diff --git "a/openai_whisper-tiny.en/AudioEncoder.mlmodelc/model.mil" "b/openai_whisper-tiny.en/AudioEncoder.mlmodelc/model.mil" new file mode 100644--- /dev/null +++ "b/openai_whisper-tiny.en/AudioEncoder.mlmodelc/model.mil" @@ -0,0 +1,1766 @@ +program(1.0) +[buildInfo = dict, tensor>({{"coremlc-component-MIL", "5.33.5"}, {"coremlc-version", "1877.40.3"}, {"coremltools-component-torch", "2.2.1"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "7.1"}})] +{ + func main(tensor melspectrogram_features) { + tensor var_34 = const()[name = tensor("op_34"), val = tensor([1, 1])]; + tensor var_40 = const()[name = tensor("op_40"), val = tensor([1, 1])]; + tensor var_45 = const()[name = tensor("op_45"), val = tensor(1)]; + tensor var_50_pad_type_0 = const()[name = tensor("op_50_pad_type_0"), val = tensor("custom")]; + tensor var_50_pad_0 = const()[name = tensor("op_50_pad_0"), val = tensor([0, 0, 1, 1])]; + tensor var_25_to_fp16 = const()[name = tensor("op_25_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64)))]; + tensor var_31_to_fp16 = const()[name = tensor("op_31_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(184448)))]; + tensor var_50_cast_fp16 = conv(bias = var_31_to_fp16, dilations = var_40, groups = var_45, pad = var_50_pad_0, pad_type = var_50_pad_type_0, strides = var_34, weight = var_25_to_fp16, x = melspectrogram_features)[name = tensor("op_50_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_50_cast_fp16)[name = tensor("hidden_states_1_cast_fp16")]; + tensor var_74 = const()[name = tensor("op_74"), val = tensor([2, 2])]; + tensor var_80 = const()[name = tensor("op_80"), val = tensor([1, 1])]; + tensor var_85 = const()[name = tensor("op_85"), val = tensor(1)]; + tensor var_90_pad_type_0 = const()[name = tensor("op_90_pad_type_0"), val = tensor("custom")]; + tensor var_90_pad_0 = const()[name = tensor("op_90_pad_0"), val = tensor([0, 0, 1, 1])]; + tensor var_65_to_fp16 = const()[name = tensor("op_65_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(185280)))]; + tensor var_71_to_fp16 = const()[name = tensor("op_71_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1070080)))]; + tensor var_90_cast_fp16 = conv(bias = var_71_to_fp16, dilations = var_80, groups = var_85, pad = var_90_pad_0, pad_type = var_90_pad_type_0, strides = var_74, weight = var_65_to_fp16, x = hidden_states_1_cast_fp16)[name = tensor("op_90_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_90_cast_fp16)[name = tensor("hidden_states_3_cast_fp16")]; + tensor var_108_to_fp16 = const()[name = tensor("op_108_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1070912)))]; + tensor inputs_1_cast_fp16 = add(x = hidden_states_3_cast_fp16, y = var_108_to_fp16)[name = tensor("inputs_1_cast_fp16")]; + tensor var_118 = const()[name = tensor("op_118"), val = tensor(3)]; + tensor var_129 = const()[name = tensor("op_129"), val = tensor(1)]; + tensor var_130 = const()[name = tensor("op_130"), val = tensor(true)]; + tensor var_140 = const()[name = tensor("op_140"), val = tensor([1])]; + tensor channels_mean_1_cast_fp16 = reduce_mean(axes = var_140, keep_dims = var_130, 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_144 = const()[name = tensor("op_144"), val = tensor([1])]; + tensor var_145_cast_fp16 = reduce_mean(axes = var_144, keep_dims = var_130, x = zero_mean_sq_1_cast_fp16)[name = tensor("op_145_cast_fp16")]; + tensor var_146_to_fp16 = const()[name = tensor("op_146_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_147_cast_fp16 = add(x = var_145_cast_fp16, y = var_146_to_fp16)[name = tensor("op_147_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_147_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(2222976)))]; + 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(2223808)))]; + 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(2224640)))]; + 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(2225472)))]; + 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 var_162 = const()[name = tensor("op_162"), val = tensor([1, 1])]; + tensor var_164 = const()[name = tensor("op_164"), val = tensor([1, 1])]; + tensor query_1_pad_type_0 = const()[name = tensor("query_1_pad_type_0"), val = tensor("custom")]; + tensor query_1_pad_0 = const()[name = tensor("query_1_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_0_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_0_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(2226304)))]; + tensor layers_0_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_0_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(2521280)))]; + tensor query_1_cast_fp16 = conv(bias = layers_0_self_attn_q_proj_bias_to_fp16, dilations = var_164, groups = var_129, pad = query_1_pad_0, pad_type = query_1_pad_type_0, strides = var_162, weight = layers_0_self_attn_q_proj_weight_to_fp16, x = obj_1_cast_fp16)[name = tensor("query_1_cast_fp16")]; + tensor var_168 = const()[name = tensor("op_168"), val = tensor([1, 1])]; + tensor var_170 = const()[name = tensor("op_170"), val = tensor([1, 1])]; + tensor key_1_pad_type_0 = const()[name = tensor("key_1_pad_type_0"), val = tensor("custom")]; + tensor key_1_pad_0 = const()[name = tensor("key_1_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_0_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_0_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(2522112)))]; + tensor key_1_cast_fp16 = conv(dilations = var_170, groups = var_129, pad = key_1_pad_0, pad_type = key_1_pad_type_0, strides = var_168, weight = layers_0_self_attn_k_proj_weight_to_fp16, x = obj_1_cast_fp16)[name = tensor("key_1_cast_fp16")]; + tensor var_175 = const()[name = tensor("op_175"), val = tensor([1, 1])]; + tensor var_177 = const()[name = tensor("op_177"), val = tensor([1, 1])]; + tensor value_1_pad_type_0 = const()[name = tensor("value_1_pad_type_0"), val = tensor("custom")]; + tensor value_1_pad_0 = const()[name = tensor("value_1_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_0_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_0_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(2817088)))]; + tensor layers_0_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_0_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3112064)))]; + tensor value_1_cast_fp16 = conv(bias = layers_0_self_attn_v_proj_bias_to_fp16, dilations = var_177, groups = var_129, pad = value_1_pad_0, pad_type = value_1_pad_type_0, strides = var_175, weight = layers_0_self_attn_v_proj_weight_to_fp16, x = obj_1_cast_fp16)[name = tensor("value_1_cast_fp16")]; + tensor var_184_begin_0 = const()[name = tensor("op_184_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_184_end_0 = const()[name = tensor("op_184_end_0"), val = tensor([1, 64, 1, 1500])]; + tensor var_184_end_mask_0 = const()[name = tensor("op_184_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_184_cast_fp16 = slice_by_index(begin = var_184_begin_0, end = var_184_end_0, end_mask = var_184_end_mask_0, x = query_1_cast_fp16)[name = tensor("op_184_cast_fp16")]; + tensor var_188_begin_0 = const()[name = tensor("op_188_begin_0"), val = tensor([0, 64, 0, 0])]; + tensor var_188_end_0 = const()[name = tensor("op_188_end_0"), val = tensor([1, 128, 1, 1500])]; + tensor var_188_end_mask_0 = const()[name = tensor("op_188_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_188_cast_fp16 = slice_by_index(begin = var_188_begin_0, end = var_188_end_0, end_mask = var_188_end_mask_0, x = query_1_cast_fp16)[name = tensor("op_188_cast_fp16")]; + tensor var_192_begin_0 = const()[name = tensor("op_192_begin_0"), val = tensor([0, 128, 0, 0])]; + tensor var_192_end_0 = const()[name = tensor("op_192_end_0"), val = tensor([1, 192, 1, 1500])]; + tensor var_192_end_mask_0 = const()[name = tensor("op_192_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_192_cast_fp16 = slice_by_index(begin = var_192_begin_0, end = var_192_end_0, end_mask = var_192_end_mask_0, x = query_1_cast_fp16)[name = tensor("op_192_cast_fp16")]; + tensor var_196_begin_0 = const()[name = tensor("op_196_begin_0"), val = tensor([0, 192, 0, 0])]; + tensor var_196_end_0 = const()[name = tensor("op_196_end_0"), val = tensor([1, 256, 1, 1500])]; + tensor var_196_end_mask_0 = const()[name = tensor("op_196_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_196_cast_fp16 = slice_by_index(begin = var_196_begin_0, end = var_196_end_0, end_mask = var_196_end_mask_0, x = query_1_cast_fp16)[name = tensor("op_196_cast_fp16")]; + tensor var_200_begin_0 = const()[name = tensor("op_200_begin_0"), val = tensor([0, 256, 0, 0])]; + tensor var_200_end_0 = const()[name = tensor("op_200_end_0"), val = tensor([1, 320, 1, 1500])]; + tensor var_200_end_mask_0 = const()[name = tensor("op_200_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_200_cast_fp16 = slice_by_index(begin = var_200_begin_0, end = var_200_end_0, end_mask = var_200_end_mask_0, x = query_1_cast_fp16)[name = tensor("op_200_cast_fp16")]; + tensor var_204_begin_0 = const()[name = tensor("op_204_begin_0"), val = tensor([0, 320, 0, 0])]; + tensor var_204_end_0 = const()[name = tensor("op_204_end_0"), val = tensor([1, 384, 1, 1500])]; + tensor var_204_end_mask_0 = const()[name = tensor("op_204_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_204_cast_fp16 = slice_by_index(begin = var_204_begin_0, end = var_204_end_0, end_mask = var_204_end_mask_0, x = query_1_cast_fp16)[name = tensor("op_204_cast_fp16")]; + tensor var_213_begin_0 = const()[name = tensor("op_213_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_213_end_0 = const()[name = tensor("op_213_end_0"), val = tensor([1, 64, 1, 375])]; + tensor var_213_end_mask_0 = const()[name = tensor("op_213_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_213_cast_fp16 = slice_by_index(begin = var_213_begin_0, end = var_213_end_0, end_mask = var_213_end_mask_0, x = var_184_cast_fp16)[name = tensor("op_213_cast_fp16")]; + tensor var_220_begin_0 = const()[name = tensor("op_220_begin_0"), val = tensor([0, 0, 0, 375])]; + tensor var_220_end_0 = const()[name = tensor("op_220_end_0"), val = tensor([1, 64, 1, 750])]; + tensor var_220_end_mask_0 = const()[name = tensor("op_220_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_220_cast_fp16 = slice_by_index(begin = var_220_begin_0, end = var_220_end_0, end_mask = var_220_end_mask_0, x = var_184_cast_fp16)[name = tensor("op_220_cast_fp16")]; + tensor var_227_begin_0 = const()[name = tensor("op_227_begin_0"), val = tensor([0, 0, 0, 750])]; + tensor var_227_end_0 = const()[name = tensor("op_227_end_0"), val = tensor([1, 64, 1, 1125])]; + tensor var_227_end_mask_0 = const()[name = tensor("op_227_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_227_cast_fp16 = slice_by_index(begin = var_227_begin_0, end = var_227_end_0, end_mask = var_227_end_mask_0, x = var_184_cast_fp16)[name = tensor("op_227_cast_fp16")]; + tensor var_234_begin_0 = const()[name = tensor("op_234_begin_0"), val = tensor([0, 0, 0, 1125])]; + tensor var_234_end_0 = const()[name = tensor("op_234_end_0"), val = tensor([1, 64, 1, 1500])]; + tensor var_234_end_mask_0 = const()[name = tensor("op_234_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_234_cast_fp16 = slice_by_index(begin = var_234_begin_0, end = var_234_end_0, end_mask = var_234_end_mask_0, x = var_184_cast_fp16)[name = tensor("op_234_cast_fp16")]; + tensor var_241_begin_0 = const()[name = tensor("op_241_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_241_end_0 = const()[name = tensor("op_241_end_0"), val = tensor([1, 64, 1, 375])]; + tensor var_241_end_mask_0 = const()[name = tensor("op_241_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_241_cast_fp16 = slice_by_index(begin = var_241_begin_0, end = var_241_end_0, end_mask = var_241_end_mask_0, x = var_188_cast_fp16)[name = tensor("op_241_cast_fp16")]; + tensor var_248_begin_0 = const()[name = tensor("op_248_begin_0"), val = tensor([0, 0, 0, 375])]; + tensor var_248_end_0 = const()[name = tensor("op_248_end_0"), val = tensor([1, 64, 1, 750])]; + tensor var_248_end_mask_0 = const()[name = tensor("op_248_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_248_cast_fp16 = slice_by_index(begin = var_248_begin_0, end = var_248_end_0, end_mask = var_248_end_mask_0, x = var_188_cast_fp16)[name = tensor("op_248_cast_fp16")]; + tensor var_255_begin_0 = const()[name = tensor("op_255_begin_0"), val = tensor([0, 0, 0, 750])]; + tensor var_255_end_0 = const()[name = tensor("op_255_end_0"), val = tensor([1, 64, 1, 1125])]; + tensor var_255_end_mask_0 = const()[name = tensor("op_255_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_255_cast_fp16 = slice_by_index(begin = var_255_begin_0, end = var_255_end_0, end_mask = var_255_end_mask_0, x = var_188_cast_fp16)[name = tensor("op_255_cast_fp16")]; + tensor var_262_begin_0 = const()[name = tensor("op_262_begin_0"), val = tensor([0, 0, 0, 1125])]; + tensor var_262_end_0 = const()[name = tensor("op_262_end_0"), val = tensor([1, 64, 1, 1500])]; + tensor var_262_end_mask_0 = const()[name = tensor("op_262_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_262_cast_fp16 = slice_by_index(begin = var_262_begin_0, end = var_262_end_0, end_mask = var_262_end_mask_0, x = var_188_cast_fp16)[name = tensor("op_262_cast_fp16")]; + tensor var_269_begin_0 = const()[name = tensor("op_269_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_269_end_0 = const()[name = tensor("op_269_end_0"), val = tensor([1, 64, 1, 375])]; + tensor var_269_end_mask_0 = const()[name = tensor("op_269_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_269_cast_fp16 = slice_by_index(begin = var_269_begin_0, end = var_269_end_0, end_mask = var_269_end_mask_0, x = var_192_cast_fp16)[name = tensor("op_269_cast_fp16")]; + tensor var_276_begin_0 = const()[name = tensor("op_276_begin_0"), val = tensor([0, 0, 0, 375])]; + tensor var_276_end_0 = const()[name = tensor("op_276_end_0"), val = tensor([1, 64, 1, 750])]; + tensor var_276_end_mask_0 = const()[name = tensor("op_276_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_276_cast_fp16 = slice_by_index(begin = var_276_begin_0, end = var_276_end_0, end_mask = var_276_end_mask_0, x = var_192_cast_fp16)[name = tensor("op_276_cast_fp16")]; + tensor var_283_begin_0 = const()[name = tensor("op_283_begin_0"), val = tensor([0, 0, 0, 750])]; + tensor var_283_end_0 = const()[name = tensor("op_283_end_0"), val = tensor([1, 64, 1, 1125])]; + tensor var_283_end_mask_0 = const()[name = tensor("op_283_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_283_cast_fp16 = slice_by_index(begin = var_283_begin_0, end = var_283_end_0, end_mask = var_283_end_mask_0, x = var_192_cast_fp16)[name = tensor("op_283_cast_fp16")]; + tensor var_290_begin_0 = const()[name = tensor("op_290_begin_0"), val = tensor([0, 0, 0, 1125])]; + tensor var_290_end_0 = const()[name = tensor("op_290_end_0"), val = tensor([1, 64, 1, 1500])]; + tensor var_290_end_mask_0 = const()[name = tensor("op_290_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_290_cast_fp16 = slice_by_index(begin = var_290_begin_0, end = var_290_end_0, end_mask = var_290_end_mask_0, x = var_192_cast_fp16)[name = tensor("op_290_cast_fp16")]; + tensor var_297_begin_0 = const()[name = tensor("op_297_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_297_end_0 = const()[name = tensor("op_297_end_0"), val = tensor([1, 64, 1, 375])]; + tensor var_297_end_mask_0 = const()[name = tensor("op_297_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_297_cast_fp16 = slice_by_index(begin = var_297_begin_0, end = var_297_end_0, end_mask = var_297_end_mask_0, x = var_196_cast_fp16)[name = tensor("op_297_cast_fp16")]; + tensor var_304_begin_0 = const()[name = tensor("op_304_begin_0"), val = tensor([0, 0, 0, 375])]; + tensor var_304_end_0 = const()[name = tensor("op_304_end_0"), val = tensor([1, 64, 1, 750])]; + tensor var_304_end_mask_0 = const()[name = tensor("op_304_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_304_cast_fp16 = slice_by_index(begin = var_304_begin_0, end = var_304_end_0, end_mask = var_304_end_mask_0, x = var_196_cast_fp16)[name = tensor("op_304_cast_fp16")]; + tensor var_311_begin_0 = const()[name = tensor("op_311_begin_0"), val = tensor([0, 0, 0, 750])]; + tensor var_311_end_0 = const()[name = tensor("op_311_end_0"), val = tensor([1, 64, 1, 1125])]; + tensor var_311_end_mask_0 = const()[name = tensor("op_311_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_311_cast_fp16 = slice_by_index(begin = var_311_begin_0, end = var_311_end_0, end_mask = var_311_end_mask_0, x = var_196_cast_fp16)[name = tensor("op_311_cast_fp16")]; + tensor var_318_begin_0 = const()[name = tensor("op_318_begin_0"), val = tensor([0, 0, 0, 1125])]; + tensor var_318_end_0 = const()[name = tensor("op_318_end_0"), val = tensor([1, 64, 1, 1500])]; + tensor var_318_end_mask_0 = const()[name = tensor("op_318_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_318_cast_fp16 = slice_by_index(begin = var_318_begin_0, end = var_318_end_0, end_mask = var_318_end_mask_0, x = var_196_cast_fp16)[name = tensor("op_318_cast_fp16")]; + tensor var_325_begin_0 = const()[name = tensor("op_325_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_325_end_0 = const()[name = tensor("op_325_end_0"), val = tensor([1, 64, 1, 375])]; + tensor var_325_end_mask_0 = const()[name = tensor("op_325_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_325_cast_fp16 = slice_by_index(begin = var_325_begin_0, end = var_325_end_0, end_mask = var_325_end_mask_0, x = var_200_cast_fp16)[name = tensor("op_325_cast_fp16")]; + tensor var_332_begin_0 = const()[name = tensor("op_332_begin_0"), val = tensor([0, 0, 0, 375])]; + tensor var_332_end_0 = const()[name = tensor("op_332_end_0"), val = tensor([1, 64, 1, 750])]; + tensor var_332_end_mask_0 = const()[name = tensor("op_332_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_332_cast_fp16 = slice_by_index(begin = var_332_begin_0, end = var_332_end_0, end_mask = var_332_end_mask_0, x = var_200_cast_fp16)[name = tensor("op_332_cast_fp16")]; + tensor var_339_begin_0 = const()[name = tensor("op_339_begin_0"), val = tensor([0, 0, 0, 750])]; + tensor var_339_end_0 = const()[name = tensor("op_339_end_0"), val = tensor([1, 64, 1, 1125])]; + tensor var_339_end_mask_0 = const()[name = tensor("op_339_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_339_cast_fp16 = slice_by_index(begin = var_339_begin_0, end = var_339_end_0, end_mask = var_339_end_mask_0, x = var_200_cast_fp16)[name = tensor("op_339_cast_fp16")]; + tensor var_346_begin_0 = const()[name = tensor("op_346_begin_0"), val = tensor([0, 0, 0, 1125])]; + tensor var_346_end_0 = const()[name = tensor("op_346_end_0"), val = tensor([1, 64, 1, 1500])]; + tensor var_346_end_mask_0 = const()[name = tensor("op_346_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_346_cast_fp16 = slice_by_index(begin = var_346_begin_0, end = var_346_end_0, end_mask = var_346_end_mask_0, x = var_200_cast_fp16)[name = tensor("op_346_cast_fp16")]; + tensor var_353_begin_0 = const()[name = tensor("op_353_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_353_end_0 = const()[name = tensor("op_353_end_0"), val = tensor([1, 64, 1, 375])]; + tensor var_353_end_mask_0 = const()[name = tensor("op_353_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_353_cast_fp16 = slice_by_index(begin = var_353_begin_0, end = var_353_end_0, end_mask = var_353_end_mask_0, x = var_204_cast_fp16)[name = tensor("op_353_cast_fp16")]; + tensor var_360_begin_0 = const()[name = tensor("op_360_begin_0"), val = tensor([0, 0, 0, 375])]; + tensor var_360_end_0 = const()[name = tensor("op_360_end_0"), val = tensor([1, 64, 1, 750])]; + tensor var_360_end_mask_0 = const()[name = tensor("op_360_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_360_cast_fp16 = slice_by_index(begin = var_360_begin_0, end = var_360_end_0, end_mask = var_360_end_mask_0, x = var_204_cast_fp16)[name = tensor("op_360_cast_fp16")]; + tensor var_367_begin_0 = const()[name = tensor("op_367_begin_0"), val = tensor([0, 0, 0, 750])]; + tensor var_367_end_0 = const()[name = tensor("op_367_end_0"), val = tensor([1, 64, 1, 1125])]; + tensor var_367_end_mask_0 = const()[name = tensor("op_367_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_367_cast_fp16 = slice_by_index(begin = var_367_begin_0, end = var_367_end_0, end_mask = var_367_end_mask_0, x = var_204_cast_fp16)[name = tensor("op_367_cast_fp16")]; + tensor var_374_begin_0 = const()[name = tensor("op_374_begin_0"), val = tensor([0, 0, 0, 1125])]; + tensor var_374_end_0 = const()[name = tensor("op_374_end_0"), val = tensor([1, 64, 1, 1500])]; + tensor var_374_end_mask_0 = const()[name = tensor("op_374_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_374_cast_fp16 = slice_by_index(begin = var_374_begin_0, end = var_374_end_0, end_mask = var_374_end_mask_0, x = var_204_cast_fp16)[name = tensor("op_374_cast_fp16")]; + tensor k_1_perm_0 = const()[name = tensor("k_1_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor var_379_begin_0 = const()[name = tensor("op_379_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_379_end_0 = const()[name = tensor("op_379_end_0"), val = tensor([1, 1500, 1, 64])]; + tensor var_379_end_mask_0 = const()[name = tensor("op_379_end_mask_0"), val = tensor([true, true, true, false])]; + tensor transpose_3 = transpose(perm = k_1_perm_0, x = key_1_cast_fp16)[name = tensor("transpose_3")]; + tensor var_379_cast_fp16 = slice_by_index(begin = var_379_begin_0, end = var_379_end_0, end_mask = var_379_end_mask_0, x = transpose_3)[name = tensor("op_379_cast_fp16")]; + tensor var_383_begin_0 = const()[name = tensor("op_383_begin_0"), val = tensor([0, 0, 0, 64])]; + tensor var_383_end_0 = const()[name = tensor("op_383_end_0"), val = tensor([1, 1500, 1, 128])]; + tensor var_383_end_mask_0 = const()[name = tensor("op_383_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_383_cast_fp16 = slice_by_index(begin = var_383_begin_0, end = var_383_end_0, end_mask = var_383_end_mask_0, x = transpose_3)[name = tensor("op_383_cast_fp16")]; + tensor var_387_begin_0 = const()[name = tensor("op_387_begin_0"), val = tensor([0, 0, 0, 128])]; + tensor var_387_end_0 = const()[name = tensor("op_387_end_0"), val = tensor([1, 1500, 1, 192])]; + tensor var_387_end_mask_0 = const()[name = tensor("op_387_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_387_cast_fp16 = slice_by_index(begin = var_387_begin_0, end = var_387_end_0, end_mask = var_387_end_mask_0, x = transpose_3)[name = tensor("op_387_cast_fp16")]; + tensor var_391_begin_0 = const()[name = tensor("op_391_begin_0"), val = tensor([0, 0, 0, 192])]; + tensor var_391_end_0 = const()[name = tensor("op_391_end_0"), val = tensor([1, 1500, 1, 256])]; + tensor var_391_end_mask_0 = const()[name = tensor("op_391_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_391_cast_fp16 = slice_by_index(begin = var_391_begin_0, end = var_391_end_0, end_mask = var_391_end_mask_0, x = transpose_3)[name = tensor("op_391_cast_fp16")]; + tensor var_395_begin_0 = const()[name = tensor("op_395_begin_0"), val = tensor([0, 0, 0, 256])]; + tensor var_395_end_0 = const()[name = tensor("op_395_end_0"), val = tensor([1, 1500, 1, 320])]; + tensor var_395_end_mask_0 = const()[name = tensor("op_395_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_395_cast_fp16 = slice_by_index(begin = var_395_begin_0, end = var_395_end_0, end_mask = var_395_end_mask_0, x = transpose_3)[name = tensor("op_395_cast_fp16")]; + tensor var_399_begin_0 = const()[name = tensor("op_399_begin_0"), val = tensor([0, 0, 0, 320])]; + tensor var_399_end_0 = const()[name = tensor("op_399_end_0"), val = tensor([1, 1500, 1, 384])]; + tensor var_399_end_mask_0 = const()[name = tensor("op_399_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_399_cast_fp16 = slice_by_index(begin = var_399_begin_0, end = var_399_end_0, end_mask = var_399_end_mask_0, x = transpose_3)[name = tensor("op_399_cast_fp16")]; + tensor var_401_begin_0 = const()[name = tensor("op_401_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_401_end_0 = const()[name = tensor("op_401_end_0"), val = tensor([1, 64, 1, 1500])]; + tensor var_401_end_mask_0 = const()[name = tensor("op_401_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_401_cast_fp16 = slice_by_index(begin = var_401_begin_0, end = var_401_end_0, end_mask = var_401_end_mask_0, x = value_1_cast_fp16)[name = tensor("op_401_cast_fp16")]; + tensor var_405_begin_0 = const()[name = tensor("op_405_begin_0"), val = tensor([0, 64, 0, 0])]; + tensor var_405_end_0 = const()[name = tensor("op_405_end_0"), val = tensor([1, 128, 1, 1500])]; + tensor var_405_end_mask_0 = const()[name = tensor("op_405_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_405_cast_fp16 = slice_by_index(begin = var_405_begin_0, end = var_405_end_0, end_mask = var_405_end_mask_0, x = value_1_cast_fp16)[name = tensor("op_405_cast_fp16")]; + tensor var_409_begin_0 = const()[name = tensor("op_409_begin_0"), val = tensor([0, 128, 0, 0])]; + tensor var_409_end_0 = const()[name = tensor("op_409_end_0"), val = tensor([1, 192, 1, 1500])]; + tensor var_409_end_mask_0 = const()[name = tensor("op_409_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_409_cast_fp16 = slice_by_index(begin = var_409_begin_0, end = var_409_end_0, end_mask = var_409_end_mask_0, x = value_1_cast_fp16)[name = tensor("op_409_cast_fp16")]; + tensor var_413_begin_0 = const()[name = tensor("op_413_begin_0"), val = tensor([0, 192, 0, 0])]; + tensor var_413_end_0 = const()[name = tensor("op_413_end_0"), val = tensor([1, 256, 1, 1500])]; + tensor var_413_end_mask_0 = const()[name = tensor("op_413_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_413_cast_fp16 = slice_by_index(begin = var_413_begin_0, end = var_413_end_0, end_mask = var_413_end_mask_0, x = value_1_cast_fp16)[name = tensor("op_413_cast_fp16")]; + tensor var_417_begin_0 = const()[name = tensor("op_417_begin_0"), val = tensor([0, 256, 0, 0])]; + tensor var_417_end_0 = const()[name = tensor("op_417_end_0"), val = tensor([1, 320, 1, 1500])]; + tensor var_417_end_mask_0 = const()[name = tensor("op_417_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_417_cast_fp16 = slice_by_index(begin = var_417_begin_0, end = var_417_end_0, end_mask = var_417_end_mask_0, x = value_1_cast_fp16)[name = tensor("op_417_cast_fp16")]; + tensor var_421_begin_0 = const()[name = tensor("op_421_begin_0"), val = tensor([0, 320, 0, 0])]; + tensor var_421_end_0 = const()[name = tensor("op_421_end_0"), val = tensor([1, 384, 1, 1500])]; + tensor var_421_end_mask_0 = const()[name = tensor("op_421_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_421_cast_fp16 = slice_by_index(begin = var_421_begin_0, end = var_421_end_0, end_mask = var_421_end_mask_0, x = value_1_cast_fp16)[name = tensor("op_421_cast_fp16")]; + tensor var_425_equation_0 = const()[name = tensor("op_425_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor var_425_cast_fp16 = einsum(equation = var_425_equation_0, values = (var_379_cast_fp16, var_213_cast_fp16))[name = tensor("op_425_cast_fp16")]; + tensor var_426_to_fp16 = const()[name = tensor("op_426_to_fp16"), val = tensor(0x1p-3)]; + tensor aw_chunk_1_cast_fp16 = mul(x = var_425_cast_fp16, y = var_426_to_fp16)[name = tensor("aw_chunk_1_cast_fp16")]; + tensor var_429_equation_0 = const()[name = tensor("op_429_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor var_429_cast_fp16 = einsum(equation = var_429_equation_0, values = (var_379_cast_fp16, var_220_cast_fp16))[name = tensor("op_429_cast_fp16")]; + tensor var_430_to_fp16 = const()[name = tensor("op_430_to_fp16"), val = tensor(0x1p-3)]; + tensor aw_chunk_3_cast_fp16 = mul(x = var_429_cast_fp16, y = var_430_to_fp16)[name = tensor("aw_chunk_3_cast_fp16")]; + tensor var_433_equation_0 = const()[name = tensor("op_433_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor var_433_cast_fp16 = einsum(equation = var_433_equation_0, values = (var_379_cast_fp16, var_227_cast_fp16))[name = tensor("op_433_cast_fp16")]; + tensor var_434_to_fp16 = const()[name = tensor("op_434_to_fp16"), val = tensor(0x1p-3)]; + tensor aw_chunk_5_cast_fp16 = mul(x = var_433_cast_fp16, y = var_434_to_fp16)[name = tensor("aw_chunk_5_cast_fp16")]; + tensor var_437_equation_0 = const()[name = tensor("op_437_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor var_437_cast_fp16 = einsum(equation = var_437_equation_0, values = (var_379_cast_fp16, var_234_cast_fp16))[name = tensor("op_437_cast_fp16")]; + tensor var_438_to_fp16 = const()[name = tensor("op_438_to_fp16"), val = tensor(0x1p-3)]; + tensor aw_chunk_7_cast_fp16 = mul(x = var_437_cast_fp16, y = var_438_to_fp16)[name = tensor("aw_chunk_7_cast_fp16")]; + tensor var_441_equation_0 = const()[name = tensor("op_441_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor var_441_cast_fp16 = einsum(equation = var_441_equation_0, values = (var_383_cast_fp16, var_241_cast_fp16))[name = tensor("op_441_cast_fp16")]; + tensor var_442_to_fp16 = const()[name = tensor("op_442_to_fp16"), val = tensor(0x1p-3)]; + tensor aw_chunk_9_cast_fp16 = mul(x = var_441_cast_fp16, y = var_442_to_fp16)[name = tensor("aw_chunk_9_cast_fp16")]; + tensor var_445_equation_0 = const()[name = tensor("op_445_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor var_445_cast_fp16 = einsum(equation = var_445_equation_0, values = (var_383_cast_fp16, var_248_cast_fp16))[name = tensor("op_445_cast_fp16")]; + tensor var_446_to_fp16 = const()[name = tensor("op_446_to_fp16"), val = tensor(0x1p-3)]; + tensor aw_chunk_11_cast_fp16 = mul(x = var_445_cast_fp16, y = var_446_to_fp16)[name = tensor("aw_chunk_11_cast_fp16")]; + tensor var_449_equation_0 = const()[name = tensor("op_449_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor var_449_cast_fp16 = einsum(equation = var_449_equation_0, values = (var_383_cast_fp16, var_255_cast_fp16))[name = tensor("op_449_cast_fp16")]; + tensor var_450_to_fp16 = const()[name = tensor("op_450_to_fp16"), val = tensor(0x1p-3)]; + tensor aw_chunk_13_cast_fp16 = mul(x = var_449_cast_fp16, y = var_450_to_fp16)[name = tensor("aw_chunk_13_cast_fp16")]; + tensor var_453_equation_0 = const()[name = tensor("op_453_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor var_453_cast_fp16 = einsum(equation = var_453_equation_0, values = (var_383_cast_fp16, var_262_cast_fp16))[name = tensor("op_453_cast_fp16")]; + tensor var_454_to_fp16 = const()[name = tensor("op_454_to_fp16"), val = tensor(0x1p-3)]; + tensor aw_chunk_15_cast_fp16 = mul(x = var_453_cast_fp16, y = var_454_to_fp16)[name = tensor("aw_chunk_15_cast_fp16")]; + tensor var_457_equation_0 = const()[name = tensor("op_457_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor var_457_cast_fp16 = einsum(equation = var_457_equation_0, values = (var_387_cast_fp16, var_269_cast_fp16))[name = tensor("op_457_cast_fp16")]; + tensor var_458_to_fp16 = const()[name = tensor("op_458_to_fp16"), val = tensor(0x1p-3)]; + tensor