, dilations = var_3497, groups = var_3458, pad = x_397_pad_0, pad_type = x_397_pad_type_0, strides = var_3495, weight = layers_22_self_attn_q_proj_module_weight_to_fp16_palettized, x = input_309_cast_fp16)[name = tensor("x_397_cast_fp16")]; + tensor layers_22_self_attn_q_proj_output_scale_to_fp16 = const()[name = tensor("layers_22_self_attn_q_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(228207168)))]; + tensor query_45_cast_fp16 = mul(x = x_397_cast_fp16, y = layers_22_self_attn_q_proj_output_scale_to_fp16)[name = tensor("query_45_cast_fp16")]; + tensor var_3507 = const()[name = tensor("op_3507"), val = tensor([1, 1])]; + tensor var_3509 = const()[name = tensor("op_3509"), val = tensor([1, 1])]; + tensor x_399_pad_type_0 = const()[name = tensor("x_399_pad_type_0"), val = tensor("custom")]; + tensor x_399_pad_0 = const()[name = tensor("x_399_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_22_self_attn_k_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(228209792))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(229029056))), name = tensor("layers_22_self_attn_k_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_22_self_attn_k_proj_module_bias_to_fp16 = const()[name = tensor("layers_22_self_attn_k_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(229029184)))]; + tensor x_399_cast_fp16 = conv(bias = layers_22_self_attn_k_proj_module_bias_to_fp16, dilations = var_3509, groups = var_3458, pad = x_399_pad_0, pad_type = x_399_pad_type_0, strides = var_3507, weight = layers_22_self_attn_k_proj_module_weight_to_fp16_palettized, x = input_309_cast_fp16)[name = tensor("x_399_cast_fp16")]; + tensor layers_22_self_attn_k_proj_output_scale_to_fp16 = const()[name = tensor("layers_22_self_attn_k_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(229031808)))]; + tensor key_45_cast_fp16 = mul(x = x_399_cast_fp16, y = layers_22_self_attn_k_proj_output_scale_to_fp16)[name = tensor("key_45_cast_fp16")]; + tensor var_3519 = const()[name = tensor("op_3519"), val = tensor([1, 1])]; + tensor var_3521 = const()[name = tensor("op_3521"), val = tensor([1, 1])]; + tensor x_401_pad_type_0 = const()[name = tensor("x_401_pad_type_0"), val = tensor("custom")]; + tensor x_401_pad_0 = const()[name = tensor("x_401_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_22_self_attn_v_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(229034432))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(229853696))), name = tensor("layers_22_self_attn_v_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_22_self_attn_v_proj_module_bias_to_fp16 = const()[name = tensor("layers_22_self_attn_v_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(229853824)))]; + tensor x_401_cast_fp16 = conv(bias = layers_22_self_attn_v_proj_module_bias_to_fp16, dilations = var_3521, groups = var_3458, pad = x_401_pad_0, pad_type = x_401_pad_type_0, strides = var_3519, weight = layers_22_self_attn_v_proj_module_weight_to_fp16_palettized, x = input_309_cast_fp16)[name = tensor("x_401_cast_fp16")]; + tensor layers_22_self_attn_v_proj_output_scale_to_fp16 = const()[name = tensor("layers_22_self_attn_v_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(229856448)))]; + tensor value_45_cast_fp16 = mul(x = x_401_cast_fp16, y = layers_22_self_attn_v_proj_output_scale_to_fp16)[name = tensor("value_45_cast_fp16")]; + tensor var_3526 = const()[name = tensor("op_3526"), val = tensor([1, 20, 64, -1])]; + tensor var_3527_cast_fp16 = reshape(shape = var_3526, x = query_45_cast_fp16)[name = tensor("op_3527_cast_fp16")]; + tensor var_3528_to_fp16 = const()[name = tensor("op_3528_to_fp16"), val = tensor(0x1p-3)]; + tensor var_3529_cast_fp16 = mul(x = var_3527_cast_fp16, y = var_3528_to_fp16)[name = tensor("op_3529_cast_fp16")]; + tensor var_3530 = const()[name = tensor("op_3530"), val = tensor([1, 20, 64, -1])]; + tensor var_3531_cast_fp16 = reshape(shape = var_3530, x = key_45_cast_fp16)[name = tensor("op_3531_cast_fp16")]; + tensor mh_w_45_transpose_x_0 = const()[name = tensor("mh_w_45_transpose_x_0"), val = tensor(true)]; + tensor mh_w_45_transpose_y_0 = const()[name = tensor("mh_w_45_transpose_y_0"), val = tensor(false)]; + tensor mh_w_45_cast_fp16 = matmul(transpose_x = mh_w_45_transpose_x_0, transpose_y = mh_w_45_transpose_y_0, x = var_3529_cast_fp16, y = var_3531_cast_fp16)[name = tensor("mh_w_45_cast_fp16")]; + tensor var_3534_cast_fp16 = softmax(axis = var_3456, x = mh_w_45_cast_fp16)[name = tensor("op_3534_cast_fp16")]; + tensor var_3535 = const()[name = tensor("op_3535"), val = tensor([1, 20, 64, -1])]; + tensor var_3536_cast_fp16 = reshape(shape = var_3535, x = value_45_cast_fp16)[name = tensor("op_3536_cast_fp16")]; + tensor attn_45_transpose_x_0 = const()[name = tensor("attn_45_transpose_x_0"), val = tensor(false)]; + tensor attn_45_transpose_y_0 = const()[name = tensor("attn_45_transpose_y_0"), val = tensor(true)]; + tensor attn_45_cast_fp16 = matmul(transpose_x = attn_45_transpose_x_0, transpose_y = attn_45_transpose_y_0, x = var_3536_cast_fp16, y = var_3534_cast_fp16)[name = tensor("attn_45_cast_fp16")]; + tensor var_3539 = const()[name = tensor("op_3539"), val = tensor([1, 1280, 1, -1])]; + tensor x_403_cast_fp16 = reshape(shape = var_3539, x = attn_45_cast_fp16)[name = tensor("x_403_cast_fp16")]; + tensor layers_22_self_attn_o_proj_input_shift_to_fp16 = const()[name = tensor("layers_22_self_attn_o_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(229859072)))]; + tensor input_315_cast_fp16 = sub(x = x_403_cast_fp16, y = layers_22_self_attn_o_proj_input_shift_to_fp16)[name = tensor("input_315_cast_fp16")]; + tensor var_3547 = const()[name = tensor("op_3547"), val = tensor([1, 1])]; + tensor var_3549 = const()[name = tensor("op_3549"), val = tensor([1, 1])]; + tensor x_405_pad_type_0 = const()[name = tensor("x_405_pad_type_0"), val = tensor("custom")]; + tensor x_405_pad_0 = const()[name = tensor("x_405_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_22_self_attn_o_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(229861696))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(230680960))), name = tensor("layers_22_self_attn_o_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_22_self_attn_o_proj_module_bias_to_fp16 = const()[name = tensor("layers_22_self_attn_o_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(230681088)))]; + tensor x_405_cast_fp16 = conv(bias = layers_22_self_attn_o_proj_module_bias_to_fp16, dilations = var_3549, groups = var_3458, pad = x_405_pad_0, pad_type = x_405_pad_type_0, strides = var_3547, weight = layers_22_self_attn_o_proj_module_weight_to_fp16_palettized, x = input_315_cast_fp16)[name = tensor("x_405_cast_fp16")]; + tensor layers_22_self_attn_o_proj_output_scale_to_fp16 = const()[name = tensor("layers_22_self_attn_o_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(230683712)))]; + tensor obj_91_cast_fp16 = mul(x = x_405_cast_fp16, y = layers_22_self_attn_o_proj_output_scale_to_fp16)[name = tensor("obj_91_cast_fp16")]; + tensor inputs_91_cast_fp16 = add(x = inputs_89_cast_fp16, y = obj_91_cast_fp16)[name = tensor("inputs_91_cast_fp16")]; + tensor var_3556 = const()[name = tensor("op_3556"), val = tensor([1])]; + tensor channels_mean_91_cast_fp16 = reduce_mean(axes = var_3556, keep_dims = var_3459, x = inputs_91_cast_fp16)[name = tensor("channels_mean_91_cast_fp16")]; + tensor zero_mean_91_cast_fp16 = sub(x = inputs_91_cast_fp16, y = channels_mean_91_cast_fp16)[name = tensor("zero_mean_91_cast_fp16")]; + tensor zero_mean_sq_91_cast_fp16 = mul(x = zero_mean_91_cast_fp16, y = zero_mean_91_cast_fp16)[name = tensor("zero_mean_sq_91_cast_fp16")]; + tensor var_3560 = const()[name = tensor("op_3560"), val = tensor([1])]; + tensor var_3561_cast_fp16 = reduce_mean(axes = var_3560, keep_dims = var_3459, x = zero_mean_sq_91_cast_fp16)[name = tensor("op_3561_cast_fp16")]; + tensor var_3562_to_fp16 = const()[name = tensor("op_3562_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_3563_cast_fp16 = add(x = var_3561_cast_fp16, y = var_3562_to_fp16)[name = tensor("op_3563_cast_fp16")]; + tensor denom_91_epsilon_0_to_fp16 = const()[name = tensor("denom_91_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_91_cast_fp16 = rsqrt(epsilon = denom_91_epsilon_0_to_fp16, x = var_3563_cast_fp16)[name = tensor("denom_91_cast_fp16")]; + tensor out_91_cast_fp16 = mul(x = zero_mean_91_cast_fp16, y = denom_91_cast_fp16)[name = tensor("out_91_cast_fp16")]; + tensor x_407_gamma_0_to_fp16 = const()[name = tensor("x_407_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(230686336)))]; + tensor x_407_beta_0_to_fp16 = const()[name = tensor("x_407_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(230688960)))]; + tensor x_407_epsilon_0_to_fp16 = const()[name = tensor("x_407_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_407_cast_fp16 = batch_norm(beta = x_407_beta_0_to_fp16, epsilon = x_407_epsilon_0_to_fp16, gamma = x_407_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_91_cast_fp16)[name = tensor("x_407_cast_fp16")]; + tensor layers_22_fc1_input_shift_to_fp16 = const()[name = tensor("layers_22_fc1_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(230691584)))]; + tensor input_317_cast_fp16 = sub(x = x_407_cast_fp16, y = layers_22_fc1_input_shift_to_fp16)[name = tensor("input_317_cast_fp16")]; + tensor var_3578 = const()[name = tensor("op_3578"), val = tensor([1, 1])]; + tensor var_3580 = const()[name = tensor("op_3580"), val = tensor([1, 1])]; + tensor x_409_pad_type_0 = const()[name = tensor("x_409_pad_type_0"), val = tensor("custom")]; + tensor x_409_pad_0 = const()[name = tensor("x_409_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_22_fc1_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(230694208))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(233971072))), name = tensor("layers_22_fc1_module_weight_to_fp16_palettized"), shape = tensor([5120, 1280, 1, 1])]; + tensor layers_22_fc1_module_bias_to_fp16 = const()[name = tensor("layers_22_fc1_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(233971200)))]; + tensor x_409_cast_fp16 = conv(bias = layers_22_fc1_module_bias_to_fp16, dilations = var_3580, groups = var_3458, pad = x_409_pad_0, pad_type = x_409_pad_type_0, strides = var_3578, weight = layers_22_fc1_module_weight_to_fp16_palettized, x = input_317_cast_fp16)[name = tensor("x_409_cast_fp16")]; + tensor layers_22_fc1_output_scale_to_fp16 = const()[name = tensor("layers_22_fc1_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(233981504)))]; + tensor input_319_cast_fp16 = mul(x = x_409_cast_fp16, y = layers_22_fc1_output_scale_to_fp16)[name = tensor("input_319_cast_fp16")]; + tensor x_411_mode_0 = const()[name = tensor("x_411_mode_0"), val = tensor("EXACT")]; + tensor x_411_cast_fp16 = gelu(mode = x_411_mode_0, x = input_319_cast_fp16)[name = tensor("x_411_cast_fp16")]; + tensor layers_22_fc2_input_shift_to_fp16 = const()[name = tensor("layers_22_fc2_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(233991808)))]; + tensor input_321_cast_fp16 = sub(x = x_411_cast_fp16, y = layers_22_fc2_input_shift_to_fp16)[name = tensor("input_321_cast_fp16")]; + tensor var_3591 = const()[name = tensor("op_3591"), val = tensor([1, 1])]; + tensor var_3593 = const()[name = tensor("op_3593"), val = tensor([1, 1])]; + tensor x_413_pad_type_0 = const()[name = tensor("x_413_pad_type_0"), val = tensor("custom")]; + tensor x_413_pad_0 = const()[name = tensor("x_413_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_22_fc2_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(234002112))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(237278976))), name = tensor("layers_22_fc2_module_weight_to_fp16_palettized"), shape = tensor([1280, 5120, 1, 1])]; + tensor layers_22_fc2_module_bias_to_fp16 = const()[name = tensor("layers_22_fc2_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(237279104)))]; + tensor x_413_cast_fp16 = conv(bias = layers_22_fc2_module_bias_to_fp16, dilations = var_3593, groups = var_3458, pad = x_413_pad_0, pad_type = x_413_pad_type_0, strides = var_3591, weight = layers_22_fc2_module_weight_to_fp16_palettized, x = input_321_cast_fp16)[name = tensor("x_413_cast_fp16")]; + tensor layers_22_fc2_output_scale_to_fp16 = const()[name = tensor("layers_22_fc2_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(237281728)))]; + tensor hidden_states_49_cast_fp16 = mul(x = x_413_cast_fp16, y = layers_22_fc2_output_scale_to_fp16)[name = tensor("hidden_states_49_cast_fp16")]; + tensor inputs_93_cast_fp16 = add(x = inputs_91_cast_fp16, y = hidden_states_49_cast_fp16)[name = tensor("inputs_93_cast_fp16")]; + tensor var_3605 = const()[name = tensor("op_3605"), val = tensor(3)]; + tensor var_3607 = const()[name = tensor("op_3607"), val = tensor(1)]; + tensor var_3608 = const()[name = tensor("op_3608"), val = tensor(true)]; + tensor var_3618 = const()[name = tensor("op_3618"), val = tensor([1])]; + tensor channels_mean_93_cast_fp16 = reduce_mean(axes = var_3618, keep_dims = var_3608, x = inputs_93_cast_fp16)[name = tensor("channels_mean_93_cast_fp16")]; + tensor zero_mean_93_cast_fp16 = sub(x = inputs_93_cast_fp16, y = channels_mean_93_cast_fp16)[name = tensor("zero_mean_93_cast_fp16")]; + tensor zero_mean_sq_93_cast_fp16 = mul(x = zero_mean_93_cast_fp16, y = zero_mean_93_cast_fp16)[name = tensor("zero_mean_sq_93_cast_fp16")]; + tensor var_3622 = const()[name = tensor("op_3622"), val = tensor([1])]; + tensor var_3623_cast_fp16 = reduce_mean(axes = var_3622, keep_dims = var_3608, x = zero_mean_sq_93_cast_fp16)[name = tensor("op_3623_cast_fp16")]; + tensor var_3624_to_fp16 = const()[name = tensor("op_3624_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_3625_cast_fp16 = add(x = var_3623_cast_fp16, y = var_3624_to_fp16)[name = tensor("op_3625_cast_fp16")]; + tensor denom_93_epsilon_0_to_fp16 = const()[name = tensor("denom_93_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_93_cast_fp16 = rsqrt(epsilon = denom_93_epsilon_0_to_fp16, x = var_3625_cast_fp16)[name = tensor("denom_93_cast_fp16")]; + tensor out_93_cast_fp16 = mul(x = zero_mean_93_cast_fp16, y = denom_93_cast_fp16)[name = tensor("out_93_cast_fp16")]; + tensor obj_93_gamma_0_to_fp16 = const()[name = tensor("obj_93_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(237284352)))]; + tensor obj_93_beta_0_to_fp16 = const()[name = tensor("obj_93_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(237286976)))]; + tensor obj_93_epsilon_0_to_fp16 = const()[name = tensor("obj_93_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_93_cast_fp16 = batch_norm(beta = obj_93_beta_0_to_fp16, epsilon = obj_93_epsilon_0_to_fp16, gamma = obj_93_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_93_cast_fp16)[name = tensor("obj_93_cast_fp16")]; + tensor layers_23_self_attn_q_proj_input_shift_to_fp16 = const()[name = tensor("layers_23_self_attn_q_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(237289600)))]; + tensor input_323_cast_fp16 = sub(x = obj_93_cast_fp16, y = layers_23_self_attn_q_proj_input_shift_to_fp16)[name = tensor("input_323_cast_fp16")]; + tensor var_3644 = const()[name = tensor("op_3644"), val = tensor([1, 1])]; + tensor var_3646 = const()[name = tensor("op_3646"), val = tensor([1, 1])]; + tensor x_415_pad_type_0 = const()[name = tensor("x_415_pad_type_0"), val = tensor("custom")]; + tensor x_415_pad_0 = const()[name = tensor("x_415_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_23_self_attn_q_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(237292224))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(238111488))), name = tensor("layers_23_self_attn_q_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_23_self_attn_q_proj_module_bias_to_fp16 = const()[name = tensor("layers_23_self_attn_q_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(238111616)))]; + tensor x_415_cast_fp16 = conv(bias = layers_23_self_attn_q_proj_module_bias_to_fp16, dilations = var_3646, groups = var_3607, pad = x_415_pad_0, pad_type = x_415_pad_type_0, strides = var_3644, weight = layers_23_self_attn_q_proj_module_weight_to_fp16_palettized, x = input_323_cast_fp16)[name = tensor("x_415_cast_fp16")]; + tensor layers_23_self_attn_q_proj_output_scale_to_fp16 = const()[name = tensor("layers_23_self_attn_q_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(238114240)))]; + tensor query_47_cast_fp16 = mul(x = x_415_cast_fp16, y = layers_23_self_attn_q_proj_output_scale_to_fp16)[name = tensor("query_47_cast_fp16")]; + tensor var_3656 = const()[name = tensor("op_3656"), val = tensor([1, 1])]; + tensor var_3658 = const()[name = tensor("op_3658"), val = tensor([1, 1])]; + tensor x_417_pad_type_0 = const()[name = tensor("x_417_pad_type_0"), val = tensor("custom")]; + tensor x_417_pad_0 = const()[name = tensor("x_417_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_23_self_attn_k_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(238116864))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(238936128))), name = tensor("layers_23_self_attn_k_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_23_self_attn_k_proj_module_bias_to_fp16 = const()[name = tensor("layers_23_self_attn_k_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(238936256)))]; + tensor x_417_cast_fp16 = conv(bias = layers_23_self_attn_k_proj_module_bias_to_fp16, dilations = var_3658, groups = var_3607, pad = x_417_pad_0, pad_type = x_417_pad_type_0, strides = var_3656, weight = layers_23_self_attn_k_proj_module_weight_to_fp16_palettized, x = input_323_cast_fp16)[name = tensor("x_417_cast_fp16")]; + tensor layers_23_self_attn_k_proj_output_scale_to_fp16 = const()[name = tensor("layers_23_self_attn_k_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(238938880)))]; + tensor key_47_cast_fp16 = mul(x = x_417_cast_fp16, y = layers_23_self_attn_k_proj_output_scale_to_fp16)[name = tensor("key_47_cast_fp16")]; + tensor var_3668 = const()[name = tensor("op_3668"), val = tensor([1, 1])]; + tensor var_3670 = const()[name = tensor("op_3670"), val = tensor([1, 1])]; + tensor x_419_pad_type_0 = const()[name = tensor("x_419_pad_type_0"), val = tensor("custom")]; + tensor x_419_pad_0 = const()[name = tensor("x_419_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_23_self_attn_v_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(238941504))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(239760768))), name = tensor("layers_23_self_attn_v_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_23_self_attn_v_proj_module_bias_to_fp16 = const()[name = tensor("layers_23_self_attn_v_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(239760896)))]; + tensor x_419_cast_fp16 = conv(bias = layers_23_self_attn_v_proj_module_bias_to_fp16, dilations = var_3670, groups = var_3607, pad = x_419_pad_0, pad_type = x_419_pad_type_0, strides = var_3668, weight = layers_23_self_attn_v_proj_module_weight_to_fp16_palettized, x = input_323_cast_fp16)[name = tensor("x_419_cast_fp16")]; + tensor layers_23_self_attn_v_proj_output_scale_to_fp16 = const()[name = tensor("layers_23_self_attn_v_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(239763520)))]; + tensor value_47_cast_fp16 = mul(x = x_419_cast_fp16, y = layers_23_self_attn_v_proj_output_scale_to_fp16)[name = tensor("value_47_cast_fp16")]; + tensor var_3675 = const()[name = tensor("op_3675"), val = tensor([1, 20, 64, -1])]; + tensor var_3676_cast_fp16 = reshape(shape = var_3675, x = query_47_cast_fp16)[name = tensor("op_3676_cast_fp16")]; + tensor var_3677_to_fp16 = const()[name = tensor("op_3677_to_fp16"), val = tensor(0x1p-3)]; + tensor var_3678_cast_fp16 = mul(x = var_3676_cast_fp16, y = var_3677_to_fp16)[name = tensor("op_3678_cast_fp16")]; + tensor var_3679 = const()[name = tensor("op_3679"), val = tensor([1, 20, 64, -1])]; + tensor var_3680_cast_fp16 = reshape(shape = var_3679, x = key_47_cast_fp16)[name = tensor("op_3680_cast_fp16")]; + tensor mh_w_47_transpose_x_0 = const()[name = tensor("mh_w_47_transpose_x_0"), val = tensor(true)]; + tensor mh_w_47_transpose_y_0 = const()[name = tensor("mh_w_47_transpose_y_0"), val = tensor(false)]; + tensor mh_w_47_cast_fp16 = matmul(transpose_x = mh_w_47_transpose_x_0, transpose_y = mh_w_47_transpose_y_0, x = var_3678_cast_fp16, y = var_3680_cast_fp16)[name = tensor("mh_w_47_cast_fp16")]; + tensor var_3683_cast_fp16 = softmax(axis = var_3605, x = mh_w_47_cast_fp16)[name = tensor("op_3683_cast_fp16")]; + tensor var_3684 = const()[name = tensor("op_3684"), val = tensor([1, 20, 64, -1])]; + tensor var_3685_cast_fp16 = reshape(shape = var_3684, x = value_47_cast_fp16)[name = tensor("op_3685_cast_fp16")]; + tensor attn_47_transpose_x_0 = const()[name = tensor("attn_47_transpose_x_0"), val = tensor(false)]; + tensor attn_47_transpose_y_0 = const()[name = tensor("attn_47_transpose_y_0"), val = tensor(true)]; + tensor attn_47_cast_fp16 = matmul(transpose_x = attn_47_transpose_x_0, transpose_y = attn_47_transpose_y_0, x = var_3685_cast_fp16, y = var_3683_cast_fp16)[name = tensor("attn_47_cast_fp16")]; + tensor var_3688 = const()[name = tensor("op_3688"), val = tensor([1, 1280, 1, -1])]; + tensor x_421_cast_fp16 = reshape(shape = var_3688, x = attn_47_cast_fp16)[name = tensor("x_421_cast_fp16")]; + tensor layers_23_self_attn_o_proj_input_shift_to_fp16 = const()[name = tensor("layers_23_self_attn_o_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(239766144)))]; + tensor input_329_cast_fp16 = sub(x = x_421_cast_fp16, y = layers_23_self_attn_o_proj_input_shift_to_fp16)[name = tensor("input_329_cast_fp16")]; + tensor var_3696 = const()[name = tensor("op_3696"), val = tensor([1, 1])]; + tensor var_3698 = const()[name = tensor("op_3698"), val = tensor([1, 1])]; + tensor x_423_pad_type_0 = const()[name = tensor("x_423_pad_type_0"), val = tensor("custom")]; + tensor x_423_pad_0 = const()[name = tensor("x_423_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_23_self_attn_o_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(239768768))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(240588032))), name = tensor("layers_23_self_attn_o_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_23_self_attn_o_proj_module_bias_to_fp16 = const()[name = tensor("layers_23_self_attn_o_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(240588160)))]; + tensor x_423_cast_fp16 = conv(bias = layers_23_self_attn_o_proj_module_bias_to_fp16, dilations = var_3698, groups = var_3607, pad = x_423_pad_0, pad_type = x_423_pad_type_0, strides = var_3696, weight = layers_23_self_attn_o_proj_module_weight_to_fp16_palettized, x = input_329_cast_fp16)[name = tensor("x_423_cast_fp16")]; + tensor layers_23_self_attn_o_proj_output_scale_to_fp16 = const()[name = tensor("layers_23_self_attn_o_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(240590784)))]; + tensor obj_95_cast_fp16 = mul(x = x_423_cast_fp16, y = layers_23_self_attn_o_proj_output_scale_to_fp16)[name = tensor("obj_95_cast_fp16")]; + tensor inputs_95_cast_fp16 = add(x = inputs_93_cast_fp16, y = obj_95_cast_fp16)[name = tensor("inputs_95_cast_fp16")]; + tensor var_3705 = const()[name = tensor("op_3705"), val = tensor([1])]; + tensor channels_mean_95_cast_fp16 = reduce_mean(axes = var_3705, keep_dims = var_3608, x = inputs_95_cast_fp16)[name = tensor("channels_mean_95_cast_fp16")]; + tensor zero_mean_95_cast_fp16 = sub(x = inputs_95_cast_fp16, y = channels_mean_95_cast_fp16)[name = tensor("zero_mean_95_cast_fp16")]; + tensor zero_mean_sq_95_cast_fp16 = mul(x = zero_mean_95_cast_fp16, y = zero_mean_95_cast_fp16)[name = tensor("zero_mean_sq_95_cast_fp16")]; + tensor var_3709 = const()[name = tensor("op_3709"), val = tensor([1])]; + tensor var_3710_cast_fp16 = reduce_mean(axes = var_3709, keep_dims = var_3608, x = zero_mean_sq_95_cast_fp16)[name = tensor("op_3710_cast_fp16")]; + tensor var_3711_to_fp16 = const()[name = tensor("op_3711_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_3712_cast_fp16 = add(x = var_3710_cast_fp16, y = var_3711_to_fp16)[name = tensor("op_3712_cast_fp16")]; + tensor denom_95_epsilon_0_to_fp16 = const()[name = tensor("denom_95_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_95_cast_fp16 = rsqrt(epsilon = denom_95_epsilon_0_to_fp16, x = var_3712_cast_fp16)[name = tensor("denom_95_cast_fp16")]; + tensor out_95_cast_fp16 = mul(x = zero_mean_95_cast_fp16, y = denom_95_cast_fp16)[name = tensor("out_95_cast_fp16")]; + tensor x_425_gamma_0_to_fp16 = const()[name = tensor("x_425_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(240593408)))]; + tensor x_425_beta_0_to_fp16 = const()[name = tensor("x_425_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(240596032)))]; + tensor x_425_epsilon_0_to_fp16 = const()[name = tensor("x_425_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_425_cast_fp16 = batch_norm(beta = x_425_beta_0_to_fp16, epsilon = x_425_epsilon_0_to_fp16, gamma = x_425_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_95_cast_fp16)[name = tensor("x_425_cast_fp16")]; + tensor layers_23_fc1_input_shift_to_fp16 = const()[name = tensor("layers_23_fc1_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(240598656)))]; + tensor input_331_cast_fp16 = sub(x = x_425_cast_fp16, y = layers_23_fc1_input_shift_to_fp16)[name = tensor("input_331_cast_fp16")]; + tensor var_3727 = const()[name = tensor("op_3727"), val = tensor([1, 1])]; + tensor var_3729 = const()[name = tensor("op_3729"), val = tensor([1, 1])]; + tensor x_427_pad_type_0 = const()[name = tensor("x_427_pad_type_0"), val = tensor("custom")]; + tensor x_427_pad_0 = const()[name = tensor("x_427_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_23_fc1_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(240601280))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(243878144))), name = tensor("layers_23_fc1_module_weight_to_fp16_palettized"), shape = tensor([5120, 1280, 1, 1])]; + tensor layers_23_fc1_module_bias_to_fp16 = const()[name = tensor("layers_23_fc1_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(243878272)))]; + tensor x_427_cast_fp16 = conv(bias = layers_23_fc1_module_bias_to_fp16, dilations = var_3729, groups = var_3607, pad = x_427_pad_0, pad_type = x_427_pad_type_0, strides = var_3727, weight = layers_23_fc1_module_weight_to_fp16_palettized, x = input_331_cast_fp16)[name = tensor("x_427_cast_fp16")]; + tensor layers_23_fc1_output_scale_to_fp16 = const()[name = tensor("layers_23_fc1_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(243888576)))]; + tensor input_333_cast_fp16 = mul(x = x_427_cast_fp16, y = layers_23_fc1_output_scale_to_fp16)[name = tensor("input_333_cast_fp16")]; + tensor x_429_mode_0 = const()[name = tensor("x_429_mode_0"), val = tensor("EXACT")]; + tensor x_429_cast_fp16 = gelu(mode = x_429_mode_0, x = input_333_cast_fp16)[name = tensor("x_429_cast_fp16")]; + tensor layers_23_fc2_input_shift_to_fp16 = const()[name = tensor("layers_23_fc2_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(243898880)))]; + tensor input_335_cast_fp16 = sub(x = x_429_cast_fp16, y = layers_23_fc2_input_shift_to_fp16)[name = tensor("input_335_cast_fp16")]; + tensor var_3740 = const()[name = tensor("op_3740"), val = tensor([1, 1])]; + tensor var_3742 = const()[name = tensor("op_3742"), val = tensor([1, 1])]; + tensor x_431_pad_type_0 = const()[name = tensor("x_431_pad_type_0"), val = tensor("custom")]; + tensor x_431_pad_0 = const()[name = tensor("x_431_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_23_fc2_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(243909184))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(247186048))), name = tensor("layers_23_fc2_module_weight_to_fp16_palettized"), shape = tensor([1280, 5120, 1, 1])]; + tensor layers_23_fc2_module_bias_to_fp16 = const()[name = tensor("layers_23_fc2_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(247186176)))]; + tensor x_431_cast_fp16 = conv(bias = layers_23_fc2_module_bias_to_fp16, dilations = var_3742, groups = var_3607, pad = x_431_pad_0, pad_type = x_431_pad_type_0, strides = var_3740, weight = layers_23_fc2_module_weight_to_fp16_palettized, x = input_335_cast_fp16)[name = tensor("x_431_cast_fp16")]; + tensor layers_23_fc2_output_scale_to_fp16 = const()[name = tensor("layers_23_fc2_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(247188800)))]; + tensor hidden_states_51_cast_fp16 = mul(x = x_431_cast_fp16, y = layers_23_fc2_output_scale_to_fp16)[name = tensor("hidden_states_51_cast_fp16")]; + tensor inputs_97_cast_fp16 = add(x = inputs_95_cast_fp16, y = hidden_states_51_cast_fp16)[name = tensor("inputs_97_cast_fp16")]; + tensor var_3754 = const()[name = tensor("op_3754"), val = tensor(3)]; + tensor var_3756 = const()[name = tensor("op_3756"), val = tensor(1)]; + tensor var_3757 = const()[name = tensor("op_3757"), val = tensor(true)]; + tensor var_3767 = const()[name = tensor("op_3767"), val = tensor([1])]; + tensor channels_mean_97_cast_fp16 = reduce_mean(axes = var_3767, keep_dims = var_3757, x = inputs_97_cast_fp16)[name = tensor("channels_mean_97_cast_fp16")]; + tensor zero_mean_97_cast_fp16 = sub(x = inputs_97_cast_fp16, y = channels_mean_97_cast_fp16)[name = tensor("zero_mean_97_cast_fp16")]; + tensor zero_mean_sq_97_cast_fp16 = mul(x = zero_mean_97_cast_fp16, y = zero_mean_97_cast_fp16)[name = tensor("zero_mean_sq_97_cast_fp16")]; + tensor var_3771 = const()[name = tensor("op_3771"), val = tensor([1])]; + tensor var_3772_cast_fp16 = reduce_mean(axes = var_3771, keep_dims = var_3757, x = zero_mean_sq_97_cast_fp16)[name = tensor("op_3772_cast_fp16")]; + tensor var_3773_to_fp16 = const()[name = tensor("op_3773_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_3774_cast_fp16 = add(x = var_3772_cast_fp16, y = var_3773_to_fp16)[name = tensor("op_3774_cast_fp16")]; + tensor denom_97_epsilon_0_to_fp16 = const()[name = tensor("denom_97_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_97_cast_fp16 = rsqrt(epsilon = denom_97_epsilon_0_to_fp16, x = var_3774_cast_fp16)[name = tensor("denom_97_cast_fp16")]; + tensor out_97_cast_fp16 = mul(x = zero_mean_97_cast_fp16, y = denom_97_cast_fp16)[name = tensor("out_97_cast_fp16")]; + tensor obj_97_gamma_0_to_fp16 = const()[name = tensor("obj_97_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(247191424)))]; + tensor obj_97_beta_0_to_fp16 = const()[name = tensor("obj_97_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(247194048)))]; + tensor obj_97_epsilon_0_to_fp16 = const()[name = tensor("obj_97_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_97_cast_fp16 = batch_norm(beta = obj_97_beta_0_to_fp16, epsilon = obj_97_epsilon_0_to_fp16, gamma = obj_97_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_97_cast_fp16)[name = tensor("obj_97_cast_fp16")]; + tensor layers_24_self_attn_q_proj_input_shift_to_fp16 = const()[name = tensor("layers_24_self_attn_q_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(247196672)))]; + tensor input_337_cast_fp16 = sub(x = obj_97_cast_fp16, y = layers_24_self_attn_q_proj_input_shift_to_fp16)[name = tensor("input_337_cast_fp16")]; + tensor var_3793 = const()[name = tensor("op_3793"), val = tensor([1, 1])]; + tensor var_3795 = const()[name = tensor("op_3795"), val = tensor([1, 1])]; + tensor x_433_pad_type_0 = const()[name = tensor("x_433_pad_type_0"), val = tensor("custom")]; + tensor x_433_pad_0 = const()[name = tensor("x_433_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_24_self_attn_q_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(247199296))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(248018560))), name = tensor("layers_24_self_attn_q_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_24_self_attn_q_proj_module_bias_to_fp16 = const()[name = tensor("layers_24_self_attn_q_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(248018688)))]; + tensor x_433_cast_fp16 = conv(bias = layers_24_self_attn_q_proj_module_bias_to_fp16, dilations = var_3795, groups = var_3756, pad = x_433_pad_0, pad_type = x_433_pad_type_0, strides = var_3793, weight = layers_24_self_attn_q_proj_module_weight_to_fp16_palettized, x = input_337_cast_fp16)[name = tensor("x_433_cast_fp16")]; + tensor layers_24_self_attn_q_proj_output_scale_to_fp16 = const()[name = tensor("layers_24_self_attn_q_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(248021312)))]; + tensor query_49_cast_fp16 = mul(x = x_433_cast_fp16, y = layers_24_self_attn_q_proj_output_scale_to_fp16)[name = tensor("query_49_cast_fp16")]; + tensor var_3805 = const()[name = tensor("op_3805"), val = tensor([1, 1])]; + tensor var_3807 = const()[name = tensor("op_3807"), val = tensor([1, 1])]; + tensor x_435_pad_type_0 = const()[name = tensor("x_435_pad_type_0"), val = tensor("custom")]; + tensor x_435_pad_0 = const()[name = tensor("x_435_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_24_self_attn_k_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(248023936))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(248843200))), name = tensor("layers_24_self_attn_k_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_24_self_attn_k_proj_module_bias_to_fp16 = const()[name = tensor("layers_24_self_attn_k_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(248843328)))]; + tensor x_435_cast_fp16 = conv(bias = layers_24_self_attn_k_proj_module_bias_to_fp16, dilations = var_3807, groups = var_3756, pad = x_435_pad_0, pad_type = x_435_pad_type_0, strides = var_3805, weight = layers_24_self_attn_k_proj_module_weight_to_fp16_palettized, x = input_337_cast_fp16)[name = tensor("x_435_cast_fp16")]; + tensor layers_24_self_attn_k_proj_output_scale_to_fp16 = const()[name = tensor("layers_24_self_attn_k_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(248845952)))]; + tensor key_49_cast_fp16 = mul(x = x_435_cast_fp16, y = layers_24_self_attn_k_proj_output_scale_to_fp16)[name = tensor("key_49_cast_fp16")]; + tensor var_3817 = const()[name = tensor("op_3817"), val = tensor([1, 1])]; + tensor var_3819 = const()[name = tensor("op_3819"), val = tensor([1, 1])]; + tensor x_437_pad_type_0 = const()[name = tensor("x_437_pad_type_0"), val = tensor("custom")]; + tensor x_437_pad_0 = const()[name = tensor("x_437_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_24_self_attn_v_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(248848576))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(249667840))), name = tensor("layers_24_self_attn_v_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_24_self_attn_v_proj_module_bias_to_fp16 = const()[name = tensor("layers_24_self_attn_v_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(249667968)))]; + tensor x_437_cast_fp16 = conv(bias = layers_24_self_attn_v_proj_module_bias_to_fp16, dilations = var_3819, groups = var_3756, pad = x_437_pad_0, pad_type = x_437_pad_type_0, strides = var_3817, weight = layers_24_self_attn_v_proj_module_weight_to_fp16_palettized, x = input_337_cast_fp16)[name = tensor("x_437_cast_fp16")]; + tensor layers_24_self_attn_v_proj_output_scale_to_fp16 = const()[name = tensor("layers_24_self_attn_v_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(249670592)))]; + tensor value_49_cast_fp16 = mul(x = x_437_cast_fp16, y = layers_24_self_attn_v_proj_output_scale_to_fp16)[name = tensor("value_49_cast_fp16")]; + tensor var_3824 = const()[name = tensor("op_3824"), val = tensor([1, 20, 64, -1])]; + tensor var_3825_cast_fp16 = reshape(shape = var_3824, x = query_49_cast_fp16)[name = tensor("op_3825_cast_fp16")]; + tensor var_3826_to_fp16 = const()[name = tensor("op_3826_to_fp16"), val = tensor(0x1p-3)]; + tensor var_3827_cast_fp16 = mul(x = var_3825_cast_fp16, y = var_3826_to_fp16)[name = tensor("op_3827_cast_fp16")]; + tensor var_3828 = const()[name = tensor("op_3828"), val = tensor([1, 20, 64, -1])]; + tensor var_3829_cast_fp16 = reshape(shape = var_3828, x = key_49_cast_fp16)[name = tensor("op_3829_cast_fp16")]; + tensor mh_w_49_transpose_x_0 = const()[name = tensor("mh_w_49_transpose_x_0"), val = tensor(true)]; + tensor mh_w_49_transpose_y_0 = const()[name = tensor("mh_w_49_transpose_y_0"), val = tensor(false)]; + tensor mh_w_49_cast_fp16 = matmul(transpose_x = mh_w_49_transpose_x_0, transpose_y = mh_w_49_transpose_y_0, x = var_3827_cast_fp16, y = var_3829_cast_fp16)[name = tensor("mh_w_49_cast_fp16")]; + tensor var_3832_cast_fp16 = softmax(axis = var_3754, x = mh_w_49_cast_fp16)[name = tensor("op_3832_cast_fp16")]; + tensor var_3833 = const()[name = tensor("op_3833"), val = tensor([1, 20, 64, -1])]; + tensor var_3834_cast_fp16 = reshape(shape = var_3833, x = value_49_cast_fp16)[name = tensor("op_3834_cast_fp16")]; + tensor attn_49_transpose_x_0 = const()[name = tensor("attn_49_transpose_x_0"), val = tensor(false)]; + tensor attn_49_transpose_y_0 = const()[name = tensor("attn_49_transpose_y_0"), val = tensor(true)]; + tensor attn_49_cast_fp16 = matmul(transpose_x = attn_49_transpose_x_0, transpose_y = attn_49_transpose_y_0, x = var_3834_cast_fp16, y = var_3832_cast_fp16)[name = tensor("attn_49_cast_fp16")]; + tensor var_3837 = const()[name = tensor("op_3837"), val = tensor([1, 1280, 1, -1])]; + tensor x_439_cast_fp16 = reshape(shape = var_3837, x = attn_49_cast_fp16)[name = tensor("x_439_cast_fp16")]; + tensor layers_24_self_attn_o_proj_input_shift_to_fp16 = const()[name = tensor("layers_24_self_attn_o_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(249673216)))]; + tensor input_343_cast_fp16 = sub(x = x_439_cast_fp16, y = layers_24_self_attn_o_proj_input_shift_to_fp16)[name = tensor("input_343_cast_fp16")]; + tensor var_3845 = const()[name = tensor("op_3845"), val = tensor([1, 1])]; + tensor var_3847 = const()[name = tensor("op_3847"), val = tensor([1, 1])]; + tensor x_441_pad_type_0 = const()[name = tensor("x_441_pad_type_0"), val = tensor("custom")]; + tensor x_441_pad_0 = const()[name = tensor("x_441_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_24_self_attn_o_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(249675840))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(250495104))), name = tensor("layers_24_self_attn_o_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_24_self_attn_o_proj_module_bias_to_fp16 = const()[name = tensor("layers_24_self_attn_o_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(250495232)))]; + tensor x_441_cast_fp16 = conv(bias = layers_24_self_attn_o_proj_module_bias_to_fp16, dilations = var_3847, groups = var_3756, pad = x_441_pad_0, pad_type = x_441_pad_type_0, strides = var_3845, weight = layers_24_self_attn_o_proj_module_weight_to_fp16_palettized, x = input_343_cast_fp16)[name = tensor("x_441_cast_fp16")]; + tensor layers_24_self_attn_o_proj_output_scale_to_fp16 = const()[name = tensor("layers_24_self_attn_o_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(250497856)))]; + tensor obj_99_cast_fp16 = mul(x = x_441_cast_fp16, y = layers_24_self_attn_o_proj_output_scale_to_fp16)[name = tensor("obj_99_cast_fp16")]; + tensor inputs_99_cast_fp16 = add(x = inputs_97_cast_fp16, y = obj_99_cast_fp16)[name = tensor("inputs_99_cast_fp16")]; + tensor var_3854 = const()[name = tensor("op_3854"), val = tensor([1])]; + tensor channels_mean_99_cast_fp16 = reduce_mean(axes = var_3854, keep_dims = var_3757, x = inputs_99_cast_fp16)[name = tensor("channels_mean_99_cast_fp16")]; + tensor zero_mean_99_cast_fp16 = sub(x = inputs_99_cast_fp16, y = channels_mean_99_cast_fp16)[name = tensor("zero_mean_99_cast_fp16")]; + tensor zero_mean_sq_99_cast_fp16 = mul(x = zero_mean_99_cast_fp16, y = zero_mean_99_cast_fp16)[name = tensor("zero_mean_sq_99_cast_fp16")]; + tensor var_3858 = const()[name = tensor("op_3858"), val = tensor([1])]; + tensor var_3859_cast_fp16 = reduce_mean(axes = var_3858, keep_dims = var_3757, x = zero_mean_sq_99_cast_fp16)[name = tensor("op_3859_cast_fp16")]; + tensor var_3860_to_fp16 = const()[name = tensor("op_3860_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_3861_cast_fp16 = add(x = var_3859_cast_fp16, y = var_3860_to_fp16)[name = tensor("op_3861_cast_fp16")]; + tensor denom_99_epsilon_0_to_fp16 = const()[name = tensor("denom_99_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_99_cast_fp16 = rsqrt(epsilon = denom_99_epsilon_0_to_fp16, x = var_3861_cast_fp16)[name = tensor("denom_99_cast_fp16")]; + tensor out_99_cast_fp16 = mul(x = zero_mean_99_cast_fp16, y = denom_99_cast_fp16)[name = tensor("out_99_cast_fp16")]; + tensor x_443_gamma_0_to_fp16 = const()[name = tensor("x_443_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(250500480)))]; + tensor x_443_beta_0_to_fp16 = const()[name = tensor("x_443_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(250503104)))]; + tensor x_443_epsilon_0_to_fp16 = const()[name = tensor("x_443_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_443_cast_fp16 = batch_norm(beta = x_443_beta_0_to_fp16, epsilon = x_443_epsilon_0_to_fp16, gamma = x_443_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_99_cast_fp16)[name = tensor("x_443_cast_fp16")]; + tensor layers_24_fc1_input_shift_to_fp16 = const()[name = tensor("layers_24_fc1_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(250505728)))]; + tensor input_345_cast_fp16 = sub(x = x_443_cast_fp16, y = layers_24_fc1_input_shift_to_fp16)[name = tensor("input_345_cast_fp16")]; + tensor var_3876 = const()[name = tensor("op_3876"), val = tensor([1, 1])]; + tensor var_3878 = const()[name = tensor("op_3878"), val = tensor([1, 1])]; + tensor x_445_pad_type_0 = const()[name = tensor("x_445_pad_type_0"), val = tensor("custom")]; + tensor x_445_pad_0 = const()[name = tensor("x_445_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_24_fc1_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(250508352))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(253785216))), name = tensor("layers_24_fc1_module_weight_to_fp16_palettized"), shape = tensor([5120, 1280, 1, 1])]; + tensor layers_24_fc1_module_bias_to_fp16 = const()[name = tensor("layers_24_fc1_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(253785344)))]; + tensor x_445_cast_fp16 = conv(bias = layers_24_fc1_module_bias_to_fp16, dilations = var_3878, groups = var_3756, pad = x_445_pad_0, pad_type = x_445_pad_type_0, strides = var_3876, weight = layers_24_fc1_module_weight_to_fp16_palettized, x = input_345_cast_fp16)[name = tensor("x_445_cast_fp16")]; + tensor layers_24_fc1_output_scale_to_fp16 = const()[name = tensor("layers_24_fc1_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(253795648)))]; + tensor input_347_cast_fp16 = mul(x = x_445_cast_fp16, y = layers_24_fc1_output_scale_to_fp16)[name = tensor("input_347_cast_fp16")]; + tensor x_447_mode_0 = const()[name = tensor("x_447_mode_0"), val = tensor("EXACT")]; + tensor x_447_cast_fp16 = gelu(mode = x_447_mode_0, x = input_347_cast_fp16)[name = tensor("x_447_cast_fp16")]; + tensor layers_24_fc2_input_shift_to_fp16 = const()[name = tensor("layers_24_fc2_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(253805952)))]; + tensor input_349_cast_fp16 = sub(x = x_447_cast_fp16, y = layers_24_fc2_input_shift_to_fp16)[name = tensor("input_349_cast_fp16")]; + tensor var_3889 = const()[name = tensor("op_3889"), val = tensor([1, 1])]; + tensor var_3891 = const()[name = tensor("op_3891"), val = tensor([1, 1])]; + tensor x_449_pad_type_0 = const()[name = tensor("x_449_pad_type_0"), val = tensor("custom")]; + tensor x_449_pad_0 = const()[name = tensor("x_449_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_24_fc2_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(253816256))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(257093120))), name = tensor("layers_24_fc2_module_weight_to_fp16_palettized"), shape = tensor([1280, 5120, 1, 1])]; + tensor layers_24_fc2_module_bias_to_fp16 = const()[name = tensor("layers_24_fc2_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(257093248)))]; + tensor x_449_cast_fp16 = conv(bias = layers_24_fc2_module_bias_to_fp16, dilations = var_3891, groups = var_3756, pad = x_449_pad_0, pad_type = x_449_pad_type_0, strides = var_3889, weight = layers_24_fc2_module_weight_to_fp16_palettized, x = input_349_cast_fp16)[name = tensor("x_449_cast_fp16")]; + tensor layers_24_fc2_output_scale_to_fp16 = const()[name = tensor("layers_24_fc2_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(257095872)))]; + tensor hidden_states_53_cast_fp16 = mul(x = x_449_cast_fp16, y = layers_24_fc2_output_scale_to_fp16)[name = tensor("hidden_states_53_cast_fp16")]; + tensor inputs_101_cast_fp16 = add(x = inputs_99_cast_fp16, y = hidden_states_53_cast_fp16)[name = tensor("inputs_101_cast_fp16")]; + tensor var_3903 = const()[name = tensor("op_3903"), val = tensor(3)]; + tensor var_3905 = const()[name = tensor("op_3905"), val = tensor(1)]; + tensor var_3906 = const()[name = tensor("op_3906"), val = tensor(true)]; + tensor var_3916 = const()[name = tensor("op_3916"), val = tensor([1])]; + tensor channels_mean_101_cast_fp16 = reduce_mean(axes = var_3916, keep_dims = var_3906, x = inputs_101_cast_fp16)[name = tensor("channels_mean_101_cast_fp16")]; + tensor zero_mean_101_cast_fp16 = sub(x = inputs_101_cast_fp16, y = channels_mean_101_cast_fp16)[name = tensor("zero_mean_101_cast_fp16")]; + tensor zero_mean_sq_101_cast_fp16 = mul(x = zero_mean_101_cast_fp16, y = zero_mean_101_cast_fp16)[name = tensor("zero_mean_sq_101_cast_fp16")]; + tensor var_3920 = const()[name = tensor("op_3920"), val = tensor([1])]; + tensor var_3921_cast_fp16 = reduce_mean(axes = var_3920, keep_dims = var_3906, x = zero_mean_sq_101_cast_fp16)[name = tensor("op_3921_cast_fp16")]; + tensor var_3922_to_fp16 = const()[name = tensor("op_3922_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_3923_cast_fp16 = add(x = var_3921_cast_fp16, y = var_3922_to_fp16)[name = tensor("op_3923_cast_fp16")]; + tensor denom_101_epsilon_0_to_fp16 = const()[name = tensor("denom_101_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_101_cast_fp16 = rsqrt(epsilon = denom_101_epsilon_0_to_fp16, x = var_3923_cast_fp16)[name = tensor("denom_101_cast_fp16")]; + tensor out_101_cast_fp16 = mul(x = zero_mean_101_cast_fp16, y = denom_101_cast_fp16)[name = tensor("out_101_cast_fp16")]; + tensor obj_101_gamma_0_to_fp16 = const()[name = tensor("obj_101_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(257098496)))]; + tensor obj_101_beta_0_to_fp16 = const()[name = tensor("obj_101_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(257101120)))]; + tensor obj_101_epsilon_0_to_fp16 = const()[name = tensor("obj_101_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_101_cast_fp16 = batch_norm(beta = obj_101_beta_0_to_fp16, epsilon = obj_101_epsilon_0_to_fp16, gamma = obj_101_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_101_cast_fp16)[name = tensor("obj_101_cast_fp16")]; + tensor layers_25_self_attn_q_proj_input_shift_to_fp16 = const()[name = tensor("layers_25_self_attn_q_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(257103744)))]; + tensor input_351_cast_fp16 = sub(x = obj_101_cast_fp16, y = layers_25_self_attn_q_proj_input_shift_to_fp16)[name = tensor("input_351_cast_fp16")]; + tensor var_3942 = const()[name = tensor("op_3942"), val = tensor([1, 1])]; + tensor var_3944 = const()[name = tensor("op_3944"), val = tensor([1, 1])]; + tensor x_451_pad_type_0 = const()[name = tensor("x_451_pad_type_0"), val = tensor("custom")]; + tensor x_451_pad_0 = const()[name = tensor("x_451_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_25_self_attn_q_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(257106368))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(257925632))), name = tensor("layers_25_self_attn_q_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_25_self_attn_q_proj_module_bias_to_fp16 = const()[name = tensor("layers_25_self_attn_q_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(257925760)))]; + tensor x_451_cast_fp16 = conv(bias = layers_25_self_attn_q_proj_module_bias_to_fp16, dilations = var_3944, groups = var_3905, pad = x_451_pad_0, pad_type = x_451_pad_type_0, strides = var_3942, weight = layers_25_self_attn_q_proj_module_weight_to_fp16_palettized, x = input_351_cast_fp16)[name = tensor("x_451_cast_fp16")]; + tensor layers_25_self_attn_q_proj_output_scale_to_fp16 = const()[name = tensor("layers_25_self_attn_q_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(257928384)))]; + tensor query_51_cast_fp16 = mul(x = x_451_cast_fp16, y = layers_25_self_attn_q_proj_output_scale_to_fp16)[name = tensor("query_51_cast_fp16")]; + tensor var_3954 = const()[name = tensor("op_3954"), val = tensor([1, 1])]; + tensor var_3956 = const()[name = tensor("op_3956"), val = tensor([1, 1])]; + tensor x_453_pad_type_0 = const()[name = tensor("x_453_pad_type_0"), val = tensor("custom")]; + tensor x_453_pad_0 = const()[name = tensor("x_453_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_25_self_attn_k_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(257931008))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(258750272))), name = tensor("layers_25_self_attn_k_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_25_self_attn_k_proj_module_bias_to_fp16 = const()[name = tensor("layers_25_self_attn_k_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(258750400)))]; + tensor x_453_cast_fp16 = conv(bias = layers_25_self_attn_k_proj_module_bias_to_fp16, dilations = var_3956, groups = var_3905, pad = x_453_pad_0, pad_type = x_453_pad_type_0, strides = var_3954, weight = layers_25_self_attn_k_proj_module_weight_to_fp16_palettized, x = input_351_cast_fp16)[name = tensor("x_453_cast_fp16")]; + tensor layers_25_self_attn_k_proj_output_scale_to_fp16 = const()[name = tensor("layers_25_self_attn_k_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(258753024)))]; + tensor key_51_cast_fp16 = mul(x = x_453_cast_fp16, y = layers_25_self_attn_k_proj_output_scale_to_fp16)[name = tensor("key_51_cast_fp16")]; + tensor var_3966 = const()[name = tensor("op_3966"), val = tensor([1, 1])]; + tensor var_3968 = const()[name = tensor("op_3968"), val = tensor([1, 1])]; + tensor x_455_pad_type_0 = const()[name = tensor("x_455_pad_type_0"), val = tensor("custom")]; + tensor x_455_pad_0 = const()[name = tensor("x_455_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_25_self_attn_v_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(258755648))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(259574912))), name = tensor("layers_25_self_attn_v_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_25_self_attn_v_proj_module_bias_to_fp16 = const()[name = tensor("layers_25_self_attn_v_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(259575040)))]; + tensor x_455_cast_fp16 = conv(bias = layers_25_self_attn_v_proj_module_bias_to_fp16, dilations = var_3968, groups = var_3905, pad = x_455_pad_0, pad_type = x_455_pad_type_0, strides = var_3966, weight = layers_25_self_attn_v_proj_module_weight_to_fp16_palettized, x = input_351_cast_fp16)[name = tensor("x_455_cast_fp16")]; + tensor layers_25_self_attn_v_proj_output_scale_to_fp16 = const()[name = tensor("layers_25_self_attn_v_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(259577664)))]; + tensor value_51_cast_fp16 = mul(x = x_455_cast_fp16, y = layers_25_self_attn_v_proj_output_scale_to_fp16)[name = tensor("value_51_cast_fp16")]; + tensor var_3973 = const()[name = tensor("op_3973"), val = tensor([1, 20, 64, -1])]; + tensor var_3974_cast_fp16 = reshape(shape = var_3973, x = query_51_cast_fp16)[name = tensor("op_3974_cast_fp16")]; + tensor var_3975_to_fp16 = const()[name = tensor("op_3975_to_fp16"), val = tensor(0x1p-3)]; + tensor var_3976_cast_fp16 = mul(x = var_3974_cast_fp16, y = var_3975_to_fp16)[name = tensor("op_3976_cast_fp16")]; + tensor var_3977 = const()[name = tensor("op_3977"), val = tensor([1, 20, 64, -1])]; + tensor var_3978_cast_fp16 = reshape(shape = var_3977, x = key_51_cast_fp16)[name = tensor("op_3978_cast_fp16")]; + tensor mh_w_51_transpose_x_0 = const()[name = tensor("mh_w_51_transpose_x_0"), val = tensor(true)]; + tensor mh_w_51_transpose_y_0 = const()[name = tensor("mh_w_51_transpose_y_0"), val = tensor(false)]; + tensor mh_w_51_cast_fp16 = matmul(transpose_x = mh_w_51_transpose_x_0, transpose_y = mh_w_51_transpose_y_0, x = var_3976_cast_fp16, y = var_3978_cast_fp16)[name = tensor("mh_w_51_cast_fp16")]; + tensor var_3981_cast_fp16 = softmax(axis = var_3903, x = mh_w_51_cast_fp16)[name = tensor("op_3981_cast_fp16")]; + tensor var_3982 = const()[name = tensor("op_3982"), val = tensor([1, 20, 64, -1])]; + tensor var_3983_cast_fp16 = reshape(shape = var_3982, x = value_51_cast_fp16)[name = tensor("op_3983_cast_fp16")]; + tensor attn_51_transpose_x_0 = const()[name = tensor("attn_51_transpose_x_0"), val = tensor(false)]; + tensor attn_51_transpose_y_0 = const()[name = tensor("attn_51_transpose_y_0"), val = tensor(true)]; + tensor attn_51_cast_fp16 = matmul(transpose_x = attn_51_transpose_x_0, transpose_y = attn_51_transpose_y_0, x = var_3983_cast_fp16, y = var_3981_cast_fp16)[name = tensor("attn_51_cast_fp16")]; + tensor var_3986 = const()[name = tensor("op_3986"), val = tensor([1, 1280, 1, -1])]; + tensor x_457_cast_fp16 = reshape(shape = var_3986, x = attn_51_cast_fp16)[name = tensor("x_457_cast_fp16")]; + tensor layers_25_self_attn_o_proj_input_shift_to_fp16 = const()[name = tensor("layers_25_self_attn_o_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(259580288)))]; + tensor input_357_cast_fp16 = sub(x = x_457_cast_fp16, y = layers_25_self_attn_o_proj_input_shift_to_fp16)[name = tensor("input_357_cast_fp16")]; + tensor var_3994 = const()[name = tensor("op_3994"), val = tensor([1, 1])]; + tensor var_3996 = const()[name = tensor("op_3996"), val = tensor([1, 1])]; + tensor x_459_pad_type_0 = const()[name = tensor("x_459_pad_type_0"), val = tensor("custom")]; + tensor x_459_pad_0 = const()[name = tensor("x_459_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_25_self_attn_o_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(259582912))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(260402176))), name = tensor("layers_25_self_attn_o_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_25_self_attn_o_proj_module_bias_to_fp16 = const()[name = tensor("layers_25_self_attn_o_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(260402304)))]; + tensor x_459_cast_fp16 = conv(bias = layers_25_self_attn_o_proj_module_bias_to_fp16, dilations = var_3996, groups = var_3905, pad = x_459_pad_0, pad_type = x_459_pad_type_0, strides = var_3994, weight = layers_25_self_attn_o_proj_module_weight_to_fp16_palettized, x = input_357_cast_fp16)[name = tensor("x_459_cast_fp16")]; + tensor layers_25_self_attn_o_proj_output_scale_to_fp16 = const()[name = tensor("layers_25_self_attn_o_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(260404928)))]; + tensor obj_103_cast_fp16 = mul(x = x_459_cast_fp16, y = layers_25_self_attn_o_proj_output_scale_to_fp16)[name = tensor("obj_103_cast_fp16")]; + tensor inputs_103_cast_fp16 = add(x = inputs_101_cast_fp16, y = obj_103_cast_fp16)[name = tensor("inputs_103_cast_fp16")]; + tensor var_4003 = const()[name = tensor("op_4003"), val = tensor([1])]; + tensor channels_mean_103_cast_fp16 = reduce_mean(axes = var_4003, keep_dims = var_3906, x = inputs_103_cast_fp16)[name = tensor("channels_mean_103_cast_fp16")]; + tensor zero_mean_103_cast_fp16 = sub(x = inputs_103_cast_fp16, y = channels_mean_103_cast_fp16)[name = tensor("zero_mean_103_cast_fp16")]; + tensor zero_mean_sq_103_cast_fp16 = mul(x = zero_mean_103_cast_fp16, y = zero_mean_103_cast_fp16)[name = tensor("zero_mean_sq_103_cast_fp16")]; + tensor var_4007 = const()[name = tensor("op_4007"), val = tensor([1])]; + tensor var_4008_cast_fp16 = reduce_mean(axes = var_4007, keep_dims = var_3906, x = zero_mean_sq_103_cast_fp16)[name = tensor("op_4008_cast_fp16")]; + tensor var_4009_to_fp16 = const()[name = tensor("op_4009_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_4010_cast_fp16 = add(x = var_4008_cast_fp16, y = var_4009_to_fp16)[name = tensor("op_4010_cast_fp16")]; + tensor denom_103_epsilon_0_to_fp16 = const()[name = tensor("denom_103_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_103_cast_fp16 = rsqrt(epsilon = denom_103_epsilon_0_to_fp16, x = var_4010_cast_fp16)[name = tensor("denom_103_cast_fp16")]; + tensor out_103_cast_fp16 = mul(x = zero_mean_103_cast_fp16, y = denom_103_cast_fp16)[name = tensor("out_103_cast_fp16")]; + tensor x_461_gamma_0_to_fp16 = const()[name = tensor("x_461_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(260407552)))]; + tensor x_461_beta_0_to_fp16 = const()[name = tensor("x_461_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(260410176)))]; + tensor x_461_epsilon_0_to_fp16 = const()[name = tensor("x_461_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_461_cast_fp16 = batch_norm(beta = x_461_beta_0_to_fp16, epsilon = x_461_epsilon_0_to_fp16, gamma = x_461_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_103_cast_fp16)[name = tensor("x_461_cast_fp16")]; + tensor layers_25_fc1_input_shift_to_fp16 = const()[name = tensor("layers_25_fc1_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(260412800)))]; + tensor input_359_cast_fp16 = sub(x = x_461_cast_fp16, y = layers_25_fc1_input_shift_to_fp16)[name = tensor("input_359_cast_fp16")]; + tensor var_4025 = const()[name = tensor("op_4025"), val = tensor([1, 1])]; + tensor var_4027 = const()[name = tensor("op_4027"), val = tensor([1, 1])]; + tensor x_463_pad_type_0 = const()[name = tensor("x_463_pad_type_0"), val = tensor("custom")]; + tensor x_463_pad_0 = const()[name = tensor("x_463_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_25_fc1_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(260415424))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(263692288))), name = tensor("layers_25_fc1_module_weight_to_fp16_palettized"), shape = tensor([5120, 1280, 1, 1])]; + tensor layers_25_fc1_module_bias_to_fp16 = const()[name = tensor("layers_25_fc1_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(263692416)))]; + tensor x_463_cast_fp16 = conv(bias = layers_25_fc1_module_bias_to_fp16, dilations = var_4027, groups = var_3905, pad = x_463_pad_0, pad_type = x_463_pad_type_0, strides = var_4025, weight = layers_25_fc1_module_weight_to_fp16_palettized, x = input_359_cast_fp16)[name = tensor("x_463_cast_fp16")]; + tensor layers_25_fc1_output_scale_to_fp16 = const()[name = tensor("layers_25_fc1_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(263702720)))]; + tensor input_361_cast_fp16 = mul(x = x_463_cast_fp16, y = layers_25_fc1_output_scale_to_fp16)[name = tensor("input_361_cast_fp16")]; + tensor x_465_mode_0 = const()[name = tensor("x_465_mode_0"), val = tensor("EXACT")]; + tensor x_465_cast_fp16 = gelu(mode = x_465_mode_0, x = input_361_cast_fp16)[name = tensor("x_465_cast_fp16")]; + tensor layers_25_fc2_input_shift_to_fp16 = const()[name = tensor("layers_25_fc2_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(263713024)))]; + tensor input_363_cast_fp16 = sub(x = x_465_cast_fp16, y = layers_25_fc2_input_shift_to_fp16)[name = tensor("input_363_cast_fp16")]; + tensor var_4038 = const()[name = tensor("op_4038"), val = tensor([1, 1])]; + tensor var_4040 = const()[name = tensor("op_4040"), val = tensor([1, 1])]; + tensor x_467_pad_type_0 = const()[name = tensor("x_467_pad_type_0"), val = tensor("custom")]; + tensor x_467_pad_0 = const()[name = tensor("x_467_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_25_fc2_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(263723328))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(267000192))), name = tensor("layers_25_fc2_module_weight_to_fp16_palettized"), shape = tensor([1280, 5120, 1, 1])]; + tensor layers_25_fc2_module_bias_to_fp16 = const()[name = tensor("layers_25_fc2_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(267000320)))]; + tensor x_467_cast_fp16 = conv(bias = layers_25_fc2_module_bias_to_fp16, dilations = var_4040, groups = var_3905, pad = x_467_pad_0, pad_type = x_467_pad_type_0, strides = var_4038, weight = layers_25_fc2_module_weight_to_fp16_palettized, x = input_363_cast_fp16)[name = tensor("x_467_cast_fp16")]; + tensor layers_25_fc2_output_scale_to_fp16 = const()[name = tensor("layers_25_fc2_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(267002944)))]; + tensor hidden_states_55_cast_fp16 = mul(x = x_467_cast_fp16, y = layers_25_fc2_output_scale_to_fp16)[name = tensor("hidden_states_55_cast_fp16")]; + tensor inputs_105_cast_fp16 = add(x = inputs_103_cast_fp16, y = hidden_states_55_cast_fp16)[name = tensor("inputs_105_cast_fp16")]; + tensor var_4052 = const()[name = tensor("op_4052"), val = tensor(3)]; + tensor var_4054 = const()[name = tensor("op_4054"), val = tensor(1)]; + tensor var_4055 = const()[name = tensor("op_4055"), val = tensor(true)]; + tensor var_4065 = const()[name = tensor("op_4065"), val = tensor([1])]; + tensor channels_mean_105_cast_fp16 = reduce_mean(axes = var_4065, keep_dims = var_4055, x = inputs_105_cast_fp16)[name = tensor("channels_mean_105_cast_fp16")]; + tensor zero_mean_105_cast_fp16 = sub(x = inputs_105_cast_fp16, y = channels_mean_105_cast_fp16)[name = tensor("zero_mean_105_cast_fp16")]; + tensor zero_mean_sq_105_cast_fp16 = mul(x = zero_mean_105_cast_fp16, y = zero_mean_105_cast_fp16)[name = tensor("zero_mean_sq_105_cast_fp16")]; + tensor var_4069 = const()[name = tensor("op_4069"), val = tensor([1])]; + tensor var_4070_cast_fp16 = reduce_mean(axes = var_4069, keep_dims = var_4055, x = zero_mean_sq_105_cast_fp16)[name = tensor("op_4070_cast_fp16")]; + tensor var_4071_to_fp16 = const()[name = tensor("op_4071_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_4072_cast_fp16 = add(x = var_4070_cast_fp16, y = var_4071_to_fp16)[name = tensor("op_4072_cast_fp16")]; + tensor denom_105_epsilon_0_to_fp16 = const()[name = tensor("denom_105_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_105_cast_fp16 = rsqrt(epsilon = denom_105_epsilon_0_to_fp16, x = var_4072_cast_fp16)[name = tensor("denom_105_cast_fp16")]; + tensor out_105_cast_fp16 = mul(x = zero_mean_105_cast_fp16, y = denom_105_cast_fp16)[name = tensor("out_105_cast_fp16")]; + tensor obj_105_gamma_0_to_fp16 = const()[name = tensor("obj_105_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(267005568)))]; + tensor obj_105_beta_0_to_fp16 = const()[name = tensor("obj_105_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(267008192)))]; + tensor obj_105_epsilon_0_to_fp16 = const()[name = tensor("obj_105_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_105_cast_fp16 = batch_norm(beta = obj_105_beta_0_to_fp16, epsilon = obj_105_epsilon_0_to_fp16, gamma = obj_105_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_105_cast_fp16)[name = tensor("obj_105_cast_fp16")]; + tensor layers_26_self_attn_q_proj_input_shift_to_fp16 = const()[name = tensor("layers_26_self_attn_q_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(267010816)))]; + tensor input_365_cast_fp16 = sub(x = obj_105_cast_fp16, y = layers_26_self_attn_q_proj_input_shift_to_fp16)[name = tensor("input_365_cast_fp16")]; + tensor var_4091 = const()[name = tensor("op_4091"), val = tensor([1, 1])]; + tensor var_4093 = const()[name = tensor("op_4093"), val = tensor([1, 1])]; + tensor x_469_pad_type_0 = const()[name = tensor("x_469_pad_type_0"), val = tensor("custom")]; + tensor x_469_pad_0 = const()[name = tensor("x_469_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_26_self_attn_q_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(267013440))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(267832704))), name = tensor("layers_26_self_attn_q_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_26_self_attn_q_proj_module_bias_to_fp16 = const()[name = tensor("layers_26_self_attn_q_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(267832832)))]; + tensor x_469_cast_fp16 = conv(bias = layers_26_self_attn_q_proj_module_bias_to_fp16, dilations = var_4093, groups = var_4054, pad = x_469_pad_0, pad_type = x_469_pad_type_0, strides = var_4091, weight = layers_26_self_attn_q_proj_module_weight_to_fp16_palettized, x = input_365_cast_fp16)[name = tensor("x_469_cast_fp16")]; + tensor layers_26_self_attn_q_proj_output_scale_to_fp16 = const()[name = tensor("layers_26_self_attn_q_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(267835456)))]; + tensor query_53_cast_fp16 = mul(x = x_469_cast_fp16, y = layers_26_self_attn_q_proj_output_scale_to_fp16)[name = tensor("query_53_cast_fp16")]; + tensor var_4103 = const()[name = tensor("op_4103"), val = tensor([1, 1])]; + tensor var_4105 = const()[name = tensor("op_4105"), val = tensor([1, 1])]; + tensor x_471_pad_type_0 = const()[name = tensor("x_471_pad_type_0"), val = tensor("custom")]; + tensor x_471_pad_0 = const()[name = tensor("x_471_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_26_self_attn_k_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(267838080))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(268657344))), name = tensor("layers_26_self_attn_k_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_26_self_attn_k_proj_module_bias_to_fp16 = const()[name = tensor("layers_26_self_attn_k_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(268657472)))]; + tensor x_471_cast_fp16 = conv(bias = layers_26_self_attn_k_proj_module_bias_to_fp16, dilations = var_4105, groups = var_4054, pad = x_471_pad_0, pad_type = x_471_pad_type_0, strides = var_4103, weight = layers_26_self_attn_k_proj_module_weight_to_fp16_palettized, x = input_365_cast_fp16)[name = tensor("x_471_cast_fp16")]; + tensor layers_26_self_attn_k_proj_output_scale_to_fp16 = const()[name = tensor("layers_26_self_attn_k_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(268660096)))]; + tensor key_53_cast_fp16 = mul(x = x_471_cast_fp16, y = layers_26_self_attn_k_proj_output_scale_to_fp16)[name = tensor("key_53_cast_fp16")]; + tensor var_4115 = const()[name = tensor("op_4115"), val = tensor([1, 1])]; + tensor var_4117 = const()[name = tensor("op_4117"), val = tensor([1, 1])]; + tensor x_473_pad_type_0 = const()[name = tensor("x_473_pad_type_0"), val = tensor("custom")]; + tensor x_473_pad_0 = const()[name = tensor("x_473_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_26_self_attn_v_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(268662720))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(269481984))), name = tensor("layers_26_self_attn_v_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_26_self_attn_v_proj_module_bias_to_fp16 = const()[name = tensor("layers_26_self_attn_v_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(269482112)))]; + tensor x_473_cast_fp16 = conv(bias = layers_26_self_attn_v_proj_module_bias_to_fp16, dilations = var_4117, groups = var_4054, pad = x_473_pad_0, pad_type = x_473_pad_type_0, strides = var_4115, weight = layers_26_self_attn_v_proj_module_weight_to_fp16_palettized, x = input_365_cast_fp16)[name = tensor("x_473_cast_fp16")]; + tensor layers_26_self_attn_v_proj_output_scale_to_fp16 = const()[name = tensor("layers_26_self_attn_v_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(269484736)))]; + tensor value_53_cast_fp16 = mul(x = x_473_cast_fp16, y = layers_26_self_attn_v_proj_output_scale_to_fp16)[name = tensor("value_53_cast_fp16")]; + tensor var_4122 = const()[name = tensor("op_4122"), val = tensor([1, 20, 64, -1])]; + tensor var_4123_cast_fp16 = reshape(shape = var_4122, x = query_53_cast_fp16)[name = tensor("op_4123_cast_fp16")]; + tensor var_4124_to_fp16 = const()[name = tensor("op_4124_to_fp16"), val = tensor(0x1p-3)]; + tensor var_4125_cast_fp16 = mul(x = var_4123_cast_fp16, y = var_4124_to_fp16)[name = tensor("op_4125_cast_fp16")]; + tensor var_4126 = const()[name = tensor("op_4126"), val = tensor([1, 20, 64, -1])]; + tensor var_4127_cast_fp16 = reshape(shape = var_4126, x = key_53_cast_fp16)[name = tensor("op_4127_cast_fp16")]; + tensor mh_w_53_transpose_x_0 = const()[name = tensor("mh_w_53_transpose_x_0"), val = tensor(true)]; + tensor mh_w_53_transpose_y_0 = const()[name = tensor("mh_w_53_transpose_y_0"), val = tensor(false)]; + tensor mh_w_53_cast_fp16 = matmul(transpose_x = mh_w_53_transpose_x_0, transpose_y = mh_w_53_transpose_y_0, x = var_4125_cast_fp16, y = var_4127_cast_fp16)[name = tensor("mh_w_53_cast_fp16")]; + tensor var_4130_cast_fp16 = softmax(axis = var_4052, x = mh_w_53_cast_fp16)[name = tensor("op_4130_cast_fp16")]; + tensor var_4131 = const()[name = tensor("op_4131"), val = tensor([1, 20, 64, -1])]; + tensor var_4132_cast_fp16 = reshape(shape = var_4131, x = value_53_cast_fp16)[name = tensor("op_4132_cast_fp16")]; + tensor attn_53_transpose_x_0 = const()[name = tensor("attn_53_transpose_x_0"), val = tensor(false)]; + tensor attn_53_transpose_y_0 = const()[name = tensor("attn_53_transpose_y_0"), val = tensor(true)]; + tensor attn_53_cast_fp16 = matmul(transpose_x = attn_53_transpose_x_0, transpose_y = attn_53_transpose_y_0, x = var_4132_cast_fp16, y = var_4130_cast_fp16)[name = tensor("attn_53_cast_fp16")]; + tensor var_4135 = const()[name = tensor("op_4135"), val = tensor([1, 1280, 1, -1])]; + tensor x_475_cast_fp16 = reshape(shape = var_4135, x = attn_53_cast_fp16)[name = tensor("x_475_cast_fp16")]; + tensor layers_26_self_attn_o_proj_input_shift_to_fp16 = const()[name = tensor("layers_26_self_attn_o_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(269487360)))]; + tensor input_371_cast_fp16 = sub(x = x_475_cast_fp16, y = layers_26_self_attn_o_proj_input_shift_to_fp16)[name = tensor("input_371_cast_fp16")]; + tensor var_4143 = const()[name = tensor("op_4143"), val = tensor([1, 1])]; + tensor var_4145 = const()[name = tensor("op_4145"), val = tensor([1, 1])]; + tensor x_477_pad_type_0 = const()[name = tensor("x_477_pad_type_0"), val = tensor("custom")]; + tensor x_477_pad_0 = const()[name = tensor("x_477_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_26_self_attn_o_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(269489984))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(270309248))), name = tensor("layers_26_self_attn_o_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_26_self_attn_o_proj_module_bias_to_fp16 = const()[name = tensor("layers_26_self_attn_o_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(270309376)))]; + tensor x_477_cast_fp16 = conv(bias = layers_26_self_attn_o_proj_module_bias_to_fp16, dilations = var_4145, groups = var_4054, pad = x_477_pad_0, pad_type = x_477_pad_type_0, strides = var_4143, weight = layers_26_self_attn_o_proj_module_weight_to_fp16_palettized, x = input_371_cast_fp16)[name = tensor("x_477_cast_fp16")]; + tensor layers_26_self_attn_o_proj_output_scale_to_fp16 = const()[name = tensor("layers_26_self_attn_o_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(270312000)))]; + tensor obj_107_cast_fp16 = mul(x = x_477_cast_fp16, y = layers_26_self_attn_o_proj_output_scale_to_fp16)[name = tensor("obj_107_cast_fp16")]; + tensor inputs_107_cast_fp16 = add(x = inputs_105_cast_fp16, y = obj_107_cast_fp16)[name = tensor("inputs_107_cast_fp16")]; + tensor var_4152 = const()[name = tensor("op_4152"), val = tensor([1])]; + tensor channels_mean_107_cast_fp16 = reduce_mean(axes = var_4152, keep_dims = var_4055, x = inputs_107_cast_fp16)[name = tensor("channels_mean_107_cast_fp16")]; + tensor zero_mean_107_cast_fp16 = sub(x = inputs_107_cast_fp16, y = channels_mean_107_cast_fp16)[name = tensor("zero_mean_107_cast_fp16")]; + tensor zero_mean_sq_107_cast_fp16 = mul(x = zero_mean_107_cast_fp16, y = zero_mean_107_cast_fp16)[name = tensor("zero_mean_sq_107_cast_fp16")]; + tensor var_4156 = const()[name = tensor("op_4156"), val = tensor([1])]; + tensor var_4157_cast_fp16 = reduce_mean(axes = var_4156, keep_dims = var_4055, x = zero_mean_sq_107_cast_fp16)[name = tensor("op_4157_cast_fp16")]; + tensor var_4158_to_fp16 = const()[name = tensor("op_4158_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_4159_cast_fp16 = add(x = var_4157_cast_fp16, y = var_4158_to_fp16)[name = tensor("op_4159_cast_fp16")]; + tensor denom_107_epsilon_0_to_fp16 = const()[name = tensor("denom_107_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_107_cast_fp16 = rsqrt(epsilon = denom_107_epsilon_0_to_fp16, x = var_4159_cast_fp16)[name = tensor("denom_107_cast_fp16")]; + tensor out_107_cast_fp16 = mul(x = zero_mean_107_cast_fp16, y = denom_107_cast_fp16)[name = tensor("out_107_cast_fp16")]; + tensor x_479_gamma_0_to_fp16 = const()[name = tensor("x_479_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(270314624)))]; + tensor x_479_beta_0_to_fp16 = const()[name = tensor("x_479_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(270317248)))]; + tensor x_479_epsilon_0_to_fp16 = const()[name = tensor("x_479_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_479_cast_fp16 = batch_norm(beta = x_479_beta_0_to_fp16, epsilon = x_479_epsilon_0_to_fp16, gamma = x_479_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_107_cast_fp16)[name = tensor("x_479_cast_fp16")]; + tensor layers_26_fc1_input_shift_to_fp16 = const()[name = tensor("layers_26_fc1_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(270319872)))]; + tensor input_373_cast_fp16 = sub(x = x_479_cast_fp16, y = layers_26_fc1_input_shift_to_fp16)[name = tensor("input_373_cast_fp16")]; + tensor var_4174 = const()[name = tensor("op_4174"), val = tensor([1, 1])]; + tensor var_4176 = const()[name = tensor("op_4176"), val = tensor([1, 1])]; + tensor x_481_pad_type_0 = const()[name = tensor("x_481_pad_type_0"), val = tensor("custom")]; + tensor x_481_pad_0 = const()[name = tensor("x_481_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_26_fc1_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(270322496))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(273599360))), name = tensor("layers_26_fc1_module_weight_to_fp16_palettized"), shape = tensor([5120, 1280, 1, 1])]; + tensor layers_26_fc1_module_bias_to_fp16 = const()[name = tensor("layers_26_fc1_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(273599488)))]; + tensor x_481_cast_fp16 = conv(bias = layers_26_fc1_module_bias_to_fp16, dilations = var_4176, groups = var_4054, pad = x_481_pad_0, pad_type = x_481_pad_type_0, strides = var_4174, weight = layers_26_fc1_module_weight_to_fp16_palettized, x = input_373_cast_fp16)[name = tensor("x_481_cast_fp16")]; + tensor layers_26_fc1_output_scale_to_fp16 = const()[name = tensor("layers_26_fc1_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(273609792)))]; + tensor input_375_cast_fp16 = mul(x = x_481_cast_fp16, y = layers_26_fc1_output_scale_to_fp16)[name = tensor("input_375_cast_fp16")]; + tensor x_483_mode_0 = const()[name = tensor("x_483_mode_0"), val = tensor("EXACT")]; + tensor x_483_cast_fp16 = gelu(mode = x_483_mode_0, x = input_375_cast_fp16)[name = tensor("x_483_cast_fp16")]; + tensor layers_26_fc2_input_shift_to_fp16 = const()[name = tensor("layers_26_fc2_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(273620096)))]; + tensor input_377_cast_fp16 = sub(x = x_483_cast_fp16, y = layers_26_fc2_input_shift_to_fp16)[name = tensor("input_377_cast_fp16")]; + tensor var_4187 = const()[name = tensor("op_4187"), val = tensor([1, 1])]; + tensor var_4189 = const()[name = tensor("op_4189"), val = tensor([1, 1])]; + tensor x_485_pad_type_0 = const()[name = tensor("x_485_pad_type_0"), val = tensor("custom")]; + tensor x_485_pad_0 = const()[name = tensor("x_485_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_26_fc2_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(273630400))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(276907264))), name = tensor("layers_26_fc2_module_weight_to_fp16_palettized"), shape = tensor([1280, 5120, 1, 1])]; + tensor layers_26_fc2_module_bias_to_fp16 = const()[name = tensor("layers_26_fc2_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(276907392)))]; + tensor x_485_cast_fp16 = conv(bias = layers_26_fc2_module_bias_to_fp16, dilations = var_4189, groups = var_4054, pad = x_485_pad_0, pad_type = x_485_pad_type_0, strides = var_4187, weight = layers_26_fc2_module_weight_to_fp16_palettized, x = input_377_cast_fp16)[name = tensor("x_485_cast_fp16")]; + tensor layers_26_fc2_output_scale_to_fp16 = const()[name = tensor("layers_26_fc2_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(276910016)))]; + tensor hidden_states_57_cast_fp16 = mul(x = x_485_cast_fp16, y = layers_26_fc2_output_scale_to_fp16)[name = tensor("hidden_states_57_cast_fp16")]; + tensor inputs_109_cast_fp16 = add(x = inputs_107_cast_fp16, y = hidden_states_57_cast_fp16)[name = tensor("inputs_109_cast_fp16")]; + tensor var_4201 = const()[name = tensor("op_4201"), val = tensor(3)]; + tensor var_4203 = const()[name = tensor("op_4203"), val = tensor(1)]; + tensor var_4204 = const()[name = tensor("op_4204"), val = tensor(true)]; + tensor var_4214 = const()[name = tensor("op_4214"), val = tensor([1])]; + tensor channels_mean_109_cast_fp16 = reduce_mean(axes = var_4214, keep_dims = var_4204, x = inputs_109_cast_fp16)[name = tensor("channels_mean_109_cast_fp16")]; + tensor zero_mean_109_cast_fp16 = sub(x = inputs_109_cast_fp16, y = channels_mean_109_cast_fp16)[name = tensor("zero_mean_109_cast_fp16")]; + tensor zero_mean_sq_109_cast_fp16 = mul(x = zero_mean_109_cast_fp16, y = zero_mean_109_cast_fp16)[name = tensor("zero_mean_sq_109_cast_fp16")]; + tensor var_4218 = const()[name = tensor("op_4218"), val = tensor([1])]; + tensor var_4219_cast_fp16 = reduce_mean(axes = var_4218, keep_dims = var_4204, x = zero_mean_sq_109_cast_fp16)[name = tensor("op_4219_cast_fp16")]; + tensor var_4220_to_fp16 = const()[name = tensor("op_4220_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_4221_cast_fp16 = add(x = var_4219_cast_fp16, y = var_4220_to_fp16)[name = tensor("op_4221_cast_fp16")]; + tensor denom_109_epsilon_0_to_fp16 = const()[name = tensor("denom_109_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_109_cast_fp16 = rsqrt(epsilon = denom_109_epsilon_0_to_fp16, x = var_4221_cast_fp16)[name = tensor("denom_109_cast_fp16")]; + tensor out_109_cast_fp16 = mul(x = zero_mean_109_cast_fp16, y = denom_109_cast_fp16)[name = tensor("out_109_cast_fp16")]; + tensor obj_109_gamma_0_to_fp16 = const()[name = tensor("obj_109_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(276912640)))]; + tensor obj_109_beta_0_to_fp16 = const()[name = tensor("obj_109_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(276915264)))]; + tensor obj_109_epsilon_0_to_fp16 = const()[name = tensor("obj_109_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_109_cast_fp16 = batch_norm(beta = obj_109_beta_0_to_fp16, epsilon = obj_109_epsilon_0_to_fp16, gamma = obj_109_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_109_cast_fp16)[name = tensor("obj_109_cast_fp16")]; + tensor layers_27_self_attn_q_proj_input_shift_to_fp16 = const()[name = tensor("layers_27_self_attn_q_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(276917888)))]; + tensor input_379_cast_fp16 = sub(x = obj_109_cast_fp16, y = layers_27_self_attn_q_proj_input_shift_to_fp16)[name = tensor("input_379_cast_fp16")]; + tensor var_4240 = const()[name = tensor("op_4240"), val = tensor([1, 1])]; + tensor var_4242 = const()[name = tensor("op_4242"), val = tensor([1, 1])]; + tensor x_487_pad_type_0 = const()[name = tensor("x_487_pad_type_0"), val = tensor("custom")]; + tensor x_487_pad_0 = const()[name = tensor("x_487_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_27_self_attn_q_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(276920512))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(277739776))), name = tensor("layers_27_self_attn_q_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_27_self_attn_q_proj_module_bias_to_fp16 = const()[name = tensor("layers_27_self_attn_q_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(277739904)))]; + tensor x_487_cast_fp16 = conv(bias = layers_27_self_attn_q_proj_module_bias_to_fp16, dilations = var_4242, groups = var_4203, pad = x_487_pad_0, pad_type = x_487_pad_type_0, strides = var_4240, weight = layers_27_self_attn_q_proj_module_weight_to_fp16_palettized, x = input_379_cast_fp16)[name = tensor("x_487_cast_fp16")]; + tensor layers_27_self_attn_q_proj_output_scale_to_fp16 = const()[name = tensor("layers_27_self_attn_q_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(277742528)))]; + tensor query_55_cast_fp16 = mul(x = x_487_cast_fp16, y = layers_27_self_attn_q_proj_output_scale_to_fp16)[name = tensor("query_55_cast_fp16")]; + tensor var_4252 = const()[name = tensor("op_4252"), val = tensor([1, 1])]; + tensor var_4254 = const()[name = tensor("op_4254"), val = tensor([1, 1])]; + tensor x_489_pad_type_0 = const()[name = tensor("x_489_pad_type_0"), val = tensor("custom")]; + tensor x_489_pad_0 = const()[name = tensor("x_489_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_27_self_attn_k_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(277745152))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(278564416))), name = tensor("layers_27_self_attn_k_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_27_self_attn_k_proj_module_bias_to_fp16 = const()[name = tensor("layers_27_self_attn_k_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(278564544)))]; + tensor x_489_cast_fp16 = conv(bias = layers_27_self_attn_k_proj_module_bias_to_fp16, dilations = var_4254, groups = var_4203, pad = x_489_pad_0, pad_type = x_489_pad_type_0, strides = var_4252, weight = layers_27_self_attn_k_proj_module_weight_to_fp16_palettized, x = input_379_cast_fp16)[name = tensor("x_489_cast_fp16")]; + tensor layers_27_self_attn_k_proj_output_scale_to_fp16 = const()[name = tensor("layers_27_self_attn_k_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(278567168)))]; + tensor key_55_cast_fp16 = mul(x = x_489_cast_fp16, y = layers_27_self_attn_k_proj_output_scale_to_fp16)[name = tensor("key_55_cast_fp16")]; + tensor var_4264 = const()[name = tensor("op_4264"), val = tensor([1, 1])]; + tensor var_4266 = const()[name = tensor("op_4266"), val = tensor([1, 1])]; + tensor x_491_pad_type_0 = const()[name = tensor("x_491_pad_type_0"), val = tensor("custom")]; + tensor x_491_pad_0 = const()[name = tensor("x_491_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_27_self_attn_v_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(278569792))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(279389056))), name = tensor("layers_27_self_attn_v_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_27_self_attn_v_proj_module_bias_to_fp16 = const()[name = tensor("layers_27_self_attn_v_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(279389184)))]; + tensor x_491_cast_fp16 = conv(bias = layers_27_self_attn_v_proj_module_bias_to_fp16, dilations = var_4266, groups = var_4203, pad = x_491_pad_0, pad_type = x_491_pad_type_0, strides = var_4264, weight = layers_27_self_attn_v_proj_module_weight_to_fp16_palettized, x = input_379_cast_fp16)[name = tensor("x_491_cast_fp16")]; + tensor layers_27_self_attn_v_proj_output_scale_to_fp16 = const()[name = tensor("layers_27_self_attn_v_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(279391808)))]; + tensor value_55_cast_fp16 = mul(x = x_491_cast_fp16, y = layers_27_self_attn_v_proj_output_scale_to_fp16)[name = tensor("value_55_cast_fp16")]; + tensor var_4271 = const()[name = tensor("op_4271"), val = tensor([1, 20, 64, -1])]; + tensor var_4272_cast_fp16 = reshape(shape = var_4271, x = query_55_cast_fp16)[name = tensor("op_4272_cast_fp16")]; + tensor var_4273_to_fp16 = const()[name = tensor("op_4273_to_fp16"), val = tensor(0x1p-3)]; + tensor var_4274_cast_fp16 = mul(x = var_4272_cast_fp16, y = var_4273_to_fp16)[name = tensor("op_4274_cast_fp16")]; + tensor var_4275 = const()[name = tensor("op_4275"), val = tensor([1, 20, 64, -1])]; + tensor var_4276_cast_fp16 = reshape(shape = var_4275, x = key_55_cast_fp16)[name = tensor("op_4276_cast_fp16")]; + tensor mh_w_55_transpose_x_0 = const()[name = tensor("mh_w_55_transpose_x_0"), val = tensor(true)]; + tensor mh_w_55_transpose_y_0 = const()[name = tensor("mh_w_55_transpose_y_0"), val = tensor(false)]; + tensor mh_w_55_cast_fp16 = matmul(transpose_x = mh_w_55_transpose_x_0, transpose_y = mh_w_55_transpose_y_0, x = var_4274_cast_fp16, y = var_4276_cast_fp16)[name = tensor("mh_w_55_cast_fp16")]; + tensor var_4279_cast_fp16 = softmax(axis = var_4201, x = mh_w_55_cast_fp16)[name = tensor("op_4279_cast_fp16")]; + tensor var_4280 = const()[name = tensor("op_4280"), val = tensor([1, 20, 64, -1])]; + tensor var_4281_cast_fp16 = reshape(shape = var_4280, x = value_55_cast_fp16)[name = tensor("op_4281_cast_fp16")]; + tensor attn_55_transpose_x_0 = const()[name = tensor("attn_55_transpose_x_0"), val = tensor(false)]; + tensor attn_55_transpose_y_0 = const()[name = tensor("attn_55_transpose_y_0"), val = tensor(true)]; + tensor attn_55_cast_fp16 = matmul(transpose_x = attn_55_transpose_x_0, transpose_y = attn_55_transpose_y_0, x = var_4281_cast_fp16, y = var_4279_cast_fp16)[name = tensor("attn_55_cast_fp16")]; + tensor var_4284 = const()[name = tensor("op_4284"), val = tensor([1, 1280, 1, -1])]; + tensor x_493_cast_fp16 = reshape(shape = var_4284, x = attn_55_cast_fp16)[name = tensor("x_493_cast_fp16")]; + tensor layers_27_self_attn_o_proj_input_shift_to_fp16 = const()[name = tensor("layers_27_self_attn_o_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(279394432)))]; + tensor input_385_cast_fp16 = sub(x = x_493_cast_fp16, y = layers_27_self_attn_o_proj_input_shift_to_fp16)[name = tensor("input_385_cast_fp16")]; + tensor var_4292 = const()[name = tensor("op_4292"), val = tensor([1, 1])]; + tensor var_4294 = const()[name = tensor("op_4294"), val = tensor([1, 1])]; + tensor x_495_pad_type_0 = const()[name = tensor("x_495_pad_type_0"), val = tensor("custom")]; + tensor x_495_pad_0 = const()[name = tensor("x_495_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_27_self_attn_o_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(279397056))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(280216320))), name = tensor("layers_27_self_attn_o_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_27_self_attn_o_proj_module_bias_to_fp16 = const()[name = tensor("layers_27_self_attn_o_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(280216448)))]; + tensor x_495_cast_fp16 = conv(bias = layers_27_self_attn_o_proj_module_bias_to_fp16, dilations = var_4294, groups = var_4203, pad = x_495_pad_0, pad_type = x_495_pad_type_0, strides = var_4292, weight = layers_27_self_attn_o_proj_module_weight_to_fp16_palettized, x = input_385_cast_fp16)[name = tensor("x_495_cast_fp16")]; + tensor layers_27_self_attn_o_proj_output_scale_to_fp16 = const()[name = tensor("layers_27_self_attn_o_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(280219072)))]; + tensor obj_111_cast_fp16 = mul(x = x_495_cast_fp16, y = layers_27_self_attn_o_proj_output_scale_to_fp16)[name = tensor("obj_111_cast_fp16")]; + tensor inputs_111_cast_fp16 = add(x = inputs_109_cast_fp16, y = obj_111_cast_fp16)[name = tensor("inputs_111_cast_fp16")]; + tensor var_4301 = const()[name = tensor("op_4301"), val = tensor([1])]; + tensor channels_mean_111_cast_fp16 = reduce_mean(axes = var_4301, keep_dims = var_4204, x = inputs_111_cast_fp16)[name = tensor("channels_mean_111_cast_fp16")]; + tensor zero_mean_111_cast_fp16 = sub(x = inputs_111_cast_fp16, y = channels_mean_111_cast_fp16)[name = tensor("zero_mean_111_cast_fp16")]; + tensor zero_mean_sq_111_cast_fp16 = mul(x = zero_mean_111_cast_fp16, y = zero_mean_111_cast_fp16)[name = tensor("zero_mean_sq_111_cast_fp16")]; + tensor var_4305 = const()[name = tensor("op_4305"), val = tensor([1])]; + tensor var_4306_cast_fp16 = reduce_mean(axes = var_4305, keep_dims = var_4204, x = zero_mean_sq_111_cast_fp16)[name = tensor("op_4306_cast_fp16")]; + tensor var_4307_to_fp16 = const()[name = tensor("op_4307_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_4308_cast_fp16 = add(x = var_4306_cast_fp16, y = var_4307_to_fp16)[name = tensor("op_4308_cast_fp16")]; + tensor denom_111_epsilon_0_to_fp16 = const()[name = tensor("denom_111_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_111_cast_fp16 = rsqrt(epsilon = denom_111_epsilon_0_to_fp16, x = var_4308_cast_fp16)[name = tensor("denom_111_cast_fp16")]; + tensor out_111_cast_fp16 = mul(x = zero_mean_111_cast_fp16, y = denom_111_cast_fp16)[name = tensor("out_111_cast_fp16")]; + tensor x_497_gamma_0_to_fp16 = const()[name = tensor("x_497_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(280221696)))]; + tensor x_497_beta_0_to_fp16 = const()[name = tensor("x_497_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(280224320)))]; + tensor x_497_epsilon_0_to_fp16 = const()[name = tensor("x_497_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_497_cast_fp16 = batch_norm(beta = x_497_beta_0_to_fp16, epsilon = x_497_epsilon_0_to_fp16, gamma = x_497_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_111_cast_fp16)[name = tensor("x_497_cast_fp16")]; + tensor layers_27_fc1_input_shift_to_fp16 = const()[name = tensor("layers_27_fc1_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(280226944)))]; + tensor input_387_cast_fp16 = sub(x = x_497_cast_fp16, y = layers_27_fc1_input_shift_to_fp16)[name = tensor("input_387_cast_fp16")]; + tensor var_4323 = const()[name = tensor("op_4323"), val = tensor([1, 1])]; + tensor var_4325 = const()[name = tensor("op_4325"), val = tensor([1, 1])]; + tensor x_499_pad_type_0 = const()[name = tensor("x_499_pad_type_0"), val = tensor("custom")]; + tensor x_499_pad_0 = const()[name = tensor("x_499_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_27_fc1_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(280229568))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(283506432))), name = tensor("layers_27_fc1_module_weight_to_fp16_palettized"), shape = tensor([5120, 1280, 1, 1])]; + tensor layers_27_fc1_module_bias_to_fp16 = const()[name = tensor("layers_27_fc1_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(283506560)))]; + tensor x_499_cast_fp16 = conv(bias = layers_27_fc1_module_bias_to_fp16, dilations = var_4325, groups = var_4203, pad = x_499_pad_0, pad_type = x_499_pad_type_0, strides = var_4323, weight = layers_27_fc1_module_weight_to_fp16_palettized, x = input_387_cast_fp16)[name = tensor("x_499_cast_fp16")]; + tensor layers_27_fc1_output_scale_to_fp16 = const()[name = tensor("layers_27_fc1_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(283516864)))]; + tensor input_389_cast_fp16 = mul(x = x_499_cast_fp16, y = layers_27_fc1_output_scale_to_fp16)[name = tensor("input_389_cast_fp16")]; + tensor x_501_mode_0 = const()[name = tensor("x_501_mode_0"), val = tensor("EXACT")]; + tensor x_501_cast_fp16 = gelu(mode = x_501_mode_0, x = input_389_cast_fp16)[name = tensor("x_501_cast_fp16")]; + tensor layers_27_fc2_input_shift_to_fp16 = const()[name = tensor("layers_27_fc2_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(283527168)))]; + tensor input_391_cast_fp16 = sub(x = x_501_cast_fp16, y = layers_27_fc2_input_shift_to_fp16)[name = tensor("input_391_cast_fp16")]; + tensor var_4336 = const()[name = tensor("op_4336"), val = tensor([1, 1])]; + tensor var_4338 = const()[name = tensor("op_4338"), val = tensor([1, 1])]; + tensor x_503_pad_type_0 = const()[name = tensor("x_503_pad_type_0"), val = tensor("custom")]; + tensor x_503_pad_0 = const()[name = tensor("x_503_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_27_fc2_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(283537472))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(286814336))), name = tensor("layers_27_fc2_module_weight_to_fp16_palettized"), shape = tensor([1280, 5120, 1, 1])]; + tensor layers_27_fc2_module_bias_to_fp16 = const()[name = tensor("layers_27_fc2_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(286814464)))]; + tensor x_503_cast_fp16 = conv(bias = layers_27_fc2_module_bias_to_fp16, dilations = var_4338, groups = var_4203, pad = x_503_pad_0, pad_type = x_503_pad_type_0, strides = var_4336, weight = layers_27_fc2_module_weight_to_fp16_palettized, x = input_391_cast_fp16)[name = tensor("x_503_cast_fp16")]; + tensor layers_27_fc2_output_scale_to_fp16 = const()[name = tensor("layers_27_fc2_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(286817088)))]; + tensor hidden_states_59_cast_fp16 = mul(x = x_503_cast_fp16, y = layers_27_fc2_output_scale_to_fp16)[name = tensor("hidden_states_59_cast_fp16")]; + tensor inputs_113_cast_fp16 = add(x = inputs_111_cast_fp16, y = hidden_states_59_cast_fp16)[name = tensor("inputs_113_cast_fp16")]; + tensor var_4350 = const()[name = tensor("op_4350"), val = tensor(3)]; + tensor var_4352 = const()[name = tensor("op_4352"), val = tensor(1)]; + tensor var_4353 = const()[name = tensor("op_4353"), val = tensor(true)]; + tensor var_4363 = const()[name = tensor("op_4363"), val = tensor([1])]; + tensor channels_mean_113_cast_fp16 = reduce_mean(axes = var_4363, keep_dims = var_4353, x = inputs_113_cast_fp16)[name = tensor("channels_mean_113_cast_fp16")]; + tensor zero_mean_113_cast_fp16 = sub(x = inputs_113_cast_fp16, y = channels_mean_113_cast_fp16)[name = tensor("zero_mean_113_cast_fp16")]; + tensor zero_mean_sq_113_cast_fp16 = mul(x = zero_mean_113_cast_fp16, y = zero_mean_113_cast_fp16)[name = tensor("zero_mean_sq_113_cast_fp16")]; + tensor var_4367 = const()[name = tensor("op_4367"), val = tensor([1])]; + tensor var_4368_cast_fp16 = reduce_mean(axes = var_4367, keep_dims = var_4353, x = zero_mean_sq_113_cast_fp16)[name = tensor("op_4368_cast_fp16")]; + tensor var_4369_to_fp16 = const()[name = tensor("op_4369_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_4370_cast_fp16 = add(x = var_4368_cast_fp16, y = var_4369_to_fp16)[name = tensor("op_4370_cast_fp16")]; + tensor denom_113_epsilon_0_to_fp16 = const()[name = tensor("denom_113_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_113_cast_fp16 = rsqrt(epsilon = denom_113_epsilon_0_to_fp16, x = var_4370_cast_fp16)[name = tensor("denom_113_cast_fp16")]; + tensor out_113_cast_fp16 = mul(x = zero_mean_113_cast_fp16, y = denom_113_cast_fp16)[name = tensor("out_113_cast_fp16")]; + tensor obj_113_gamma_0_to_fp16 = const()[name = tensor("obj_113_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(286819712)))]; + tensor obj_113_beta_0_to_fp16 = const()[name = tensor("obj_113_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(286822336)))]; + tensor obj_113_epsilon_0_to_fp16 = const()[name = tensor("obj_113_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_113_cast_fp16 = batch_norm(beta = obj_113_beta_0_to_fp16, epsilon = obj_113_epsilon_0_to_fp16, gamma = obj_113_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_113_cast_fp16)[name = tensor("obj_113_cast_fp16")]; + tensor layers_28_self_attn_q_proj_input_shift_to_fp16 = const()[name = tensor("layers_28_self_attn_q_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(286824960)))]; + tensor input_393_cast_fp16 = sub(x = obj_113_cast_fp16, y = layers_28_self_attn_q_proj_input_shift_to_fp16)[name = tensor("input_393_cast_fp16")]; + tensor var_4389 = const()[name = tensor("op_4389"), val = tensor([1, 1])]; + tensor var_4391 = const()[name = tensor("op_4391"), val = tensor([1, 1])]; + tensor x_505_pad_type_0 = const()[name = tensor("x_505_pad_type_0"), val = tensor("custom")]; + tensor x_505_pad_0 = const()[name = tensor("x_505_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_28_self_attn_q_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(286827584))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(287646848))), name = tensor("layers_28_self_attn_q_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_28_self_attn_q_proj_module_bias_to_fp16 = const()[name = tensor("layers_28_self_attn_q_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(287646976)))]; + tensor x_505_cast_fp16 = conv(bias = layers_28_self_attn_q_proj_module_bias_to_fp16, dilations = var_4391, groups = var_4352, pad = x_505_pad_0, pad_type = x_505_pad_type_0, strides = var_4389, weight = layers_28_self_attn_q_proj_module_weight_to_fp16_palettized, x = input_393_cast_fp16)[name = tensor("x_505_cast_fp16")]; + tensor layers_28_self_attn_q_proj_output_scale_to_fp16 = const()[name = tensor("layers_28_self_attn_q_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(287649600)))]; + tensor query_57_cast_fp16 = mul(x = x_505_cast_fp16, y = layers_28_self_attn_q_proj_output_scale_to_fp16)[name = tensor("query_57_cast_fp16")]; + tensor var_4401 = const()[name = tensor("op_4401"), val = tensor([1, 1])]; + tensor var_4403 = const()[name = tensor("op_4403"), val = tensor([1, 1])]; + tensor x_507_pad_type_0 = const()[name = tensor("x_507_pad_type_0"), val = tensor("custom")]; + tensor x_507_pad_0 = const()[name = tensor("x_507_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_28_self_attn_k_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(287652224))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(288471488))), name = tensor("layers_28_self_attn_k_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_28_self_attn_k_proj_module_bias_to_fp16 = const()[name = tensor("layers_28_self_attn_k_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(288471616)))]; + tensor x_507_cast_fp16 = conv(bias = layers_28_self_attn_k_proj_module_bias_to_fp16, dilations = var_4403, groups = var_4352, pad = x_507_pad_0, pad_type = x_507_pad_type_0, strides = var_4401, weight = layers_28_self_attn_k_proj_module_weight_to_fp16_palettized, x = input_393_cast_fp16)[name = tensor("x_507_cast_fp16")]; + tensor layers_28_self_attn_k_proj_output_scale_to_fp16 = const()[name = tensor("layers_28_self_attn_k_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(288474240)))]; + tensor key_57_cast_fp16 = mul(x = x_507_cast_fp16, y = layers_28_self_attn_k_proj_output_scale_to_fp16)[name = tensor("key_57_cast_fp16")]; + tensor var_4413 = const()[name = tensor("op_4413"), val = tensor([1, 1])]; + tensor var_4415 = const()[name = tensor("op_4415"), val = tensor([1, 1])]; + tensor x_509_pad_type_0 = const()[name = tensor("x_509_pad_type_0"), val = tensor("custom")]; + tensor x_509_pad_0 = const()[name = tensor("x_509_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_28_self_attn_v_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(288476864))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(289296128))), name = tensor("layers_28_self_attn_v_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_28_self_attn_v_proj_module_bias_to_fp16 = const()[name = tensor("layers_28_self_attn_v_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(289296256)))]; + tensor x_509_cast_fp16 = conv(bias = layers_28_self_attn_v_proj_module_bias_to_fp16, dilations = var_4415, groups = var_4352, pad = x_509_pad_0, pad_type = x_509_pad_type_0, strides = var_4413, weight = layers_28_self_attn_v_proj_module_weight_to_fp16_palettized, x = input_393_cast_fp16)[name = tensor("x_509_cast_fp16")]; + tensor layers_28_self_attn_v_proj_output_scale_to_fp16 = const()[name = tensor("layers_28_self_attn_v_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(289298880)))]; + tensor value_57_cast_fp16 = mul(x = x_509_cast_fp16, y = layers_28_self_attn_v_proj_output_scale_to_fp16)[name = tensor("value_57_cast_fp16")]; + tensor var_4420 = const()[name = tensor("op_4420"), val = tensor([1, 20, 64, -1])]; + tensor var_4421_cast_fp16 = reshape(shape = var_4420, x = query_57_cast_fp16)[name = tensor("op_4421_cast_fp16")]; + tensor var_4422_to_fp16 = const()[name = tensor("op_4422_to_fp16"), val = tensor(0x1p-3)]; + tensor var_4423_cast_fp16 = mul(x = var_4421_cast_fp16, y = var_4422_to_fp16)[name = tensor("op_4423_cast_fp16")]; + tensor var_4424 = const()[name = tensor("op_4424"), val = tensor([1, 20, 64, -1])]; + tensor var_4425_cast_fp16 = reshape(shape = var_4424, x = key_57_cast_fp16)[name = tensor("op_4425_cast_fp16")]; + tensor mh_w_57_transpose_x_0 = const()[name = tensor("mh_w_57_transpose_x_0"), val = tensor(true)]; + tensor mh_w_57_transpose_y_0 = const()[name = tensor("mh_w_57_transpose_y_0"), val = tensor(false)]; + tensor mh_w_57_cast_fp16 = matmul(transpose_x = mh_w_57_transpose_x_0, transpose_y = mh_w_57_transpose_y_0, x = var_4423_cast_fp16, y = var_4425_cast_fp16)[name = tensor("mh_w_57_cast_fp16")]; + tensor var_4428_cast_fp16 = softmax(axis = var_4350, x = mh_w_57_cast_fp16)[name = tensor("op_4428_cast_fp16")]; + tensor var_4429 = const()[name = tensor("op_4429"), val = tensor([1, 20, 64, -1])]; + tensor var_4430_cast_fp16 = reshape(shape = var_4429, x = value_57_cast_fp16)[name = tensor("op_4430_cast_fp16")]; + tensor attn_57_transpose_x_0 = const()[name = tensor("attn_57_transpose_x_0"), val = tensor(false)]; + tensor attn_57_transpose_y_0 = const()[name = tensor("attn_57_transpose_y_0"), val = tensor(true)]; + tensor attn_57_cast_fp16 = matmul(transpose_x = attn_57_transpose_x_0, transpose_y = attn_57_transpose_y_0, x = var_4430_cast_fp16, y = var_4428_cast_fp16)[name = tensor("attn_57_cast_fp16")]; + tensor var_4433 = const()[name = tensor("op_4433"), val = tensor([1, 1280, 1, -1])]; + tensor x_511_cast_fp16 = reshape(shape = var_4433, x = attn_57_cast_fp16)[name = tensor("x_511_cast_fp16")]; + tensor layers_28_self_attn_o_proj_input_shift_to_fp16 = const()[name = tensor("layers_28_self_attn_o_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(289301504)))]; + tensor input_399_cast_fp16 = sub(x = x_511_cast_fp16, y = layers_28_self_attn_o_proj_input_shift_to_fp16)[name = tensor("input_399_cast_fp16")]; + tensor var_4441 = const()[name = tensor("op_4441"), val = tensor([1, 1])]; + tensor var_4443 = const()[name = tensor("op_4443"), val = tensor([1, 1])]; + tensor x_513_pad_type_0 = const()[name = tensor("x_513_pad_type_0"), val = tensor("custom")]; + tensor x_513_pad_0 = const()[name = tensor("x_513_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_28_self_attn_o_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(289304128))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(290123392))), name = tensor("layers_28_self_attn_o_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_28_self_attn_o_proj_module_bias_to_fp16 = const()[name = tensor("layers_28_self_attn_o_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(290123520)))]; + tensor x_513_cast_fp16 = conv(bias = layers_28_self_attn_o_proj_module_bias_to_fp16, dilations = var_4443, groups = var_4352, pad = x_513_pad_0, pad_type = x_513_pad_type_0, strides = var_4441, weight = layers_28_self_attn_o_proj_module_weight_to_fp16_palettized, x = input_399_cast_fp16)[name = tensor("x_513_cast_fp16")]; + tensor layers_28_self_attn_o_proj_output_scale_to_fp16 = const()[name = tensor("layers_28_self_attn_o_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(290126144)))]; + tensor obj_115_cast_fp16 = mul(x = x_513_cast_fp16, y = layers_28_self_attn_o_proj_output_scale_to_fp16)[name = tensor("obj_115_cast_fp16")]; + tensor inputs_115_cast_fp16 = add(x = inputs_113_cast_fp16, y = obj_115_cast_fp16)[name = tensor("inputs_115_cast_fp16")]; + tensor var_4450 = const()[name = tensor("op_4450"), val = tensor([1])]; + tensor channels_mean_115_cast_fp16 = reduce_mean(axes = var_4450, keep_dims = var_4353, x = inputs_115_cast_fp16)[name = tensor("channels_mean_115_cast_fp16")]; + tensor zero_mean_115_cast_fp16 = sub(x = inputs_115_cast_fp16, y = channels_mean_115_cast_fp16)[name = tensor("zero_mean_115_cast_fp16")]; + tensor zero_mean_sq_115_cast_fp16 = mul(x = zero_mean_115_cast_fp16, y = zero_mean_115_cast_fp16)[name = tensor("zero_mean_sq_115_cast_fp16")]; + tensor var_4454 = const()[name = tensor("op_4454"), val = tensor([1])]; + tensor var_4455_cast_fp16 = reduce_mean(axes = var_4454, keep_dims = var_4353, x = zero_mean_sq_115_cast_fp16)[name = tensor("op_4455_cast_fp16")]; + tensor var_4456_to_fp16 = const()[name = tensor("op_4456_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_4457_cast_fp16 = add(x = var_4455_cast_fp16, y = var_4456_to_fp16)[name = tensor("op_4457_cast_fp16")]; + tensor denom_115_epsilon_0_to_fp16 = const()[name = tensor("denom_115_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_115_cast_fp16 = rsqrt(epsilon = denom_115_epsilon_0_to_fp16, x = var_4457_cast_fp16)[name = tensor("denom_115_cast_fp16")]; + tensor out_115_cast_fp16 = mul(x = zero_mean_115_cast_fp16, y = denom_115_cast_fp16)[name = tensor("out_115_cast_fp16")]; + tensor x_515_gamma_0_to_fp16 = const()[name = tensor("x_515_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(290128768)))]; + tensor x_515_beta_0_to_fp16 = const()[name = tensor("x_515_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(290131392)))]; + tensor x_515_epsilon_0_to_fp16 = const()[name = tensor("x_515_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_515_cast_fp16 = batch_norm(beta = x_515_beta_0_to_fp16, epsilon = x_515_epsilon_0_to_fp16, gamma = x_515_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_115_cast_fp16)[name = tensor("x_515_cast_fp16")]; + tensor layers_28_fc1_input_shift_to_fp16 = const()[name = tensor("layers_28_fc1_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(290134016)))]; + tensor input_401_cast_fp16 = sub(x = x_515_cast_fp16, y = layers_28_fc1_input_shift_to_fp16)[name = tensor("input_401_cast_fp16")]; + tensor var_4472 = const()[name = tensor("op_4472"), val = tensor([1, 1])]; + tensor var_4474 = const()[name = tensor("op_4474"), val = tensor([1, 1])]; + tensor x_517_pad_type_0 = const()[name = tensor("x_517_pad_type_0"), val = tensor("custom")]; + tensor x_517_pad_0 = const()[name = tensor("x_517_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_28_fc1_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(290136640))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(293413504))), name = tensor("layers_28_fc1_module_weight_to_fp16_palettized"), shape = tensor([5120, 1280, 1, 1])]; + tensor layers_28_fc1_module_bias_to_fp16 = const()[name = tensor("layers_28_fc1_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(293413632)))]; + tensor x_517_cast_fp16 = conv(bias = layers_28_fc1_module_bias_to_fp16, dilations = var_4474, groups = var_4352, pad = x_517_pad_0, pad_type = x_517_pad_type_0, strides = var_4472, weight = layers_28_fc1_module_weight_to_fp16_palettized, x = input_401_cast_fp16)[name = tensor("x_517_cast_fp16")]; + tensor layers_28_fc1_output_scale_to_fp16 = const()[name = tensor("layers_28_fc1_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(293423936)))]; + tensor input_403_cast_fp16 = mul(x = x_517_cast_fp16, y = layers_28_fc1_output_scale_to_fp16)[name = tensor("input_403_cast_fp16")]; + tensor x_519_mode_0 = const()[name = tensor("x_519_mode_0"), val = tensor("EXACT")]; + tensor x_519_cast_fp16 = gelu(mode = x_519_mode_0, x = input_403_cast_fp16)[name = tensor("x_519_cast_fp16")]; + tensor layers_28_fc2_input_shift_to_fp16 = const()[name = tensor("layers_28_fc2_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(293434240)))]; + tensor input_405_cast_fp16 = sub(x = x_519_cast_fp16, y = layers_28_fc2_input_shift_to_fp16)[name = tensor("input_405_cast_fp16")]; + tensor var_4485 = const()[name = tensor("op_4485"), val = tensor([1, 1])]; + tensor var_4487 = const()[name = tensor("op_4487"), val = tensor([1, 1])]; + tensor x_521_pad_type_0 = const()[name = tensor("x_521_pad_type_0"), val = tensor("custom")]; + tensor x_521_pad_0 = const()[name = tensor("x_521_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_28_fc2_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(293444544))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(296721408))), name = tensor("layers_28_fc2_module_weight_to_fp16_palettized"), shape = tensor([1280, 5120, 1, 1])]; + tensor layers_28_fc2_module_bias_to_fp16 = const()[name = tensor("layers_28_fc2_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(296721536)))]; + tensor x_521_cast_fp16 = conv(bias = layers_28_fc2_module_bias_to_fp16, dilations = var_4487, groups = var_4352, pad = x_521_pad_0, pad_type = x_521_pad_type_0, strides = var_4485, weight = layers_28_fc2_module_weight_to_fp16_palettized, x = input_405_cast_fp16)[name = tensor("x_521_cast_fp16")]; + tensor layers_28_fc2_output_scale_to_fp16 = const()[name = tensor("layers_28_fc2_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(296724160)))]; + tensor hidden_states_61_cast_fp16 = mul(x = x_521_cast_fp16, y = layers_28_fc2_output_scale_to_fp16)[name = tensor("hidden_states_61_cast_fp16")]; + tensor inputs_117_cast_fp16 = add(x = inputs_115_cast_fp16, y = hidden_states_61_cast_fp16)[name = tensor("inputs_117_cast_fp16")]; + tensor var_4499 = const()[name = tensor("op_4499"), val = tensor(3)]; + tensor var_4501 = const()[name = tensor("op_4501"), val = tensor(1)]; + tensor var_4502 = const()[name = tensor("op_4502"), val = tensor(true)]; + tensor var_4512 = const()[name = tensor("op_4512"), val = tensor([1])]; + tensor channels_mean_117_cast_fp16 = reduce_mean(axes = var_4512, keep_dims = var_4502, x = inputs_117_cast_fp16)[name = tensor("channels_mean_117_cast_fp16")]; + tensor zero_mean_117_cast_fp16 = sub(x = inputs_117_cast_fp16, y = channels_mean_117_cast_fp16)[name = tensor("zero_mean_117_cast_fp16")]; + tensor zero_mean_sq_117_cast_fp16 = mul(x = zero_mean_117_cast_fp16, y = zero_mean_117_cast_fp16)[name = tensor("zero_mean_sq_117_cast_fp16")]; + tensor var_4516 = const()[name = tensor("op_4516"), val = tensor([1])]; + tensor var_4517_cast_fp16 = reduce_mean(axes = var_4516, keep_dims = var_4502, x = zero_mean_sq_117_cast_fp16)[name = tensor("op_4517_cast_fp16")]; + tensor var_4518_to_fp16 = const()[name = tensor("op_4518_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_4519_cast_fp16 = add(x = var_4517_cast_fp16, y = var_4518_to_fp16)[name = tensor("op_4519_cast_fp16")]; + tensor denom_117_epsilon_0_to_fp16 = const()[name = tensor("denom_117_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_117_cast_fp16 = rsqrt(epsilon = denom_117_epsilon_0_to_fp16, x = var_4519_cast_fp16)[name = tensor("denom_117_cast_fp16")]; + tensor out_117_cast_fp16 = mul(x = zero_mean_117_cast_fp16, y = denom_117_cast_fp16)[name = tensor("out_117_cast_fp16")]; + tensor obj_117_gamma_0_to_fp16 = const()[name = tensor("obj_117_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(296726784)))]; + tensor obj_117_beta_0_to_fp16 = const()[name = tensor("obj_117_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(296729408)))]; + tensor obj_117_epsilon_0_to_fp16 = const()[name = tensor("obj_117_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_117_cast_fp16 = batch_norm(beta = obj_117_beta_0_to_fp16, epsilon = obj_117_epsilon_0_to_fp16, gamma = obj_117_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_117_cast_fp16)[name = tensor("obj_117_cast_fp16")]; + tensor layers_29_self_attn_q_proj_input_shift_to_fp16 = const()[name = tensor("layers_29_self_attn_q_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(296732032)))]; + tensor input_407_cast_fp16 = sub(x = obj_117_cast_fp16, y = layers_29_self_attn_q_proj_input_shift_to_fp16)[name = tensor("input_407_cast_fp16")]; + tensor var_4538 = const()[name = tensor("op_4538"), val = tensor([1, 1])]; + tensor var_4540 = const()[name = tensor("op_4540"), val = tensor([1, 1])]; + tensor x_523_pad_type_0 = const()[name = tensor("x_523_pad_type_0"), val = tensor("custom")]; + tensor x_523_pad_0 = const()[name = tensor("x_523_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_29_self_attn_q_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(296734656))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(297553920))), name = tensor("layers_29_self_attn_q_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_29_self_attn_q_proj_module_bias_to_fp16 = const()[name = tensor("layers_29_self_attn_q_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(297554048)))]; + tensor x_523_cast_fp16 = conv(bias = layers_29_self_attn_q_proj_module_bias_to_fp16, dilations = var_4540, groups = var_4501, pad = x_523_pad_0, pad_type = x_523_pad_type_0, strides = var_4538, weight = layers_29_self_attn_q_proj_module_weight_to_fp16_palettized, x = input_407_cast_fp16)[name = tensor("x_523_cast_fp16")]; + tensor layers_29_self_attn_q_proj_output_scale_to_fp16 = const()[name = tensor("layers_29_self_attn_q_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(297556672)))]; + tensor query_59_cast_fp16 = mul(x = x_523_cast_fp16, y = layers_29_self_attn_q_proj_output_scale_to_fp16)[name = tensor("query_59_cast_fp16")]; + tensor var_4550 = const()[name = tensor("op_4550"), val = tensor([1, 1])]; + tensor var_4552 = const()[name = tensor("op_4552"), val = tensor([1, 1])]; + tensor x_525_pad_type_0 = const()[name = tensor("x_525_pad_type_0"), val = tensor("custom")]; + tensor x_525_pad_0 = const()[name = tensor("x_525_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_29_self_attn_k_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(297559296))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(298378560))), name = tensor("layers_29_self_attn_k_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_29_self_attn_k_proj_module_bias_to_fp16 = const()[name = tensor("layers_29_self_attn_k_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(298378688)))]; + tensor x_525_cast_fp16 = conv(bias = layers_29_self_attn_k_proj_module_bias_to_fp16, dilations = var_4552, groups = var_4501, pad = x_525_pad_0, pad_type = x_525_pad_type_0, strides = var_4550, weight = layers_29_self_attn_k_proj_module_weight_to_fp16_palettized, x = input_407_cast_fp16)[name = tensor("x_525_cast_fp16")]; + tensor layers_29_self_attn_k_proj_output_scale_to_fp16 = const()[name = tensor("layers_29_self_attn_k_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(298381312)))]; + tensor key_59_cast_fp16 = mul(x = x_525_cast_fp16, y = layers_29_self_attn_k_proj_output_scale_to_fp16)[name = tensor("key_59_cast_fp16")]; + tensor var_4562 = const()[name = tensor("op_4562"), val = tensor([1, 1])]; + tensor var_4564 = const()[name = tensor("op_4564"), val = tensor([1, 1])]; + tensor x_527_pad_type_0 = const()[name = tensor("x_527_pad_type_0"), val = tensor("custom")]; + tensor x_527_pad_0 = const()[name = tensor("x_527_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_29_self_attn_v_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(298383936))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(299203200))), name = tensor("layers_29_self_attn_v_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_29_self_attn_v_proj_module_bias_to_fp16 = const()[name = tensor("layers_29_self_attn_v_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(299203328)))]; + tensor x_527_cast_fp16 = conv(bias = layers_29_self_attn_v_proj_module_bias_to_fp16, dilations = var_4564, groups = var_4501, pad = x_527_pad_0, pad_type = x_527_pad_type_0, strides = var_4562, weight = layers_29_self_attn_v_proj_module_weight_to_fp16_palettized, x = input_407_cast_fp16)[name = tensor("x_527_cast_fp16")]; + tensor layers_29_self_attn_v_proj_output_scale_to_fp16 = const()[name = tensor("layers_29_self_attn_v_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(299205952)))]; + tensor value_59_cast_fp16 = mul(x = x_527_cast_fp16, y = layers_29_self_attn_v_proj_output_scale_to_fp16)[name = tensor("value_59_cast_fp16")]; + tensor var_4569 = const()[name = tensor("op_4569"), val = tensor([1, 20, 64, -1])]; + tensor var_4570_cast_fp16 = reshape(shape = var_4569, x = query_59_cast_fp16)[name = tensor("op_4570_cast_fp16")]; + tensor var_4571_to_fp16 = const()[name = tensor("op_4571_to_fp16"), val = tensor(0x1p-3)]; + tensor var_4572_cast_fp16 = mul(x = var_4570_cast_fp16, y = var_4571_to_fp16)[name = tensor("op_4572_cast_fp16")]; + tensor var_4573 = const()[name = tensor("op_4573"), val = tensor([1, 20, 64, -1])]; + tensor var_4574_cast_fp16 = reshape(shape = var_4573, x = key_59_cast_fp16)[name = tensor("op_4574_cast_fp16")]; + tensor mh_w_59_transpose_x_0 = const()[name = tensor("mh_w_59_transpose_x_0"), val = tensor(true)]; + tensor mh_w_59_transpose_y_0 = const()[name = tensor("mh_w_59_transpose_y_0"), val = tensor(false)]; + tensor mh_w_59_cast_fp16 = matmul(transpose_x = mh_w_59_transpose_x_0, transpose_y = mh_w_59_transpose_y_0, x = var_4572_cast_fp16, y = var_4574_cast_fp16)[name = tensor("mh_w_59_cast_fp16")]; + tensor var_4577_cast_fp16 = softmax(axis = var_4499, x = mh_w_59_cast_fp16)[name = tensor("op_4577_cast_fp16")]; + tensor var_4578 = const()[name = tensor("op_4578"), val = tensor([1, 20, 64, -1])]; + tensor var_4579_cast_fp16 = reshape(shape = var_4578, x = value_59_cast_fp16)[name = tensor("op_4579_cast_fp16")]; + tensor attn_59_transpose_x_0 = const()[name = tensor("attn_59_transpose_x_0"), val = tensor(false)]; + tensor attn_59_transpose_y_0 = const()[name = tensor("attn_59_transpose_y_0"), val = tensor(true)]; + tensor attn_59_cast_fp16 = matmul(transpose_x = attn_59_transpose_x_0, transpose_y = attn_59_transpose_y_0, x = var_4579_cast_fp16, y = var_4577_cast_fp16)[name = tensor("attn_59_cast_fp16")]; + tensor var_4582 = const()[name = tensor("op_4582"), val = tensor([1, 1280, 1, -1])]; + tensor x_529_cast_fp16 = reshape(shape = var_4582, x = attn_59_cast_fp16)[name = tensor("x_529_cast_fp16")]; + tensor layers_29_self_attn_o_proj_input_shift_to_fp16 = const()[name = tensor("layers_29_self_attn_o_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(299208576)))]; + tensor input_413_cast_fp16 = sub(x = x_529_cast_fp16, y = layers_29_self_attn_o_proj_input_shift_to_fp16)[name = tensor("input_413_cast_fp16")]; + tensor var_4590 = const()[name = tensor("op_4590"), val = tensor([1, 1])]; + tensor var_4592 = const()[name = tensor("op_4592"), val = tensor([1, 1])]; + tensor x_531_pad_type_0 = const()[name = tensor("x_531_pad_type_0"), val = tensor("custom")]; + tensor x_531_pad_0 = const()[name = tensor("x_531_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_29_self_attn_o_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(299211200))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(300030464))), name = tensor("layers_29_self_attn_o_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_29_self_attn_o_proj_module_bias_to_fp16 = const()[name = tensor("layers_29_self_attn_o_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(300030592)))]; + tensor x_531_cast_fp16 = conv(bias = layers_29_self_attn_o_proj_module_bias_to_fp16, dilations = var_4592, groups = var_4501, pad = x_531_pad_0, pad_type = x_531_pad_type_0, strides = var_4590, weight = layers_29_self_attn_o_proj_module_weight_to_fp16_palettized, x = input_413_cast_fp16)[name = tensor("x_531_cast_fp16")]; + tensor layers_29_self_attn_o_proj_output_scale_to_fp16 = const()[name = tensor("layers_29_self_attn_o_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(300033216)))]; + tensor obj_119_cast_fp16 = mul(x = x_531_cast_fp16, y = layers_29_self_attn_o_proj_output_scale_to_fp16)[name = tensor("obj_119_cast_fp16")]; + tensor inputs_119_cast_fp16 = add(x = inputs_117_cast_fp16, y = obj_119_cast_fp16)[name = tensor("inputs_119_cast_fp16")]; + tensor var_4599 = const()[name = tensor("op_4599"), val = tensor([1])]; + tensor channels_mean_119_cast_fp16 = reduce_mean(axes = var_4599, keep_dims = var_4502, x = inputs_119_cast_fp16)[name = tensor("channels_mean_119_cast_fp16")]; + tensor zero_mean_119_cast_fp16 = sub(x = inputs_119_cast_fp16, y = channels_mean_119_cast_fp16)[name = tensor("zero_mean_119_cast_fp16")]; + tensor zero_mean_sq_119_cast_fp16 = mul(x = zero_mean_119_cast_fp16, y = zero_mean_119_cast_fp16)[name = tensor("zero_mean_sq_119_cast_fp16")]; + tensor var_4603 = const()[name = tensor("op_4603"), val = tensor([1])]; + tensor var_4604_cast_fp16 = reduce_mean(axes = var_4603, keep_dims = var_4502, x = zero_mean_sq_119_cast_fp16)[name = tensor("op_4604_cast_fp16")]; + tensor var_4605_to_fp16 = const()[name = tensor("op_4605_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_4606_cast_fp16 = add(x = var_4604_cast_fp16, y = var_4605_to_fp16)[name = tensor("op_4606_cast_fp16")]; + tensor denom_119_epsilon_0_to_fp16 = const()[name = tensor("denom_119_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_119_cast_fp16 = rsqrt(epsilon = denom_119_epsilon_0_to_fp16, x = var_4606_cast_fp16)[name = tensor("denom_119_cast_fp16")]; + tensor out_119_cast_fp16 = mul(x = zero_mean_119_cast_fp16, y = denom_119_cast_fp16)[name = tensor("out_119_cast_fp16")]; + tensor x_533_gamma_0_to_fp16 = const()[name = tensor("x_533_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(300035840)))]; + tensor x_533_beta_0_to_fp16 = const()[name = tensor("x_533_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(300038464)))]; + tensor x_533_epsilon_0_to_fp16 = const()[name = tensor("x_533_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_533_cast_fp16 = batch_norm(beta = x_533_beta_0_to_fp16, epsilon = x_533_epsilon_0_to_fp16, gamma = x_533_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_119_cast_fp16)[name = tensor("x_533_cast_fp16")]; + tensor layers_29_fc1_input_shift_to_fp16 = const()[name = tensor("layers_29_fc1_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(300041088)))]; + tensor input_415_cast_fp16 = sub(x = x_533_cast_fp16, y = layers_29_fc1_input_shift_to_fp16)[name = tensor("input_415_cast_fp16")]; + tensor var_4621 = const()[name = tensor("op_4621"), val = tensor([1, 1])]; + tensor var_4623 = const()[name = tensor("op_4623"), val = tensor([1, 1])]; + tensor x_535_pad_type_0 = const()[name = tensor("x_535_pad_type_0"), val = tensor("custom")]; + tensor x_535_pad_0 = const()[name = tensor("x_535_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_29_fc1_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(300043712))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303320576))), name = tensor("layers_29_fc1_module_weight_to_fp16_palettized"), shape = tensor([5120, 1280, 1, 1])]; + tensor layers_29_fc1_module_bias_to_fp16 = const()[name = tensor("layers_29_fc1_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303320704)))]; + tensor x_535_cast_fp16 = conv(bias = layers_29_fc1_module_bias_to_fp16, dilations = var_4623, groups = var_4501, pad = x_535_pad_0, pad_type = x_535_pad_type_0, strides = var_4621, weight = layers_29_fc1_module_weight_to_fp16_palettized, x = input_415_cast_fp16)[name = tensor("x_535_cast_fp16")]; + tensor layers_29_fc1_output_scale_to_fp16 = const()[name = tensor("layers_29_fc1_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303331008)))]; + tensor input_417_cast_fp16 = mul(x = x_535_cast_fp16, y = layers_29_fc1_output_scale_to_fp16)[name = tensor("input_417_cast_fp16")]; + tensor x_537_mode_0 = const()[name = tensor("x_537_mode_0"), val = tensor("EXACT")]; + tensor x_537_cast_fp16 = gelu(mode = x_537_mode_0, x = input_417_cast_fp16)[name = tensor("x_537_cast_fp16")]; + tensor layers_29_fc2_input_shift_to_fp16 = const()[name = tensor("layers_29_fc2_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303341312)))]; + tensor input_419_cast_fp16 = sub(x = x_537_cast_fp16, y = layers_29_fc2_input_shift_to_fp16)[name = tensor("input_419_cast_fp16")]; + tensor var_4634 = const()[name = tensor("op_4634"), val = tensor([1, 1])]; + tensor var_4636 = const()[name = tensor("op_4636"), val = tensor([1, 1])]; + tensor x_539_pad_type_0 = const()[name = tensor("x_539_pad_type_0"), val = tensor("custom")]; + tensor x_539_pad_0 = const()[name = tensor("x_539_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_29_fc2_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303351616))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(306628480))), name = tensor("layers_29_fc2_module_weight_to_fp16_palettized"), shape = tensor([1280, 5120, 1, 1])]; + tensor layers_29_fc2_module_bias_to_fp16 = const()[name = tensor("layers_29_fc2_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(306628608)))]; + tensor x_539_cast_fp16 = conv(bias = layers_29_fc2_module_bias_to_fp16, dilations = var_4636, groups = var_4501, pad = x_539_pad_0, pad_type = x_539_pad_type_0, strides = var_4634, weight = layers_29_fc2_module_weight_to_fp16_palettized, x = input_419_cast_fp16)[name = tensor("x_539_cast_fp16")]; + tensor layers_29_fc2_output_scale_to_fp16 = const()[name = tensor("layers_29_fc2_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(306631232)))]; + tensor hidden_states_63_cast_fp16 = mul(x = x_539_cast_fp16, y = layers_29_fc2_output_scale_to_fp16)[name = tensor("hidden_states_63_cast_fp16")]; + tensor inputs_121_cast_fp16 = add(x = inputs_119_cast_fp16, y = hidden_states_63_cast_fp16)[name = tensor("inputs_121_cast_fp16")]; + tensor var_4648 = const()[name = tensor("op_4648"), val = tensor(3)]; + tensor var_4650 = const()[name = tensor("op_4650"), val = tensor(1)]; + tensor var_4651 = const()[name = tensor("op_4651"), val = tensor(true)]; + tensor var_4661 = const()[name = tensor("op_4661"), val = tensor([1])]; + tensor channels_mean_121_cast_fp16 = reduce_mean(axes = var_4661, keep_dims = var_4651, x = inputs_121_cast_fp16)[name = tensor("channels_mean_121_cast_fp16")]; + tensor zero_mean_121_cast_fp16 = sub(x = inputs_121_cast_fp16, y = channels_mean_121_cast_fp16)[name = tensor("zero_mean_121_cast_fp16")]; + tensor zero_mean_sq_121_cast_fp16 = mul(x = zero_mean_121_cast_fp16, y = zero_mean_121_cast_fp16)[name = tensor("zero_mean_sq_121_cast_fp16")]; + tensor var_4665 = const()[name = tensor("op_4665"), val = tensor([1])]; + tensor var_4666_cast_fp16 = reduce_mean(axes = var_4665, keep_dims = var_4651, x = zero_mean_sq_121_cast_fp16)[name = tensor("op_4666_cast_fp16")]; + tensor var_4667_to_fp16 = const()[name = tensor("op_4667_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_4668_cast_fp16 = add(x = var_4666_cast_fp16, y = var_4667_to_fp16)[name = tensor("op_4668_cast_fp16")]; + tensor denom_121_epsilon_0_to_fp16 = const()[name = tensor("denom_121_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_121_cast_fp16 = rsqrt(epsilon = denom_121_epsilon_0_to_fp16, x = var_4668_cast_fp16)[name = tensor("denom_121_cast_fp16")]; + tensor out_121_cast_fp16 = mul(x = zero_mean_121_cast_fp16, y = denom_121_cast_fp16)[name = tensor("out_121_cast_fp16")]; + tensor obj_121_gamma_0_to_fp16 = const()[name = tensor("obj_121_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(306633856)))]; + tensor obj_121_beta_0_to_fp16 = const()[name = tensor("obj_121_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(306636480)))]; + tensor obj_121_epsilon_0_to_fp16 = const()[name = tensor("obj_121_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_121_cast_fp16 = batch_norm(beta = obj_121_beta_0_to_fp16, epsilon = obj_121_epsilon_0_to_fp16, gamma = obj_121_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_121_cast_fp16)[name = tensor("obj_121_cast_fp16")]; + tensor layers_30_self_attn_q_proj_input_shift_to_fp16 = const()[name = tensor("layers_30_self_attn_q_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(306639104)))]; + tensor input_421_cast_fp16 = sub(x = obj_121_cast_fp16, y = layers_30_self_attn_q_proj_input_shift_to_fp16)[name = tensor("input_421_cast_fp16")]; + tensor var_4687 = const()[name = tensor("op_4687"), val = tensor([1, 1])]; + tensor var_4689 = const()[name = tensor("op_4689"), val = tensor([1, 1])]; + tensor x_541_pad_type_0 = const()[name = tensor("x_541_pad_type_0"), val = tensor("custom")]; + tensor x_541_pad_0 = const()[name = tensor("x_541_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_30_self_attn_q_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(306641728))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(307460992))), name = tensor("layers_30_self_attn_q_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_30_self_attn_q_proj_module_bias_to_fp16 = const()[name = tensor("layers_30_self_attn_q_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(307461120)))]; + tensor x_541_cast_fp16 = conv(bias = layers_30_self_attn_q_proj_module_bias_to_fp16, dilations = var_4689, groups = var_4650, pad = x_541_pad_0, pad_type = x_541_pad_type_0, strides = var_4687, weight = layers_30_self_attn_q_proj_module_weight_to_fp16_palettized, x = input_421_cast_fp16)[name = tensor("x_541_cast_fp16")]; + tensor layers_30_self_attn_q_proj_output_scale_to_fp16 = const()[name = tensor("layers_30_self_attn_q_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(307463744)))]; + tensor query_61_cast_fp16 = mul(x = x_541_cast_fp16, y = layers_30_self_attn_q_proj_output_scale_to_fp16)[name = tensor("query_61_cast_fp16")]; + tensor var_4699 = const()[name = tensor("op_4699"), val = tensor([1, 1])]; + tensor var_4701 = const()[name = tensor("op_4701"), val = tensor([1, 1])]; + tensor x_543_pad_type_0 = const()[name = tensor("x_543_pad_type_0"), val = tensor("custom")]; + tensor x_543_pad_0 = const()[name = tensor("x_543_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_30_self_attn_k_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(307466368))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(308285632))), name = tensor("layers_30_self_attn_k_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_30_self_attn_k_proj_module_bias_to_fp16 = const()[name = tensor("layers_30_self_attn_k_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(308285760)))]; + tensor x_543_cast_fp16 = conv(bias = layers_30_self_attn_k_proj_module_bias_to_fp16, dilations = var_4701, groups = var_4650, pad = x_543_pad_0, pad_type = x_543_pad_type_0, strides = var_4699, weight = layers_30_self_attn_k_proj_module_weight_to_fp16_palettized, x = input_421_cast_fp16)[name = tensor("x_543_cast_fp16")]; + tensor layers_30_self_attn_k_proj_output_scale_to_fp16 = const()[name = tensor("layers_30_self_attn_k_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(308288384)))]; + tensor key_61_cast_fp16 = mul(x = x_543_cast_fp16, y = layers_30_self_attn_k_proj_output_scale_to_fp16)[name = tensor("key_61_cast_fp16")]; + tensor var_4711 = const()[name = tensor("op_4711"), val = tensor([1, 1])]; + tensor var_4713 = const()[name = tensor("op_4713"), val = tensor([1, 1])]; + tensor x_545_pad_type_0 = const()[name = tensor("x_545_pad_type_0"), val = tensor("custom")]; + tensor x_545_pad_0 = const()[name = tensor("x_545_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_30_self_attn_v_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(308291008))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(309110272))), name = tensor("layers_30_self_attn_v_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_30_self_attn_v_proj_module_bias_to_fp16 = const()[name = tensor("layers_30_self_attn_v_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(309110400)))]; + tensor x_545_cast_fp16 = conv(bias = layers_30_self_attn_v_proj_module_bias_to_fp16, dilations = var_4713, groups = var_4650, pad = x_545_pad_0, pad_type = x_545_pad_type_0, strides = var_4711, weight = layers_30_self_attn_v_proj_module_weight_to_fp16_palettized, x = input_421_cast_fp16)[name = tensor("x_545_cast_fp16")]; + tensor layers_30_self_attn_v_proj_output_scale_to_fp16 = const()[name = tensor("layers_30_self_attn_v_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(309113024)))]; + tensor value_61_cast_fp16 = mul(x = x_545_cast_fp16, y = layers_30_self_attn_v_proj_output_scale_to_fp16)[name = tensor("value_61_cast_fp16")]; + tensor var_4718 = const()[name = tensor("op_4718"), val = tensor([1, 20, 64, -1])]; + tensor var_4719_cast_fp16 = reshape(shape = var_4718, x = query_61_cast_fp16)[name = tensor("op_4719_cast_fp16")]; + tensor var_4720_to_fp16 = const()[name = tensor("op_4720_to_fp16"), val = tensor(0x1p-3)]; + tensor var_4721_cast_fp16 = mul(x = var_4719_cast_fp16, y = var_4720_to_fp16)[name = tensor("op_4721_cast_fp16")]; + tensor var_4722 = const()[name = tensor("op_4722"), val = tensor([1, 20, 64, -1])]; + tensor var_4723_cast_fp16 = reshape(shape = var_4722, x = key_61_cast_fp16)[name = tensor("op_4723_cast_fp16")]; + tensor mh_w_61_transpose_x_0 = const()[name = tensor("mh_w_61_transpose_x_0"), val = tensor(true)]; + tensor mh_w_61_transpose_y_0 = const()[name = tensor("mh_w_61_transpose_y_0"), val = tensor(false)]; + tensor mh_w_61_cast_fp16 = matmul(transpose_x = mh_w_61_transpose_x_0, transpose_y = mh_w_61_transpose_y_0, x = var_4721_cast_fp16, y = var_4723_cast_fp16)[name = tensor("mh_w_61_cast_fp16")]; + tensor var_4726_cast_fp16 = softmax(axis = var_4648, x = mh_w_61_cast_fp16)[name = tensor("op_4726_cast_fp16")]; + tensor var_4727 = const()[name = tensor("op_4727"), val = tensor([1, 20, 64, -1])]; + tensor var_4728_cast_fp16 = reshape(shape = var_4727, x = value_61_cast_fp16)[name = tensor("op_4728_cast_fp16")]; + tensor attn_61_transpose_x_0 = const()[name = tensor("attn_61_transpose_x_0"), val = tensor(false)]; + tensor attn_61_transpose_y_0 = const()[name = tensor("attn_61_transpose_y_0"), val = tensor(true)]; + tensor attn_61_cast_fp16 = matmul(transpose_x = attn_61_transpose_x_0, transpose_y = attn_61_transpose_y_0, x = var_4728_cast_fp16, y = var_4726_cast_fp16)[name = tensor("attn_61_cast_fp16")]; + tensor var_4731 = const()[name = tensor("op_4731"), val = tensor([1, 1280, 1, -1])]; + tensor x_547_cast_fp16 = reshape(shape = var_4731, x = attn_61_cast_fp16)[name = tensor("x_547_cast_fp16")]; + tensor layers_30_self_attn_o_proj_input_shift_to_fp16 = const()[name = tensor("layers_30_self_attn_o_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(309115648)))]; + tensor input_427_cast_fp16 = sub(x = x_547_cast_fp16, y = layers_30_self_attn_o_proj_input_shift_to_fp16)[name = tensor("input_427_cast_fp16")]; + tensor var_4739 = const()[name = tensor("op_4739"), val = tensor([1, 1])]; + tensor var_4741 = const()[name = tensor("op_4741"), val = tensor([1, 1])]; + tensor x_549_pad_type_0 = const()[name = tensor("x_549_pad_type_0"), val = tensor("custom")]; + tensor x_549_pad_0 = const()[name = tensor("x_549_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_30_self_attn_o_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(309118272))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(309937536))), name = tensor("layers_30_self_attn_o_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_30_self_attn_o_proj_module_bias_to_fp16 = const()[name = tensor("layers_30_self_attn_o_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(309937664)))]; + tensor x_549_cast_fp16 = conv(bias = layers_30_self_attn_o_proj_module_bias_to_fp16, dilations = var_4741, groups = var_4650, pad = x_549_pad_0, pad_type = x_549_pad_type_0, strides = var_4739, weight = layers_30_self_attn_o_proj_module_weight_to_fp16_palettized, x = input_427_cast_fp16)[name = tensor("x_549_cast_fp16")]; + tensor layers_30_self_attn_o_proj_output_scale_to_fp16 = const()[name = tensor("layers_30_self_attn_o_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(309940288)))]; + tensor obj_123_cast_fp16 = mul(x = x_549_cast_fp16, y = layers_30_self_attn_o_proj_output_scale_to_fp16)[name = tensor("obj_123_cast_fp16")]; + tensor inputs_123_cast_fp16 = add(x = inputs_121_cast_fp16, y = obj_123_cast_fp16)[name = tensor("inputs_123_cast_fp16")]; + tensor var_4748 = const()[name = tensor("op_4748"), val = tensor([1])]; + tensor channels_mean_123_cast_fp16 = reduce_mean(axes = var_4748, keep_dims = var_4651, x = inputs_123_cast_fp16)[name = tensor("channels_mean_123_cast_fp16")]; + tensor zero_mean_123_cast_fp16 = sub(x = inputs_123_cast_fp16, y = channels_mean_123_cast_fp16)[name = tensor("zero_mean_123_cast_fp16")]; + tensor zero_mean_sq_123_cast_fp16 = mul(x = zero_mean_123_cast_fp16, y = zero_mean_123_cast_fp16)[name = tensor("zero_mean_sq_123_cast_fp16")]; + tensor var_4752 = const()[name = tensor("op_4752"), val = tensor([1])]; + tensor var_4753_cast_fp16 = reduce_mean(axes = var_4752, keep_dims = var_4651, x = zero_mean_sq_123_cast_fp16)[name = tensor("op_4753_cast_fp16")]; + tensor var_4754_to_fp16 = const()[name = tensor("op_4754_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_4755_cast_fp16 = add(x = var_4753_cast_fp16, y = var_4754_to_fp16)[name = tensor("op_4755_cast_fp16")]; + tensor denom_123_epsilon_0_to_fp16 = const()[name = tensor("denom_123_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_123_cast_fp16 = rsqrt(epsilon = denom_123_epsilon_0_to_fp16, x = var_4755_cast_fp16)[name = tensor("denom_123_cast_fp16")]; + tensor out_123_cast_fp16 = mul(x = zero_mean_123_cast_fp16, y = denom_123_cast_fp16)[name = tensor("out_123_cast_fp16")]; + tensor x_551_gamma_0_to_fp16 = const()[name = tensor("x_551_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(309942912)))]; + tensor x_551_beta_0_to_fp16 = const()[name = tensor("x_551_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(309945536)))]; + tensor x_551_epsilon_0_to_fp16 = const()[name = tensor("x_551_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_551_cast_fp16 = batch_norm(beta = x_551_beta_0_to_fp16, epsilon = x_551_epsilon_0_to_fp16, gamma = x_551_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_123_cast_fp16)[name = tensor("x_551_cast_fp16")]; + tensor layers_30_fc1_input_shift_to_fp16 = const()[name = tensor("layers_30_fc1_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(309948160)))]; + tensor input_429_cast_fp16 = sub(x = x_551_cast_fp16, y = layers_30_fc1_input_shift_to_fp16)[name = tensor("input_429_cast_fp16")]; + tensor var_4770 = const()[name = tensor("op_4770"), val = tensor([1, 1])]; + tensor var_4772 = const()[name = tensor("op_4772"), val = tensor([1, 1])]; + tensor x_553_pad_type_0 = const()[name = tensor("x_553_pad_type_0"), val = tensor("custom")]; + tensor x_553_pad_0 = const()[name = tensor("x_553_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_30_fc1_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(309950784))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(313227648))), name = tensor("layers_30_fc1_module_weight_to_fp16_palettized"), shape = tensor([5120, 1280, 1, 1])]; + tensor layers_30_fc1_module_bias_to_fp16 = const()[name = tensor("layers_30_fc1_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(313227776)))]; + tensor x_553_cast_fp16 = conv(bias = layers_30_fc1_module_bias_to_fp16, dilations = var_4772, groups = var_4650, pad = x_553_pad_0, pad_type = x_553_pad_type_0, strides = var_4770, weight = layers_30_fc1_module_weight_to_fp16_palettized, x = input_429_cast_fp16)[name = tensor("x_553_cast_fp16")]; + tensor layers_30_fc1_output_scale_to_fp16 = const()[name = tensor("layers_30_fc1_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(313238080)))]; + tensor input_431_cast_fp16 = mul(x = x_553_cast_fp16, y = layers_30_fc1_output_scale_to_fp16)[name = tensor("input_431_cast_fp16")]; + tensor x_555_mode_0 = const()[name = tensor("x_555_mode_0"), val = tensor("EXACT")]; + tensor x_555_cast_fp16 = gelu(mode = x_555_mode_0, x = input_431_cast_fp16)[name = tensor("x_555_cast_fp16")]; + tensor layers_30_fc2_input_shift_to_fp16 = const()[name = tensor("layers_30_fc2_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(313248384)))]; + tensor input_433_cast_fp16 = sub(x = x_555_cast_fp16, y = layers_30_fc2_input_shift_to_fp16)[name = tensor("input_433_cast_fp16")]; + tensor var_4783 = const()[name = tensor("op_4783"), val = tensor([1, 1])]; + tensor var_4785 = const()[name = tensor("op_4785"), val = tensor([1, 1])]; + tensor x_557_pad_type_0 = const()[name = tensor("x_557_pad_type_0"), val = tensor("custom")]; + tensor x_557_pad_0 = const()[name = tensor("x_557_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_30_fc2_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(313258688))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(316535552))), name = tensor("layers_30_fc2_module_weight_to_fp16_palettized"), shape = tensor([1280, 5120, 1, 1])]; + tensor layers_30_fc2_module_bias_to_fp16 = const()[name = tensor("layers_30_fc2_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(316535680)))]; + tensor x_557_cast_fp16 = conv(bias = layers_30_fc2_module_bias_to_fp16, dilations = var_4785, groups = var_4650, pad = x_557_pad_0, pad_type = x_557_pad_type_0, strides = var_4783, weight = layers_30_fc2_module_weight_to_fp16_palettized, x = input_433_cast_fp16)[name = tensor("x_557_cast_fp16")]; + tensor layers_30_fc2_output_scale_to_fp16 = const()[name = tensor("layers_30_fc2_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(316538304)))]; + tensor hidden_states_65_cast_fp16 = mul(x = x_557_cast_fp16, y = layers_30_fc2_output_scale_to_fp16)[name = tensor("hidden_states_65_cast_fp16")]; + tensor inputs_125_cast_fp16 = add(x = inputs_123_cast_fp16, y = hidden_states_65_cast_fp16)[name = tensor("inputs_125_cast_fp16")]; + tensor var_4797 = const()[name = tensor("op_4797"), val = tensor(3)]; + tensor var_4799 = const()[name = tensor("op_4799"), val = tensor(1)]; + tensor var_4800 = const()[name = tensor("op_4800"), val = tensor(true)]; + tensor var_4810 = const()[name = tensor("op_4810"), val = tensor([1])]; + tensor channels_mean_125_cast_fp16 = reduce_mean(axes = var_4810, keep_dims = var_4800, x = inputs_125_cast_fp16)[name = tensor("channels_mean_125_cast_fp16")]; + tensor zero_mean_125_cast_fp16 = sub(x = inputs_125_cast_fp16, y = channels_mean_125_cast_fp16)[name = tensor("zero_mean_125_cast_fp16")]; + tensor zero_mean_sq_125_cast_fp16 = mul(x = zero_mean_125_cast_fp16, y = zero_mean_125_cast_fp16)[name = tensor("zero_mean_sq_125_cast_fp16")]; + tensor var_4814 = const()[name = tensor("op_4814"), val = tensor([1])]; + tensor var_4815_cast_fp16 = reduce_mean(axes = var_4814, keep_dims = var_4800, x = zero_mean_sq_125_cast_fp16)[name = tensor("op_4815_cast_fp16")]; + tensor var_4816_to_fp16 = const()[name = tensor("op_4816_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_4817_cast_fp16 = add(x = var_4815_cast_fp16, y = var_4816_to_fp16)[name = tensor("op_4817_cast_fp16")]; + tensor denom_125_epsilon_0_to_fp16 = const()[name = tensor("denom_125_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_125_cast_fp16 = rsqrt(epsilon = denom_125_epsilon_0_to_fp16, x = var_4817_cast_fp16)[name = tensor("denom_125_cast_fp16")]; + tensor out_125_cast_fp16 = mul(x = zero_mean_125_cast_fp16, y = denom_125_cast_fp16)[name = tensor("out_125_cast_fp16")]; + tensor obj_125_gamma_0_to_fp16 = const()[name = tensor("obj_125_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(316540928)))]; + tensor obj_125_beta_0_to_fp16 = const()[name = tensor("obj_125_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(316543552)))]; + tensor obj_125_epsilon_0_to_fp16 = const()[name = tensor("obj_125_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_125_cast_fp16 = batch_norm(beta = obj_125_beta_0_to_fp16, epsilon = obj_125_epsilon_0_to_fp16, gamma = obj_125_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_125_cast_fp16)[name = tensor("obj_125_cast_fp16")]; + tensor layers_31_self_attn_q_proj_input_shift_to_fp16 = const()[name = tensor("layers_31_self_attn_q_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(316546176)))]; + tensor input_435_cast_fp16 = sub(x = obj_125_cast_fp16, y = layers_31_self_attn_q_proj_input_shift_to_fp16)[name = tensor("input_435_cast_fp16")]; + tensor var_4836 = const()[name = tensor("op_4836"), val = tensor([1, 1])]; + tensor var_4838 = const()[name = tensor("op_4838"), val = tensor([1, 1])]; + tensor x_559_pad_type_0 = const()[name = tensor("x_559_pad_type_0"), val = tensor("custom")]; + tensor x_559_pad_0 = const()[name = tensor("x_559_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_31_self_attn_q_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(316548800))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(317368064))), name = tensor("layers_31_self_attn_q_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_31_self_attn_q_proj_module_bias_to_fp16 = const()[name = tensor("layers_31_self_attn_q_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(317368192)))]; + tensor x_559_cast_fp16 = conv(bias = layers_31_self_attn_q_proj_module_bias_to_fp16, dilations = var_4838, groups = var_4799, pad = x_559_pad_0, pad_type = x_559_pad_type_0, strides = var_4836, weight = layers_31_self_attn_q_proj_module_weight_to_fp16_palettized, x = input_435_cast_fp16)[name = tensor("x_559_cast_fp16")]; + tensor layers_31_self_attn_q_proj_output_scale_to_fp16 = const()[name = tensor("layers_31_self_attn_q_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(317370816)))]; + tensor query_cast_fp16 = mul(x = x_559_cast_fp16, y = layers_31_self_attn_q_proj_output_scale_to_fp16)[name = tensor("query_cast_fp16")]; + tensor var_4848 = const()[name = tensor("op_4848"), val = tensor([1, 1])]; + tensor var_4850 = const()[name = tensor("op_4850"), val = tensor([1, 1])]; + tensor x_561_pad_type_0 = const()[name = tensor("x_561_pad_type_0"), val = tensor("custom")]; + tensor x_561_pad_0 = const()[name = tensor("x_561_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_31_self_attn_k_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(317373440))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(318192704))), name = tensor("layers_31_self_attn_k_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_31_self_attn_k_proj_module_bias_to_fp16 = const()[name = tensor("layers_31_self_attn_k_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(318192832)))]; + tensor x_561_cast_fp16 = conv(bias = layers_31_self_attn_k_proj_module_bias_to_fp16, dilations = var_4850, groups = var_4799, pad = x_561_pad_0, pad_type = x_561_pad_type_0, strides = var_4848, weight = layers_31_self_attn_k_proj_module_weight_to_fp16_palettized, x = input_435_cast_fp16)[name = tensor("x_561_cast_fp16")]; + tensor layers_31_self_attn_k_proj_output_scale_to_fp16 = const()[name = tensor("layers_31_self_attn_k_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(318195456)))]; + tensor key_cast_fp16 = mul(x = x_561_cast_fp16, y = layers_31_self_attn_k_proj_output_scale_to_fp16)[name = tensor("key_cast_fp16")]; + tensor var_4860 = const()[name = tensor("op_4860"), val = tensor([1, 1])]; + tensor var_4862 = const()[name = tensor("op_4862"), val = tensor([1, 1])]; + tensor x_563_pad_type_0 = const()[name = tensor("x_563_pad_type_0"), val = tensor("custom")]; + tensor x_563_pad_0 = const()[name = tensor("x_563_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_31_self_attn_v_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(318198080))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(319017344))), name = tensor("layers_31_self_attn_v_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_31_self_attn_v_proj_module_bias_to_fp16 = const()[name = tensor("layers_31_self_attn_v_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(319017472)))]; + tensor x_563_cast_fp16 = conv(bias = layers_31_self_attn_v_proj_module_bias_to_fp16, dilations = var_4862, groups = var_4799, pad = x_563_pad_0, pad_type = x_563_pad_type_0, strides = var_4860, weight = layers_31_self_attn_v_proj_module_weight_to_fp16_palettized, x = input_435_cast_fp16)[name = tensor("x_563_cast_fp16")]; + tensor layers_31_self_attn_v_proj_output_scale_to_fp16 = const()[name = tensor("layers_31_self_attn_v_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(319020096)))]; + tensor value_cast_fp16 = mul(x = x_563_cast_fp16, y = layers_31_self_attn_v_proj_output_scale_to_fp16)[name = tensor("value_cast_fp16")]; + tensor var_4867 = const()[name = tensor("op_4867"), val = tensor([1, 20, 64, -1])]; + tensor var_4868_cast_fp16 = reshape(shape = var_4867, x = query_cast_fp16)[name = tensor("op_4868_cast_fp16")]; + tensor var_4869_to_fp16 = const()[name = tensor("op_4869_to_fp16"), val = tensor(0x1p-3)]; + tensor var_4870_cast_fp16 = mul(x = var_4868_cast_fp16, y = var_4869_to_fp16)[name = tensor("op_4870_cast_fp16")]; + tensor var_4871 = const()[name = tensor("op_4871"), val = tensor([1, 20, 64, -1])]; + tensor var_4872_cast_fp16 = reshape(shape = var_4871, x = key_cast_fp16)[name = tensor("op_4872_cast_fp16")]; + tensor mh_w_transpose_x_0 = const()[name = tensor("mh_w_transpose_x_0"), val = tensor(true)]; + tensor mh_w_transpose_y_0 = const()[name = tensor("mh_w_transpose_y_0"), val = tensor(false)]; + tensor mh_w_cast_fp16 = matmul(transpose_x = mh_w_transpose_x_0, transpose_y = mh_w_transpose_y_0, x = var_4870_cast_fp16, y = var_4872_cast_fp16)[name = tensor("mh_w_cast_fp16")]; + tensor var_4875_cast_fp16 = softmax(axis = var_4797, x = mh_w_cast_fp16)[name = tensor("op_4875_cast_fp16")]; + tensor var_4876 = const()[name = tensor("op_4876"), val = tensor([1, 20, 64, -1])]; + tensor var_4877_cast_fp16 = reshape(shape = var_4876, x = value_cast_fp16)[name = tensor("op_4877_cast_fp16")]; + tensor attn_transpose_x_0 = const()[name = tensor("attn_transpose_x_0"), val = tensor(false)]; + tensor attn_transpose_y_0 = const()[name = tensor("attn_transpose_y_0"), val = tensor(true)]; + tensor attn_cast_fp16 = matmul(transpose_x = attn_transpose_x_0, transpose_y = attn_transpose_y_0, x = var_4877_cast_fp16, y = var_4875_cast_fp16)[name = tensor("attn_cast_fp16")]; + tensor var_4880 = const()[name = tensor("op_4880"), val = tensor([1, 1280, 1, -1])]; + tensor x_565_cast_fp16 = reshape(shape = var_4880, x = attn_cast_fp16)[name = tensor("x_565_cast_fp16")]; + tensor layers_31_self_attn_o_proj_input_shift_to_fp16 = const()[name = tensor("layers_31_self_attn_o_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(319022720)))]; + tensor input_441_cast_fp16 = sub(x = x_565_cast_fp16, y = layers_31_self_attn_o_proj_input_shift_to_fp16)[name = tensor("input_441_cast_fp16")]; + tensor var_4888 = const()[name = tensor("op_4888"), val = tensor([1, 1])]; + tensor var_4890 = const()[name = tensor("op_4890"), val = tensor([1, 1])]; + tensor x_567_pad_type_0 = const()[name = tensor("x_567_pad_type_0"), val = tensor("custom")]; + tensor x_567_pad_0 = const()[name = tensor("x_567_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_31_self_attn_o_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(319025344))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(319844608))), name = tensor("layers_31_self_attn_o_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_31_self_attn_o_proj_module_bias_to_fp16 = const()[name = tensor("layers_31_self_attn_o_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(319844736)))]; + tensor x_567_cast_fp16 = conv(bias = layers_31_self_attn_o_proj_module_bias_to_fp16, dilations = var_4890, groups = var_4799, pad = x_567_pad_0, pad_type = x_567_pad_type_0, strides = var_4888, weight = layers_31_self_attn_o_proj_module_weight_to_fp16_palettized, x = input_441_cast_fp16)[name = tensor("x_567_cast_fp16")]; + tensor layers_31_self_attn_o_proj_output_scale_to_fp16 = const()[name = tensor("layers_31_self_attn_o_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(319847360)))]; + tensor obj_cast_fp16 = mul(x = x_567_cast_fp16, y = layers_31_self_attn_o_proj_output_scale_to_fp16)[name = tensor("obj_cast_fp16")]; + tensor inputs_127_cast_fp16 = add(x = inputs_125_cast_fp16, y = obj_cast_fp16)[name = tensor("inputs_127_cast_fp16")]; + tensor var_4897 = const()[name = tensor("op_4897"), val = tensor([1])]; + tensor channels_mean_127_cast_fp16 = reduce_mean(axes = var_4897, keep_dims = var_4800, x = inputs_127_cast_fp16)[name = tensor("channels_mean_127_cast_fp16")]; + tensor zero_mean_127_cast_fp16 = sub(x = inputs_127_cast_fp16, y = channels_mean_127_cast_fp16)[name = tensor("zero_mean_127_cast_fp16")]; + tensor zero_mean_sq_127_cast_fp16 = mul(x = zero_mean_127_cast_fp16, y = zero_mean_127_cast_fp16)[name = tensor("zero_mean_sq_127_cast_fp16")]; + tensor var_4901 = const()[name = tensor("op_4901"), val = tensor([1])]; + tensor var_4902_cast_fp16 = reduce_mean(axes = var_4901, keep_dims = var_4800, x = zero_mean_sq_127_cast_fp16)[name = tensor("op_4902_cast_fp16")]; + tensor var_4903_to_fp16 = const()[name = tensor("op_4903_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_4904_cast_fp16 = add(x = var_4902_cast_fp16, y = var_4903_to_fp16)[name = tensor("op_4904_cast_fp16")]; + tensor denom_127_epsilon_0_to_fp16 = const()[name = tensor("denom_127_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_127_cast_fp16 = rsqrt(epsilon = denom_127_epsilon_0_to_fp16, x = var_4904_cast_fp16)[name = tensor("denom_127_cast_fp16")]; + tensor out_127_cast_fp16 = mul(x = zero_mean_127_cast_fp16, y = denom_127_cast_fp16)[name = tensor("out_127_cast_fp16")]; + tensor x_569_gamma_0_to_fp16 = const()[name = tensor("x_569_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(319849984)))]; + tensor x_569_beta_0_to_fp16 = const()[name = tensor("x_569_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(319852608)))]; + tensor x_569_epsilon_0_to_fp16 = const()[name = tensor("x_569_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_569_cast_fp16 = batch_norm(beta = x_569_beta_0_to_fp16, epsilon = x_569_epsilon_0_to_fp16, gamma = x_569_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_127_cast_fp16)[name = tensor("x_569_cast_fp16")]; + tensor layers_31_fc1_input_shift_to_fp16 = const()[name = tensor("layers_31_fc1_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(319855232)))]; + tensor input_443_cast_fp16 = sub(x = x_569_cast_fp16, y = layers_31_fc1_input_shift_to_fp16)[name = tensor("input_443_cast_fp16")]; + tensor var_4919 = const()[name = tensor("op_4919"), val = tensor([1, 1])]; + tensor var_4921 = const()[name = tensor("op_4921"), val = tensor([1, 1])]; + tensor x_571_pad_type_0 = const()[name = tensor("x_571_pad_type_0"), val = tensor("custom")]; + tensor x_571_pad_0 = const()[name = tensor("x_571_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_31_fc1_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(319857856))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(323134720))), name = tensor("layers_31_fc1_module_weight_to_fp16_palettized"), shape = tensor([5120, 1280, 1, 1])]; + tensor layers_31_fc1_module_bias_to_fp16 = const()[name = tensor("layers_31_fc1_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(323134848)))]; + tensor x_571_cast_fp16 = conv(bias = layers_31_fc1_module_bias_to_fp16, dilations = var_4921, groups = var_4799, pad = x_571_pad_0, pad_type = x_571_pad_type_0, strides = var_4919, weight = layers_31_fc1_module_weight_to_fp16_palettized, x = input_443_cast_fp16)[name = tensor("x_571_cast_fp16")]; + tensor layers_31_fc1_output_scale_to_fp16 = const()[name = tensor("layers_31_fc1_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(323145152)))]; + tensor input_445_cast_fp16 = mul(x = x_571_cast_fp16, y = layers_31_fc1_output_scale_to_fp16)[name = tensor("input_445_cast_fp16")]; + tensor x_573_mode_0 = const()[name = tensor("x_573_mode_0"), val = tensor("EXACT")]; + tensor x_573_cast_fp16 = gelu(mode = x_573_mode_0, x = input_445_cast_fp16)[name = tensor("x_573_cast_fp16")]; + tensor layers_31_fc2_input_shift_to_fp16 = const()[name = tensor("layers_31_fc2_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(323155456)))]; + tensor input_cast_fp16 = sub(x = x_573_cast_fp16, y = layers_31_fc2_input_shift_to_fp16)[name = tensor("input_cast_fp16")]; + tensor var_4932 = const()[name = tensor("op_4932"), val = tensor([1, 1])]; + tensor var_4934 = const()[name = tensor("op_4934"), val = tensor([1, 1])]; + tensor x_pad_type_0 = const()[name = tensor("x_pad_type_0"), val = tensor("custom")]; + tensor x_pad_0 = const()[name = tensor("x_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_31_fc2_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(323165760))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(326442624))), name = tensor("layers_31_fc2_module_weight_to_fp16_palettized"), shape = tensor([1280, 5120, 1, 1])]; + tensor layers_31_fc2_module_bias_to_fp16 = const()[name = tensor("layers_31_fc2_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(326442752)))]; + tensor x_cast_fp16 = conv(bias = layers_31_fc2_module_bias_to_fp16, dilations = var_4934, groups = var_4799, pad = x_pad_0, pad_type = x_pad_type_0, strides = var_4932, weight = layers_31_fc2_module_weight_to_fp16_palettized, x = input_cast_fp16)[name = tensor("x_cast_fp16")]; + tensor layers_31_fc2_output_scale_to_fp16 = const()[name = tensor("layers_31_fc2_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(326445376)))]; + tensor hidden_states_cast_fp16 = mul(x = x_cast_fp16, y = layers_31_fc2_output_scale_to_fp16)[name = tensor("hidden_states_cast_fp16")]; + tensor inputs_cast_fp16 = add(x = inputs_127_cast_fp16, y = hidden_states_cast_fp16)[name = tensor("inputs_cast_fp16")]; + tensor var_4941 = const()[name = tensor("op_4941"), val = tensor(true)]; + tensor var_4945 = const()[name = tensor("op_4945"), val = tensor([1])]; + tensor channels_mean_cast_fp16 = reduce_mean(axes = var_4945, keep_dims = var_4941, x = inputs_cast_fp16)[name = tensor("channels_mean_cast_fp16")]; + tensor zero_mean_cast_fp16 = sub(x = inputs_cast_fp16, y = channels_mean_cast_fp16)[name = tensor("zero_mean_cast_fp16")]; + tensor zero_mean_sq_cast_fp16 = mul(x = zero_mean_cast_fp16, y = zero_mean_cast_fp16)[name = tensor("zero_mean_sq_cast_fp16")]; + tensor var_4949 = const()[name = tensor("op_4949"), val = tensor([1])]; + tensor var_4950_cast_fp16 = reduce_mean(axes = var_4949, keep_dims = var_4941, x = zero_mean_sq_cast_fp16)[name = tensor("op_4950_cast_fp16")]; + tensor var_4951_to_fp16 = const()[name = tensor("op_4951_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_4952_cast_fp16 = add(x = var_4950_cast_fp16, y = var_4951_to_fp16)[name = tensor("op_4952_cast_fp16")]; + tensor denom_epsilon_0_to_fp16 = const()[name = tensor("denom_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_cast_fp16 = rsqrt(epsilon = denom_epsilon_0_to_fp16, x = var_4952_cast_fp16)[name = tensor("denom_cast_fp16")]; + tensor out_cast_fp16 = mul(x = zero_mean_cast_fp16, y = denom_cast_fp16)[name = tensor("out_cast_fp16")]; + tensor encoder_output_embeds_type_fp32_gamma_0_to_fp16 = const()[name = tensor("encoder_output_embeds_type_fp32_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(326448000)))]; + tensor encoder_output_embeds_type_fp32_beta_0_to_fp16 = const()[name = tensor("encoder_output_embeds_type_fp32_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(326450624)))]; + tensor encoder_output_embeds_type_fp32_epsilon_0_to_fp16 = const()[name = tensor("encoder_output_embeds_type_fp32_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor encoder_output_embeds = batch_norm(beta = encoder_output_embeds_type_fp32_beta_0_to_fp16, epsilon = encoder_output_embeds_type_fp32_epsilon_0_to_fp16, gamma = encoder_output_embeds_type_fp32_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_cast_fp16)[name = tensor("encoder_output_embeds_type_fp32_cast_fp16")]; + } -> (encoder_output_embeds); +} \ No newline at end of file