aw_chunk_17_cast_fp16 = mul(x = var_457_cast_fp16, y = var_458_to_fp16)[name = tensor("aw_chunk_17_cast_fp16")]; + tensor var_461_equation_0 = const()[name = tensor("op_461_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor var_461_cast_fp16 = einsum(equation = var_461_equation_0, values = (var_387_cast_fp16, var_276_cast_fp16))[name = tensor("op_461_cast_fp16")]; + tensor var_462_to_fp16 = const()[name = tensor("op_462_to_fp16"), val = tensor(0x1p-3)]; + tensor aw_chunk_19_cast_fp16 = mul(x = var_461_cast_fp16, y = var_462_to_fp16)[name = tensor("aw_chunk_19_cast_fp16")]; + tensor var_465_equation_0 = const()[name = tensor("op_465_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor var_465_cast_fp16 = einsum(equation = var_465_equation_0, values = (var_387_cast_fp16, var_283_cast_fp16))[name = tensor("op_465_cast_fp16")]; + tensor var_466_to_fp16 = const()[name = tensor("op_466_to_fp16"), val = tensor(0x1p-3)]; + tensor aw_chunk_21_cast_fp16 = mul(x = var_465_cast_fp16, y = var_466_to_fp16)[name = tensor("aw_chunk_21_cast_fp16")]; + tensor var_469_equation_0 = const()[name = tensor("op_469_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor var_469_cast_fp16 = einsum(equation = var_469_equation_0, values = (var_387_cast_fp16, var_290_cast_fp16))[name = tensor("op_469_cast_fp16")]; + tensor var_470_to_fp16 = const()[name = tensor("op_470_to_fp16"), val = tensor(0x1p-3)]; + tensor aw_chunk_23_cast_fp16 = mul(x = var_469_cast_fp16, y = var_470_to_fp16)[name = tensor("aw_chunk_23_cast_fp16")]; + tensor var_473_equation_0 = const()[name = tensor("op_473_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor var_473_cast_fp16 = einsum(equation = var_473_equation_0, values = (var_391_cast_fp16, var_297_cast_fp16))[name = tensor("op_473_cast_fp16")]; + tensor var_474_to_fp16 = const()[name = tensor("op_474_to_fp16"), val = tensor(0x1p-3)]; + tensor aw_chunk_25_cast_fp16 = mul(x = var_473_cast_fp16, y = var_474_to_fp16)[name = tensor("aw_chunk_25_cast_fp16")]; + tensor var_477_equation_0 = const()[name = tensor("op_477_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor var_477_cast_fp16 = einsum(equation = var_477_equation_0, values = (var_391_cast_fp16, var_304_cast_fp16))[name = tensor("op_477_cast_fp16")]; + tensor var_478_to_fp16 = const()[name = tensor("op_478_to_fp16"), val = tensor(0x1p-3)]; + tensor aw_chunk_27_cast_fp16 = mul(x = var_477_cast_fp16, y = var_478_to_fp16)[name = tensor("aw_chunk_27_cast_fp16")]; + tensor var_481_equation_0 = const()[name = tensor("op_481_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor var_481_cast_fp16 = einsum(equation = var_481_equation_0, values = (var_391_cast_fp16, var_311_cast_fp16))[name = tensor("op_481_cast_fp16")]; + tensor var_482_to_fp16 = const()[name = tensor("op_482_to_fp16"), val = tensor(0x1p-3)]; + tensor aw_chunk_29_cast_fp16 = mul(x = var_481_cast_fp16, y = var_482_to_fp16)[name = tensor("aw_chunk_29_cast_fp16")]; + tensor var_485_equation_0 = const()[name = tensor("op_485_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor var_485_cast_fp16 = einsum(equation = var_485_equation_0, values = (var_391_cast_fp16, var_318_cast_fp16))[name = tensor("op_485_cast_fp16")]; + tensor var_486_to_fp16 = const()[name = tensor("op_486_to_fp16"), val = tensor(0x1p-3)]; + tensor aw_chunk_31_cast_fp16 = mul(x = var_485_cast_fp16, y = var_486_to_fp16)[name = tensor("aw_chunk_31_cast_fp16")]; + tensor var_489_equation_0 = const()[name = tensor("op_489_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor var_489_cast_fp16 = einsum(equation = var_489_equation_0, values = (var_395_cast_fp16, var_325_cast_fp16))[name = tensor("op_489_cast_fp16")]; + tensor var_490_to_fp16 = const()[name = tensor("op_490_to_fp16"), val = tensor(0x1p-3)]; + tensor aw_chunk_33_cast_fp16 = mul(x = var_489_cast_fp16, y = var_490_to_fp16)[name = tensor("aw_chunk_33_cast_fp16")]; + tensor var_493_equation_0 = const()[name = tensor("op_493_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor var_493_cast_fp16 = einsum(equation = var_493_equation_0, values = (var_395_cast_fp16, var_332_cast_fp16))[name = tensor("op_493_cast_fp16")]; + tensor var_494_to_fp16 = const()[name = tensor("op_494_to_fp16"), val = tensor(0x1p-3)]; + tensor aw_chunk_35_cast_fp16 = mul(x = var_493_cast_fp16, y = var_494_to_fp16)[name = tensor("aw_chunk_35_cast_fp16")]; + tensor var_497_equation_0 = const()[name = tensor("op_497_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor var_497_cast_fp16 = einsum(equation = var_497_equation_0, values = (var_395_cast_fp16, var_339_cast_fp16))[name = tensor("op_497_cast_fp16")]; + tensor var_498_to_fp16 = const()[name = tensor("op_498_to_fp16"), val = tensor(0x1p-3)]; + tensor aw_chunk_37_cast_fp16 = mul(x = var_497_cast_fp16, y = var_498_to_fp16)[name = tensor("aw_chunk_37_cast_fp16")]; + tensor var_501_equation_0 = const()[name = tensor("op_501_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor var_501_cast_fp16 = einsum(equation = var_501_equation_0, values = (var_395_cast_fp16, var_346_cast_fp16))[name = tensor("op_501_cast_fp16")]; + tensor var_502_to_fp16 = const()[name = tensor("op_502_to_fp16"), val = tensor(0x1p-3)]; + tensor aw_chunk_39_cast_fp16 = mul(x = var_501_cast_fp16, y = var_502_to_fp16)[name = tensor("aw_chunk_39_cast_fp16")]; + tensor var_505_equation_0 = const()[name = tensor("op_505_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor var_505_cast_fp16 = einsum(equation = var_505_equation_0, values = (var_399_cast_fp16, var_353_cast_fp16))[name = tensor("op_505_cast_fp16")]; + tensor var_506_to_fp16 = const()[name = tensor("op_506_to_fp16"), val = tensor(0x1p-3)]; + tensor aw_chunk_41_cast_fp16 = mul(x = var_505_cast_fp16, y = var_506_to_fp16)[name = tensor("aw_chunk_41_cast_fp16")]; + tensor var_509_equation_0 = const()[name = tensor("op_509_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor var_509_cast_fp16 = einsum(equation = var_509_equation_0, values = (var_399_cast_fp16, var_360_cast_fp16))[name = tensor("op_509_cast_fp16")]; + tensor var_510_to_fp16 = const()[name = tensor("op_510_to_fp16"), val = tensor(0x1p-3)]; + tensor aw_chunk_43_cast_fp16 = mul(x = var_509_cast_fp16, y = var_510_to_fp16)[name = tensor("aw_chunk_43_cast_fp16")]; + tensor var_513_equation_0 = const()[name = tensor("op_513_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor var_513_cast_fp16 = einsum(equation = var_513_equation_0, values = (var_399_cast_fp16, var_367_cast_fp16))[name = tensor("op_513_cast_fp16")]; + tensor var_514_to_fp16 = const()[name = tensor("op_514_to_fp16"), val = tensor(0x1p-3)]; + tensor aw_chunk_45_cast_fp16 = mul(x = var_513_cast_fp16, y = var_514_to_fp16)[name = tensor("aw_chunk_45_cast_fp16")]; + tensor var_517_equation_0 = const()[name = tensor("op_517_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor var_517_cast_fp16 = einsum(equation = var_517_equation_0, values = (var_399_cast_fp16, var_374_cast_fp16))[name = tensor("op_517_cast_fp16")]; + tensor var_518_to_fp16 = const()[name = tensor("op_518_to_fp16"), val = tensor(0x1p-3)]; + tensor aw_chunk_47_cast_fp16 = mul(x = var_517_cast_fp16, y = var_518_to_fp16)[name = tensor("aw_chunk_47_cast_fp16")]; + tensor var_520_cast_fp16 = softmax(axis = var_129, x = aw_chunk_1_cast_fp16)[name = tensor("op_520_cast_fp16")]; + tensor var_521_cast_fp16 = softmax(axis = var_129, x = aw_chunk_3_cast_fp16)[name = tensor("op_521_cast_fp16")]; + tensor var_522_cast_fp16 = softmax(axis = var_129, x = aw_chunk_5_cast_fp16)[name = tensor("op_522_cast_fp16")]; + tensor var_523_cast_fp16 = softmax(axis = var_129, x = aw_chunk_7_cast_fp16)[name = tensor("op_523_cast_fp16")]; + tensor var_524_cast_fp16 = softmax(axis = var_129, x = aw_chunk_9_cast_fp16)[name = tensor("op_524_cast_fp16")]; + tensor var_525_cast_fp16 = softmax(axis = var_129, x = aw_chunk_11_cast_fp16)[name = tensor("op_525_cast_fp16")]; + tensor var_526_cast_fp16 = softmax(axis = var_129, x = aw_chunk_13_cast_fp16)[name = tensor("op_526_cast_fp16")]; + tensor var_527_cast_fp16 = softmax(axis = var_129, x = aw_chunk_15_cast_fp16)[name = tensor("op_527_cast_fp16")]; + tensor var_528_cast_fp16 = softmax(axis = var_129, x = aw_chunk_17_cast_fp16)[name = tensor("op_528_cast_fp16")]; + tensor var_529_cast_fp16 = softmax(axis = var_129, x = aw_chunk_19_cast_fp16)[name = tensor("op_529_cast_fp16")]; + tensor var_530_cast_fp16 = softmax(axis = var_129, x = aw_chunk_21_cast_fp16)[name = tensor("op_530_cast_fp16")]; + tensor var_531_cast_fp16 = softmax(axis = var_129, x = aw_chunk_23_cast_fp16)[name = tensor("op_531_cast_fp16")]; + tensor var_532_cast_fp16 = softmax(axis = var_129, x = aw_chunk_25_cast_fp16)[name = tensor("op_532_cast_fp16")]; + tensor var_533_cast_fp16 = softmax(axis = var_129, x = aw_chunk_27_cast_fp16)[name = tensor("op_533_cast_fp16")]; + tensor var_534_cast_fp16 = softmax(axis = var_129, x = aw_chunk_29_cast_fp16)[name = tensor("op_534_cast_fp16")]; + tensor var_535_cast_fp16 = softmax(axis = var_129, x = aw_chunk_31_cast_fp16)[name = tensor("op_535_cast_fp16")]; + tensor var_536_cast_fp16 = softmax(axis = var_129, x = aw_chunk_33_cast_fp16)[name = tensor("op_536_cast_fp16")]; + tensor var_537_cast_fp16 = softmax(axis = var_129, x = aw_chunk_35_cast_fp16)[name = tensor("op_537_cast_fp16")]; + tensor var_538_cast_fp16 = softmax(axis = var_129, x = aw_chunk_37_cast_fp16)[name = tensor("op_538_cast_fp16")]; + tensor var_539_cast_fp16 = softmax(axis = var_129, x = aw_chunk_39_cast_fp16)[name = tensor("op_539_cast_fp16")]; + tensor var_540_cast_fp16 = softmax(axis = var_129, x = aw_chunk_41_cast_fp16)[name = tensor("op_540_cast_fp16")]; + tensor var_541_cast_fp16 = softmax(axis = var_129, x = aw_chunk_43_cast_fp16)[name = tensor("op_541_cast_fp16")]; + tensor var_542_cast_fp16 = softmax(axis = var_129, x = aw_chunk_45_cast_fp16)[name = tensor("op_542_cast_fp16")]; + tensor var_543_cast_fp16 = softmax(axis = var_129, x = aw_chunk_47_cast_fp16)[name = tensor("op_543_cast_fp16")]; + tensor var_545_equation_0 = const()[name = tensor("op_545_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_545_cast_fp16 = einsum(equation = var_545_equation_0, values = (var_401_cast_fp16, var_520_cast_fp16))[name = tensor("op_545_cast_fp16")]; + tensor var_547_equation_0 = const()[name = tensor("op_547_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_547_cast_fp16 = einsum(equation = var_547_equation_0, values = (var_401_cast_fp16, var_521_cast_fp16))[name = tensor("op_547_cast_fp16")]; + tensor var_549_equation_0 = const()[name = tensor("op_549_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_549_cast_fp16 = einsum(equation = var_549_equation_0, values = (var_401_cast_fp16, var_522_cast_fp16))[name = tensor("op_549_cast_fp16")]; + tensor var_551_equation_0 = const()[name = tensor("op_551_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_551_cast_fp16 = einsum(equation = var_551_equation_0, values = (var_401_cast_fp16, var_523_cast_fp16))[name = tensor("op_551_cast_fp16")]; + tensor var_553_equation_0 = const()[name = tensor("op_553_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_553_cast_fp16 = einsum(equation = var_553_equation_0, values = (var_405_cast_fp16, var_524_cast_fp16))[name = tensor("op_553_cast_fp16")]; + tensor var_555_equation_0 = const()[name = tensor("op_555_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_555_cast_fp16 = einsum(equation = var_555_equation_0, values = (var_405_cast_fp16, var_525_cast_fp16))[name = tensor("op_555_cast_fp16")]; + tensor var_557_equation_0 = const()[name = tensor("op_557_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_557_cast_fp16 = einsum(equation = var_557_equation_0, values = (var_405_cast_fp16, var_526_cast_fp16))[name = tensor("op_557_cast_fp16")]; + tensor var_559_equation_0 = const()[name = tensor("op_559_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_559_cast_fp16 = einsum(equation = var_559_equation_0, values = (var_405_cast_fp16, var_527_cast_fp16))[name = tensor("op_559_cast_fp16")]; + tensor var_561_equation_0 = const()[name = tensor("op_561_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_561_cast_fp16 = einsum(equation = var_561_equation_0, values = (var_409_cast_fp16, var_528_cast_fp16))[name = tensor("op_561_cast_fp16")]; + tensor var_563_equation_0 = const()[name = tensor("op_563_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_563_cast_fp16 = einsum(equation = var_563_equation_0, values = (var_409_cast_fp16, var_529_cast_fp16))[name = tensor("op_563_cast_fp16")]; + tensor var_565_equation_0 = const()[name = tensor("op_565_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_565_cast_fp16 = einsum(equation = var_565_equation_0, values = (var_409_cast_fp16, var_530_cast_fp16))[name = tensor("op_565_cast_fp16")]; + tensor var_567_equation_0 = const()[name = tensor("op_567_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_567_cast_fp16 = einsum(equation = var_567_equation_0, values = (var_409_cast_fp16, var_531_cast_fp16))[name = tensor("op_567_cast_fp16")]; + tensor var_569_equation_0 = const()[name = tensor("op_569_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_569_cast_fp16 = einsum(equation = var_569_equation_0, values = (var_413_cast_fp16, var_532_cast_fp16))[name = tensor("op_569_cast_fp16")]; + tensor var_571_equation_0 = const()[name = tensor("op_571_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_571_cast_fp16 = einsum(equation = var_571_equation_0, values = (var_413_cast_fp16, var_533_cast_fp16))[name = tensor("op_571_cast_fp16")]; + tensor var_573_equation_0 = const()[name = tensor("op_573_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_573_cast_fp16 = einsum(equation = var_573_equation_0, values = (var_413_cast_fp16, var_534_cast_fp16))[name = tensor("op_573_cast_fp16")]; + tensor var_575_equation_0 = const()[name = tensor("op_575_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_575_cast_fp16 = einsum(equation = var_575_equation_0, values = (var_413_cast_fp16, var_535_cast_fp16))[name = tensor("op_575_cast_fp16")]; + tensor var_577_equation_0 = const()[name = tensor("op_577_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_577_cast_fp16 = einsum(equation = var_577_equation_0, values = (var_417_cast_fp16, var_536_cast_fp16))[name = tensor("op_577_cast_fp16")]; + tensor var_579_equation_0 = const()[name = tensor("op_579_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_579_cast_fp16 = einsum(equation = var_579_equation_0, values = (var_417_cast_fp16, var_537_cast_fp16))[name = tensor("op_579_cast_fp16")]; + tensor var_581_equation_0 = const()[name = tensor("op_581_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_581_cast_fp16 = einsum(equation = var_581_equation_0, values = (var_417_cast_fp16, var_538_cast_fp16))[name = tensor("op_581_cast_fp16")]; + tensor var_583_equation_0 = const()[name = tensor("op_583_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_583_cast_fp16 = einsum(equation = var_583_equation_0, values = (var_417_cast_fp16, var_539_cast_fp16))[name = tensor("op_583_cast_fp16")]; + tensor var_585_equation_0 = const()[name = tensor("op_585_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_585_cast_fp16 = einsum(equation = var_585_equation_0, values = (var_421_cast_fp16, var_540_cast_fp16))[name = tensor("op_585_cast_fp16")]; + tensor var_587_equation_0 = const()[name = tensor("op_587_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_587_cast_fp16 = einsum(equation = var_587_equation_0, values = (var_421_cast_fp16, var_541_cast_fp16))[name = tensor("op_587_cast_fp16")]; + tensor var_589_equation_0 = const()[name = tensor("op_589_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_589_cast_fp16 = einsum(equation = var_589_equation_0, values = (var_421_cast_fp16, var_542_cast_fp16))[name = tensor("op_589_cast_fp16")]; + tensor var_591_equation_0 = const()[name = tensor("op_591_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_591_cast_fp16 = einsum(equation = var_591_equation_0, values = (var_421_cast_fp16, var_543_cast_fp16))[name = tensor("op_591_cast_fp16")]; + tensor var_593_interleave_0 = const()[name = tensor("op_593_interleave_0"), val = tensor(false)]; + tensor var_593_cast_fp16 = concat(axis = var_118, interleave = var_593_interleave_0, values = (var_545_cast_fp16, var_547_cast_fp16, var_549_cast_fp16, var_551_cast_fp16))[name = tensor("op_593_cast_fp16")]; + tensor var_595_interleave_0 = const()[name = tensor("op_595_interleave_0"), val = tensor(false)]; + tensor var_595_cast_fp16 = concat(axis = var_118, interleave = var_595_interleave_0, values = (var_553_cast_fp16, var_555_cast_fp16, var_557_cast_fp16, var_559_cast_fp16))[name = tensor("op_595_cast_fp16")]; + tensor var_597_interleave_0 = const()[name = tensor("op_597_interleave_0"), val = tensor(false)]; + tensor var_597_cast_fp16 = concat(axis = var_118, interleave = var_597_interleave_0, values = (var_561_cast_fp16, var_563_cast_fp16, var_565_cast_fp16, var_567_cast_fp16))[name = tensor("op_597_cast_fp16")]; + tensor var_599_interleave_0 = const()[name = tensor("op_599_interleave_0"), val = tensor(false)]; + tensor var_599_cast_fp16 = concat(axis = var_118, interleave = var_599_interleave_0, values = (var_569_cast_fp16, var_571_cast_fp16, var_573_cast_fp16, var_575_cast_fp16))[name = tensor("op_599_cast_fp16")]; + tensor var_601_interleave_0 = const()[name = tensor("op_601_interleave_0"), val = tensor(false)]; + tensor var_601_cast_fp16 = concat(axis = var_118, interleave = var_601_interleave_0, values = (var_577_cast_fp16, var_579_cast_fp16, var_581_cast_fp16, var_583_cast_fp16))[name = tensor("op_601_cast_fp16")]; + tensor var_603_interleave_0 = const()[name = tensor("op_603_interleave_0"), val = tensor(false)]; + tensor var_603_cast_fp16 = concat(axis = var_118, interleave = var_603_interleave_0, values = (var_585_cast_fp16, var_587_cast_fp16, var_589_cast_fp16, var_591_cast_fp16))[name = tensor("op_603_cast_fp16")]; + tensor input_1_interleave_0 = const()[name = tensor("input_1_interleave_0"), val = tensor(false)]; + tensor input_1_cast_fp16 = concat(axis = var_129, interleave = input_1_interleave_0, values = (var_593_cast_fp16, var_595_cast_fp16, var_597_cast_fp16, var_599_cast_fp16, var_601_cast_fp16, var_603_cast_fp16))[name = tensor("input_1_cast_fp16")]; + tensor var_608 = const()[name = tensor("op_608"), val = tensor([1, 1])]; + tensor var_610 = const()[name = tensor("op_610"), val = tensor([1, 1])]; + tensor obj_3_pad_type_0 = const()[name = tensor("obj_3_pad_type_0"), val = tensor("custom")]; + tensor obj_3_pad_0 = const()[name = tensor("obj_3_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_0_self_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_0_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3112896)))]; + tensor layers_0_self_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_0_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3407872)))]; + tensor obj_3_cast_fp16 = conv(bias = layers_0_self_attn_o_proj_bias_to_fp16, dilations = var_610, groups = var_129, pad = obj_3_pad_0, pad_type = obj_3_pad_type_0, strides = var_608, weight = layers_0_self_attn_o_proj_weight_to_fp16, x = input_1_cast_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_616 = const()[name = tensor("op_616"), val = tensor([1])]; + tensor channels_mean_3_cast_fp16 = reduce_mean(axes = var_616, keep_dims = var_130, 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_620 = const()[name = tensor("op_620"), val = tensor([1])]; + tensor var_621_cast_fp16 = reduce_mean(axes = var_620, keep_dims = var_130, x = zero_mean_sq_3_cast_fp16)[name = tensor("op_621_cast_fp16")]; + tensor var_622_to_fp16 = const()[name = tensor("op_622_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_623_cast_fp16 = add(x = var_621_cast_fp16, y = var_622_to_fp16)[name = tensor("op_623_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_623_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 input_3_gamma_0_to_fp16 = const()[name = tensor("input_3_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3408704)))]; + tensor input_3_beta_0_to_fp16 = const()[name = tensor("input_3_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3409536)))]; + tensor input_3_epsilon_0_to_fp16 = const()[name = tensor("input_3_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_3_cast_fp16 = batch_norm(beta = input_3_beta_0_to_fp16, epsilon = input_3_epsilon_0_to_fp16, gamma = input_3_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("input_3_cast_fp16")]; + tensor var_634 = const()[name = tensor("op_634"), val = tensor([1, 1])]; + tensor var_636 = const()[name = tensor("op_636"), val = tensor([1, 1])]; + tensor input_5_pad_type_0 = const()[name = tensor("input_5_pad_type_0"), val = tensor("custom")]; + tensor input_5_pad_0 = const()[name = tensor("input_5_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_0_fc1_weight_to_fp16 = const()[name = tensor("layers_0_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3410368)))]; + tensor layers_0_fc1_bias_to_fp16 = const()[name = tensor("layers_0_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4590080)))]; + tensor input_5_cast_fp16 = conv(bias = layers_0_fc1_bias_to_fp16, dilations = var_636, groups = var_129, pad = input_5_pad_0, pad_type = input_5_pad_type_0, strides = var_634, weight = layers_0_fc1_weight_to_fp16, x = input_3_cast_fp16)[name = tensor("input_5_cast_fp16")]; + tensor input_7_mode_0 = const()[name = tensor("input_7_mode_0"), val = tensor("EXACT")]; + tensor input_7_cast_fp16 = gelu(mode = input_7_mode_0, x = input_5_cast_fp16)[name = tensor("input_7_cast_fp16")]; + tensor var_642 = const()[name = tensor("op_642"), val = tensor([1, 1])]; + tensor var_644 = const()[name = tensor("op_644"), val = tensor([1, 1])]; + tensor hidden_states_5_pad_type_0 = const()[name = tensor("hidden_states_5_pad_type_0"), val = tensor("custom")]; + tensor hidden_states_5_pad_0 = const()[name = tensor("hidden_states_5_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_0_fc2_weight_to_fp16 = const()[name = tensor("layers_0_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4593216)))]; + tensor layers_0_fc2_bias_to_fp16 = const()[name = tensor("layers_0_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5772928)))]; + tensor hidden_states_5_cast_fp16 = conv(bias = layers_0_fc2_bias_to_fp16, dilations = var_644, groups = var_129, pad = hidden_states_5_pad_0, pad_type = hidden_states_5_pad_type_0, strides = var_642, weight = layers_0_fc2_weight_to_fp16, x = input_7_cast_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_651 = const()[name = tensor("op_651"), val = tensor(3)]; + tensor var_662 = const()[name = tensor("op_662"), val = tensor(1)]; + tensor var_663 = const()[name = tensor("op_663"), val = tensor(true)]; + tensor var_673 = const()[name = tensor("op_673"), val = tensor([1])]; + tensor channels_mean_5_cast_fp16 = reduce_mean(axes = var_673, keep_dims = var_663, 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_677 = const()[name = tensor("op_677"), val = tensor([1])]; + tensor var_678_cast_fp16 = reduce_mean(axes = var_677, keep_dims = var_663, x = zero_mean_sq_5_cast_fp16)[name = tensor("op_678_cast_fp16")]; + tensor var_679_to_fp16 = const()[name = tensor("op_679_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_680_cast_fp16 = add(x = var_678_cast_fp16, y = var_679_to_fp16)[name = tensor("op_680_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_680_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(5773760)))]; + 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(5774592)))]; + 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 var_695 = const()[name = tensor("op_695"), val = tensor([1, 1])]; + tensor var_697 = const()[name = tensor("op_697"), val = tensor([1, 1])]; + tensor query_3_pad_type_0 = const()[name = tensor("query_3_pad_type_0"), val = tensor("custom")]; + tensor query_3_pad_0 = const()[name = tensor("query_3_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_1_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_1_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5775424)))]; + tensor layers_1_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_1_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6070400)))]; + tensor query_3_cast_fp16 = conv(bias = layers_1_self_attn_q_proj_bias_to_fp16, dilations = var_697, groups = var_662, pad = query_3_pad_0, pad_type = query_3_pad_type_0, strides = var_695, weight = layers_1_self_attn_q_proj_weight_to_fp16, x = obj_5_cast_fp16)[name = tensor("query_3_cast_fp16")]; + tensor var_701 = const()[name = tensor("op_701"), val = tensor([1, 1])]; + tensor var_703 = const()[name = tensor("op_703"), val = tensor([1, 1])]; + tensor key_3_pad_type_0 = const()[name = tensor("key_3_pad_type_0"), val = tensor("custom")]; + tensor key_3_pad_0 = const()[name = tensor("key_3_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_1_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_1_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6071232)))]; + tensor key_3_cast_fp16 = conv(dilations = var_703, groups = var_662, pad = key_3_pad_0, pad_type = key_3_pad_type_0, strides = var_701, weight = layers_1_self_attn_k_proj_weight_to_fp16, x = obj_5_cast_fp16)[name = tensor("key_3_cast_fp16")]; + tensor var_708 = const()[name = tensor("op_708"), val = tensor([1, 1])]; + tensor var_710 = const()[name = tensor("op_710"), val = tensor([1, 1])]; + tensor value_3_pad_type_0 = const()[name = tensor("value_3_pad_type_0"), val = tensor("custom")]; + tensor value_3_pad_0 = const()[name = tensor("value_3_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_1_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_1_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6366208)))]; + tensor layers_1_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_1_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6661184)))]; + tensor value_3_cast_fp16 = conv(bias = layers_1_self_attn_v_proj_bias_to_fp16, dilations = var_710, groups = var_662, pad = value_3_pad_0, pad_type = value_3_pad_type_0, strides = var_708, weight = layers_1_self_attn_v_proj_weight_to_fp16, x = obj_5_cast_fp16)[name = tensor("value_3_cast_fp16")]; + tensor var_717_begin_0 = const()[name = tensor("op_717_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_717_end_0 = const()[name = tensor("op_717_end_0"), val = tensor([1, 64, 1, 1500])]; + tensor var_717_end_mask_0 = const()[name = tensor("op_717_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_717_cast_fp16 = slice_by_index(begin = var_717_begin_0, end = var_717_end_0, end_mask = var_717_end_mask_0, x = query_3_cast_fp16)[name = tensor("op_717_cast_fp16")]; + tensor var_721_begin_0 = const()[name = tensor("op_721_begin_0"), val = tensor([0, 64, 0, 0])]; + tensor var_721_end_0 = const()[name = tensor("op_721_end_0"), val = tensor([1, 128, 1, 1500])]; + tensor var_721_end_mask_0 = const()[name = tensor("op_721_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_721_cast_fp16 = slice_by_index(begin = var_721_begin_0, end = var_721_end_0, end_mask = var_721_end_mask_0, x = query_3_cast_fp16)[name = tensor("op_721_cast_fp16")]; + tensor var_725_begin_0 = const()[name = tensor("op_725_begin_0"), val = tensor([0, 128, 0, 0])]; + tensor var_725_end_0 = const()[name = tensor("op_725_end_0"), val = tensor([1, 192, 1, 1500])]; + tensor var_725_end_mask_0 = const()[name = tensor("op_725_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_725_cast_fp16 = slice_by_index(begin = var_725_begin_0, end = var_725_end_0, end_mask = var_725_end_mask_0, x = query_3_cast_fp16)[name = tensor("op_725_cast_fp16")]; + tensor var_729_begin_0 = const()[name = tensor("op_729_begin_0"), val = tensor([0, 192, 0, 0])]; + tensor var_729_end_0 = const()[name = tensor("op_729_end_0"), val = tensor([1, 256, 1, 1500])]; + tensor var_729_end_mask_0 = const()[name = tensor("op_729_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_729_cast_fp16 = slice_by_index(begin = var_729_begin_0, end = var_729_end_0, end_mask = var_729_end_mask_0, x = query_3_cast_fp16)[name = tensor("op_729_cast_fp16")]; + tensor var_733_begin_0 = const()[name = tensor("op_733_begin_0"), val = tensor([0, 256, 0, 0])]; + tensor var_733_end_0 = const()[name = tensor("op_733_end_0"), val = tensor([1, 320, 1, 1500])]; + tensor var_733_end_mask_0 = const()[name = tensor("op_733_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_733_cast_fp16 = slice_by_index(begin = var_733_begin_0, end = var_733_end_0, end_mask = var_733_end_mask_0, x = query_3_cast_fp16)[name = tensor("op_733_cast_fp16")]; + tensor var_737_begin_0 = const()[name = tensor("op_737_begin_0"), val = tensor([0, 320, 0, 0])]; + tensor var_737_end_0 = const()[name = tensor("op_737_end_0"), val = tensor([1, 384, 1, 1500])]; + tensor var_737_end_mask_0 = const()[name = tensor("op_737_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_737_cast_fp16 = slice_by_index(begin = var_737_begin_0, end = var_737_end_0, end_mask = var_737_end_mask_0, x = query_3_cast_fp16)[name = tensor("op_737_cast_fp16")]; + tensor var_746_begin_0 = const()[name = tensor("op_746_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_746_end_0 = const()[name = tensor("op_746_end_0"), val = tensor([1, 64, 1, 375])]; + tensor var_746_end_mask_0 = const()[name = tensor("op_746_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_746_cast_fp16 = slice_by_index(begin = var_746_begin_0, end = var_746_end_0, end_mask = var_746_end_mask_0, x = var_717_cast_fp16)[name = tensor("op_746_cast_fp16")]; + tensor var_753_begin_0 = const()[name = tensor("op_753_begin_0"), val = tensor([0, 0, 0, 375])]; + tensor var_753_end_0 = const()[name = tensor("op_753_end_0"), val = tensor([1, 64, 1, 750])]; + tensor var_753_end_mask_0 = const()[name = tensor("op_753_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_753_cast_fp16 = slice_by_index(begin = var_753_begin_0, end = var_753_end_0, end_mask = var_753_end_mask_0, x = var_717_cast_fp16)[name = tensor("op_753_cast_fp16")]; + tensor var_760_begin_0 = const()[name = tensor("op_760_begin_0"), val = tensor([0, 0, 0, 750])]; + tensor var_760_end_0 = const()[name = tensor("op_760_end_0"), val = tensor([1, 64, 1, 1125])]; + tensor var_760_end_mask_0 = const()[name = tensor("op_760_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_760_cast_fp16 = slice_by_index(begin = var_760_begin_0, end = var_760_end_0, end_mask = var_760_end_mask_0, x = var_717_cast_fp16)[name = tensor("op_760_cast_fp16")]; + tensor var_767_begin_0 = const()[name = tensor("op_767_begin_0"), val = tensor([0, 0, 0, 1125])]; + tensor var_767_end_0 = const()[name = tensor("op_767_end_0"), val = tensor([1, 64, 1, 1500])]; + tensor var_767_end_mask_0 = const()[name = tensor("op_767_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_767_cast_fp16 = slice_by_index(begin = var_767_begin_0, end = var_767_end_0, end_mask = var_767_end_mask_0, x = var_717_cast_fp16)[name = tensor("op_767_cast_fp16")]; + tensor var_774_begin_0 = const()[name = tensor("op_774_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_774_end_0 = const()[name = tensor("op_774_end_0"), val = tensor([1, 64, 1, 375])]; + tensor var_774_end_mask_0 = const()[name = tensor("op_774_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_774_cast_fp16 = slice_by_index(begin = var_774_begin_0, end = var_774_end_0, end_mask = var_774_end_mask_0, x = var_721_cast_fp16)[name = tensor("op_774_cast_fp16")]; + tensor var_781_begin_0 = const()[name = tensor("op_781_begin_0"), val = tensor([0, 0, 0, 375])]; + tensor var_781_end_0 = const()[name = tensor("op_781_end_0"), val = tensor([1, 64, 1, 750])]; + tensor var_781_end_mask_0 = const()[name = tensor("op_781_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_781_cast_fp16 = slice_by_index(begin = var_781_begin_0, end = var_781_end_0, end_mask = var_781_end_mask_0, x = var_721_cast_fp16)[name = tensor("op_781_cast_fp16")]; + tensor var_788_begin_0 = const()[name = tensor("op_788_begin_0"), val = tensor([0, 0, 0, 750])]; + tensor var_788_end_0 = const()[name = tensor("op_788_end_0"), val = tensor([1, 64, 1, 1125])]; + tensor var_788_end_mask_0 = const()[name = tensor("op_788_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_788_cast_fp16 = slice_by_index(begin = var_788_begin_0, end = var_788_end_0, end_mask = var_788_end_mask_0, x = var_721_cast_fp16)[name = tensor("op_788_cast_fp16")]; + tensor var_795_begin_0 = const()[name = tensor("op_795_begin_0"), val = tensor([0, 0, 0, 1125])]; + tensor var_795_end_0 = const()[name = tensor("op_795_end_0"), val = tensor([1, 64, 1, 1500])]; + tensor var_795_end_mask_0 = const()[name = tensor("op_795_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_795_cast_fp16 = slice_by_index(begin = var_795_begin_0, end = var_795_end_0, end_mask = var_795_end_mask_0, x = var_721_cast_fp16)[name = tensor("op_795_cast_fp16")]; + tensor var_802_begin_0 = const()[name = tensor("op_802_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_802_end_0 = const()[name = tensor("op_802_end_0"), val = tensor([1, 64, 1, 375])]; + tensor var_802_end_mask_0 = const()[name = tensor("op_802_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_802_cast_fp16 = slice_by_index(begin = var_802_begin_0, end = var_802_end_0, end_mask = var_802_end_mask_0, x = var_725_cast_fp16)[name = tensor("op_802_cast_fp16")]; + tensor var_809_begin_0 = const()[name = tensor("op_809_begin_0"), val = tensor([0, 0, 0, 375])]; + tensor var_809_end_0 = const()[name = tensor("op_809_end_0"), val = tensor([1, 64, 1, 750])]; + tensor var_809_end_mask_0 = const()[name = tensor("op_809_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_809_cast_fp16 = slice_by_index(begin = var_809_begin_0, end = var_809_end_0, end_mask = var_809_end_mask_0, x = var_725_cast_fp16)[name = tensor("op_809_cast_fp16")]; + tensor var_816_begin_0 = const()[name = tensor("op_816_begin_0"), val = tensor([0, 0, 0, 750])]; + tensor var_816_end_0 = const()[name = tensor("op_816_end_0"), val = tensor([1, 64, 1, 1125])]; + tensor var_816_end_mask_0 = const()[name = tensor("op_816_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_816_cast_fp16 = slice_by_index(begin = var_816_begin_0, end = var_816_end_0, end_mask = var_816_end_mask_0, x = var_725_cast_fp16)[name = tensor("op_816_cast_fp16")]; + tensor var_823_begin_0 = const()[name = tensor("op_823_begin_0"), val = tensor([0, 0, 0, 1125])]; + tensor var_823_end_0 = const()[name = tensor("op_823_end_0"), val = tensor([1, 64, 1, 1500])]; + tensor var_823_end_mask_0 = const()[name = tensor("op_823_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_823_cast_fp16 = slice_by_index(begin = var_823_begin_0, end = var_823_end_0, end_mask = var_823_end_mask_0, x = var_725_cast_fp16)[name = tensor("op_823_cast_fp16")]; + tensor var_830_begin_0 = const()[name = tensor("op_830_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_830_end_0 = const()[name = tensor("op_830_end_0"), val = tensor([1, 64, 1, 375])]; + tensor var_830_end_mask_0 = const()[name = tensor("op_830_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_830_cast_fp16 = slice_by_index(begin = var_830_begin_0, end = var_830_end_0, end_mask = var_830_end_mask_0, x = var_729_cast_fp16)[name = tensor("op_830_cast_fp16")]; + tensor var_837_begin_0 = const()[name = tensor("op_837_begin_0"), val = tensor([0, 0, 0, 375])]; + tensor var_837_end_0 = const()[name = tensor("op_837_end_0"), val = tensor([1, 64, 1, 750])]; + tensor var_837_end_mask_0 = const()[name = tensor("op_837_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_837_cast_fp16 = slice_by_index(begin = var_837_begin_0, end = var_837_end_0, end_mask = var_837_end_mask_0, x = var_729_cast_fp16)[name = tensor("op_837_cast_fp16")]; + tensor var_844_begin_0 = const()[name = tensor("op_844_begin_0"), val = tensor([0, 0, 0, 750])]; + tensor var_844_end_0 = const()[name = tensor("op_844_end_0"), val = tensor([1, 64, 1, 1125])]; + tensor var_844_end_mask_0 = const()[name = tensor("op_844_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_844_cast_fp16 = slice_by_index(begin = var_844_begin_0, end = var_844_end_0, end_mask = var_844_end_mask_0, x = var_729_cast_fp16)[name = tensor("op_844_cast_fp16")]; + tensor var_851_begin_0 = const()[name = tensor("op_851_begin_0"), val = tensor([0, 0, 0, 1125])]; + tensor var_851_end_0 = const()[name = tensor("op_851_end_0"), val = tensor([1, 64, 1, 1500])]; + tensor var_851_end_mask_0 = const()[name = tensor("op_851_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_851_cast_fp16 = slice_by_index(begin = var_851_begin_0, end = var_851_end_0, end_mask = var_851_end_mask_0, x = var_729_cast_fp16)[name = tensor("op_851_cast_fp16")]; + tensor var_858_begin_0 = const()[name = tensor("op_858_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_858_end_0 = const()[name = tensor("op_858_end_0"), val = tensor([1, 64, 1, 375])]; + tensor var_858_end_mask_0 = const()[name = tensor("op_858_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_858_cast_fp16 = slice_by_index(begin = var_858_begin_0, end = var_858_end_0, end_mask = var_858_end_mask_0, x = var_733_cast_fp16)[name = tensor("op_858_cast_fp16")]; + tensor var_865_begin_0 = const()[name = tensor("op_865_begin_0"), val = tensor([0, 0, 0, 375])]; + tensor var_865_end_0 = const()[name = tensor("op_865_end_0"), val = tensor([1, 64, 1, 750])]; + tensor var_865_end_mask_0 = const()[name = tensor("op_865_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_865_cast_fp16 = slice_by_index(begin = var_865_begin_0, end = var_865_end_0, end_mask = var_865_end_mask_0, x = var_733_cast_fp16)[name = tensor("op_865_cast_fp16")]; + tensor var_872_begin_0 = const()[name = tensor("op_872_begin_0"), val = tensor([0, 0, 0, 750])]; + tensor var_872_end_0 = const()[name = tensor("op_872_end_0"), val = tensor([1, 64, 1, 1125])]; + tensor var_872_end_mask_0 = const()[name = tensor("op_872_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_872_cast_fp16 = slice_by_index(begin = var_872_begin_0, end = var_872_end_0, end_mask = var_872_end_mask_0, x = var_733_cast_fp16)[name = tensor("op_872_cast_fp16")]; + tensor var_879_begin_0 = const()[name = tensor("op_879_begin_0"), val = tensor([0, 0, 0, 1125])]; + tensor var_879_end_0 = const()[name = tensor("op_879_end_0"), val = tensor([1, 64, 1, 1500])]; + tensor var_879_end_mask_0 = const()[name = tensor("op_879_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_879_cast_fp16 = slice_by_index(begin = var_879_begin_0, end = var_879_end_0, end_mask = var_879_end_mask_0, x = var_733_cast_fp16)[name = tensor("op_879_cast_fp16")]; + tensor var_886_begin_0 = const()[name = tensor("op_886_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_886_end_0 = const()[name = tensor("op_886_end_0"), val = tensor([1, 64, 1, 375])]; + tensor var_886_end_mask_0 = const()[name = tensor("op_886_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_886_cast_fp16 = slice_by_index(begin = var_886_begin_0, end = var_886_end_0, end_mask = var_886_end_mask_0, x = var_737_cast_fp16)[name = tensor("op_886_cast_fp16")]; + tensor var_893_begin_0 = const()[name = tensor("op_893_begin_0"), val = tensor([0, 0, 0, 375])]; + tensor var_893_end_0 = const()[name = tensor("op_893_end_0"), val = tensor([1, 64, 1, 750])]; + tensor var_893_end_mask_0 = const()[name = tensor("op_893_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_893_cast_fp16 = slice_by_index(begin = var_893_begin_0, end = var_893_end_0, end_mask = var_893_end_mask_0, x = var_737_cast_fp16)[name = tensor("op_893_cast_fp16")]; + tensor var_900_begin_0 = const()[name = tensor("op_900_begin_0"), val = tensor([0, 0, 0, 750])]; + tensor var_900_end_0 = const()[name = tensor("op_900_end_0"), val = tensor([1, 64, 1, 1125])]; + tensor var_900_end_mask_0 = const()[name = tensor("op_900_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_900_cast_fp16 = slice_by_index(begin = var_900_begin_0, end = var_900_end_0, end_mask = var_900_end_mask_0, x = var_737_cast_fp16)[name = tensor("op_900_cast_fp16")]; + tensor var_907_begin_0 = const()[name = tensor("op_907_begin_0"), val = tensor([0, 0, 0, 1125])]; + tensor var_907_end_0 = const()[name = tensor("op_907_end_0"), val = tensor([1, 64, 1, 1500])]; + tensor var_907_end_mask_0 = const()[name = tensor("op_907_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_907_cast_fp16 = slice_by_index(begin = var_907_begin_0, end = var_907_end_0, end_mask = var_907_end_mask_0, x = var_737_cast_fp16)[name = tensor("op_907_cast_fp16")]; + tensor k_3_perm_0 = const()[name = tensor("k_3_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor var_912_begin_0 = const()[name = tensor("op_912_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_912_end_0 = const()[name = tensor("op_912_end_0"), val = tensor([1, 1500, 1, 64])]; + tensor var_912_end_mask_0 = const()[name = tensor("op_912_end_mask_0"), val = tensor([true, true, true, false])]; + tensor transpose_2 = transpose(perm = k_3_perm_0, x = key_3_cast_fp16)[name = tensor("transpose_2")]; + tensor var_912_cast_fp16 = slice_by_index(begin = var_912_begin_0, end = var_912_end_0, end_mask = var_912_end_mask_0, x = transpose_2)[name = tensor("op_912_cast_fp16")]; + tensor var_916_begin_0 = const()[name = tensor("op_916_begin_0"), val = tensor([0, 0, 0, 64])]; + tensor var_916_end_0 = const()[name = tensor("op_916_end_0"), val = tensor([1, 1500, 1, 128])]; + tensor var_916_end_mask_0 = const()[name = tensor("op_916_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_916_cast_fp16 = slice_by_index(begin = var_916_begin_0, end = var_916_end_0, end_mask = var_916_end_mask_0, x = transpose_2)[name = tensor("op_916_cast_fp16")]; + tensor var_920_begin_0 = const()[name = tensor("op_920_begin_0"), val = tensor([0, 0, 0, 128])]; + tensor var_920_end_0 = const()[name = tensor("op_920_end_0"), val = tensor([1, 1500, 1, 192])]; + tensor var_920_end_mask_0 = const()[name = tensor("op_920_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_920_cast_fp16 = slice_by_index(begin = var_920_begin_0, end = var_920_end_0, end_mask = var_920_end_mask_0, x = transpose_2)[name = tensor("op_920_cast_fp16")]; + tensor var_924_begin_0 = const()[name = tensor("op_924_begin_0"), val = tensor([0, 0, 0, 192])]; + tensor var_924_end_0 = const()[name = tensor("op_924_end_0"), val = tensor([1, 1500, 1, 256])]; + tensor var_924_end_mask_0 = const()[name = tensor("op_924_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_924_cast_fp16 = slice_by_index(begin = var_924_begin_0, end = var_924_end_0, end_mask = var_924_end_mask_0, x = transpose_2)[name = tensor("op_924_cast_fp16")]; + tensor var_928_begin_0 = const()[name = tensor("op_928_begin_0"), val = tensor([0, 0, 0, 256])]; + tensor var_928_end_0 = const()[name = tensor("op_928_end_0"), val = tensor([1, 1500, 1, 320])]; + tensor var_928_end_mask_0 = const()[name = tensor("op_928_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_928_cast_fp16 = slice_by_index(begin = var_928_begin_0, end = var_928_end_0, end_mask = var_928_end_mask_0, x = transpose_2)[name = tensor("op_928_cast_fp16")]; + tensor var_932_begin_0 = const()[name = tensor("op_932_begin_0"), val = tensor([0, 0, 0, 320])]; + tensor var_932_end_0 = const()[name = tensor("op_932_end_0"), val = tensor([1, 1500, 1, 384])]; + tensor var_932_end_mask_0 = const()[name = tensor("op_932_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_932_cast_fp16 = slice_by_index(begin = var_932_begin_0, end = var_932_end_0, end_mask = var_932_end_mask_0, x = transpose_2)[name = tensor("op_932_cast_fp16")]; + tensor var_934_begin_0 = const()[name = tensor("op_934_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_934_end_0 = const()[name = tensor("op_934_end_0"), val = tensor([1, 64, 1, 1500])]; + tensor var_934_end_mask_0 = const()[name = tensor("op_934_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_934_cast_fp16 = slice_by_index(begin = var_934_begin_0, end = var_934_end_0, end_mask = var_934_end_mask_0, x = value_3_cast_fp16)[name = tensor("op_934_cast_fp16")]; + tensor var_938_begin_0 = const()[name = tensor("op_938_begin_0"), val = tensor([0, 64, 0, 0])]; + tensor var_938_end_0 = const()[name = tensor("op_938_end_0"), val = tensor([1, 128, 1, 1500])]; + tensor var_938_end_mask_0 = const()[name = tensor("op_938_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_938_cast_fp16 = slice_by_index(begin = var_938_begin_0, end = var_938_end_0, end_mask = var_938_end_mask_0, x = value_3_cast_fp16)[name = tensor("op_938_cast_fp16")]; + tensor var_942_begin_0 = const()[name = tensor("op_942_begin_0"), val = tensor([0, 128, 0, 0])]; + tensor var_942_end_0 = const()[name = tensor("op_942_end_0"), val = tensor([1, 192, 1, 1500])]; + tensor var_942_end_mask_0 = const()[name = tensor("op_942_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_942_cast_fp16 = slice_by_index(begin = var_942_begin_0, end = var_942_end_0, end_mask = var_942_end_mask_0, x = value_3_cast_fp16)[name = tensor("op_942_cast_fp16")]; + tensor var_946_begin_0 = const()[name = tensor("op_946_begin_0"), val = tensor([0, 192, 0, 0])]; + tensor var_946_end_0 = const()[name = tensor("op_946_end_0"), val = tensor([1, 256, 1, 1500])]; + tensor var_946_end_mask_0 = const()[name = tensor("op_946_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_946_cast_fp16 = slice_by_index(begin = var_946_begin_0, end = var_946_end_0, end_mask = var_946_end_mask_0, x = value_3_cast_fp16)[name = tensor("op_946_cast_fp16")]; + tensor var_950_begin_0 = const()[name = tensor("op_950_begin_0"), val = tensor([0, 256, 0, 0])]; + tensor var_950_end_0 = const()[name = tensor("op_950_end_0"), val = tensor([1, 320, 1, 1500])]; + tensor var_950_end_mask_0 = const()[name = tensor("op_950_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_950_cast_fp16 = slice_by_index(begin = var_950_begin_0, end = var_950_end_0, end_mask = var_950_end_mask_0, x = value_3_cast_fp16)[name = tensor("op_950_cast_fp16")]; + tensor var_954_begin_0 = const()[name = tensor("op_954_begin_0"), val = tensor([0, 320, 0, 0])]; + tensor var_954_end_0 = const()[name = tensor("op_954_end_0"), val = tensor([1, 384, 1, 1500])]; + tensor var_954_end_mask_0 = const()[name = tensor("op_954_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_954_cast_fp16 = slice_by_index(begin = var_954_begin_0, end = var_954_end_0, end_mask = var_954_end_mask_0, x = value_3_cast_fp16)[name = tensor("op_954_cast_fp16")]; + tensor var_958_equation_0 = const()[name = tensor("op_958_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor var_958_cast_fp16 = einsum(equation = var_958_equation_0, values = (var_912_cast_fp16, var_746_cast_fp16))[name = tensor("op_958_cast_fp16")]; + tensor var_959_to_fp16 = const()[name = tensor("op_959_to_fp16"), val = tensor(0x1p-3)]; + tensor aw_chunk_49_cast_fp16 = mul(x = var_958_cast_fp16, y = var_959_to_fp16)[name = tensor("aw_chunk_49_cast_fp16")]; + tensor var_962_equation_0 = const()[name = tensor("op_962_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor var_962_cast_fp16 = einsum(equation = var_962_equation_0, values = (var_912_cast_fp16, var_753_cast_fp16))[name = tensor("op_962_cast_fp16")]; + tensor var_963_to_fp16 = const()[name = tensor("op_963_to_fp16"), val = tensor(0x1p-3)]; + tensor aw_chunk_51_cast_fp16 = mul(x = var_962_cast_fp16, y = var_963_to_fp16)[name = tensor("aw_chunk_51_cast_fp16")]; + tensor var_966_equation_0 = const()[name = tensor("op_966_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor var_966_cast_fp16 = einsum(equation = var_966_equation_0, values = (var_912_cast_fp16, var_760_cast_fp16))[name = tensor("op_966_cast_fp16")]; + tensor var_967_to_fp16 = const()[name = tensor("op_967_to_fp16"), val = tensor(0x1p-3)]; + tensor aw_chunk_53_cast_fp16 = mul(x = var_966_cast_fp16, y = var_967_to_fp16)[name = tensor("aw_chunk_53_cast_fp16")]; + tensor var_970_equation_0 = const()[name = tensor("op_970_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor var_970_cast_fp16 = einsum(equation = var_970_equation_0, values = (var_912_cast_fp16, var_767_cast_fp16))[name = tensor("op_970_cast_fp16")]; + tensor var_971_to_fp16 = const()[name = tensor("op_971_to_fp16"), val = tensor(0x1p-3)]; + tensor aw_chunk_55_cast_fp16 = mul(x = var_970_cast_fp16, y = var_971_to_fp16)[name = tensor("aw_chunk_55_cast_fp16")]; + tensor var_974_equation_0 = const()[name = tensor("op_974_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor var_974_cast_fp16 = einsum(equation = var_974_equation_0, values = (var_916_cast_fp16, var_774_cast_fp16))[name = tensor("op_974_cast_fp16")]; + tensor var_975_to_fp16 = const()[name = tensor("op_975_to_fp16"), val = tensor(0x1p-3)]; + tensor aw_chunk_57_cast_fp16 = mul(x = var_974_cast_fp16, y = var_975_to_fp16)[name = tensor("aw_chunk_57_cast_fp16")]; + tensor var_978_equation_0 = const()[name = tensor("op_978_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor var_978_cast_fp16 = einsum(equation = var_978_equation_0, values = (var_916_cast_fp16, var_781_cast_fp16))[name = tensor("op_978_cast_fp16")]; + tensor var_979_to_fp16 = const()[name = tensor("op_979_to_fp16"), val = tensor(0x1p-3)]; + tensor aw_chunk_59_cast_fp16 = mul(x = var_978_cast_fp16, y = var_979_to_fp16)[name = tensor("aw_chunk_59_cast_fp16")]; + tensor var_982_equation_0 = const()[name = tensor("op_982_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor var_982_cast_fp16 = einsum(equation = var_982_equation_0, values = (var_916_cast_fp16, var_788_cast_fp16))[name = tensor("op_982_cast_fp16")]; + tensor var_983_to_fp16 = const()[name = tensor("op_983_to_fp16"), val = tensor(0x1p-3)]; + tensor aw_chunk_61_cast_fp16 = mul(x = var_982_cast_fp16, y = var_983_to_fp16)[name = tensor("aw_chunk_61_cast_fp16")]; + tensor var_986_equation_0 = const()[name = tensor("op_986_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor var_986_cast_fp16 = einsum(equation = var_986_equation_0, values = (var_916_cast_fp16, var_795_cast_fp16))[name = tensor("op_986_cast_fp16")]; + tensor var_987_to_fp16 = const()[name = tensor("op_987_to_fp16"), val = tensor(0x1p-3)]; + tensor aw_chunk_63_cast_fp16 = mul(x = var_986_cast_fp16, y = var_987_to_fp16)[name = tensor("aw_chunk_63_cast_fp16")]; + tensor var_990_equation_0 = const()[name = tensor("op_990_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor var_990_cast_fp16 = einsum(equation = var_990_equation_0, values = (var_920_cast_fp16, var_802_cast_fp16))[name = tensor("op_990_cast_fp16")]; + tensor var_991_to_fp16 = const()[name = tensor("op_991_to_fp16"), val = tensor(0x1p-3)]; + tensor aw_chunk_65_cast_fp16 = mul(x = var_990_cast_fp16, y = var_991_to_fp16)[name = tensor("aw_chunk_65_cast_fp16")]; + tensor var_994_equation_0 = const()[name = tensor("op_994_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor var_994_cast_fp16 = einsum(equation = var_994_equation_0, values = (var_920_cast_fp16, var_809_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 aw_chunk_67_cast_fp16 = mul(x = var_994_cast_fp16, y = var_995_to_fp16)[name = tensor("aw_chunk_67_cast_fp16")]; + tensor var_998_equation_0 = const()[name = tensor("op_998_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor var_998_cast_fp16 = einsum(equation = var_998_equation_0, values = (var_920_cast_fp16, var_816_cast_fp16))[name = tensor("op_998_cast_fp16")]; + tensor var_999_to_fp16 = const()[name = tensor("op_999_to_fp16"), val = tensor(0x1p-3)]; + tensor aw_chunk_69_cast_fp16 = mul(x = var_998_cast_fp16, y = var_999_to_fp16)[name = tensor("aw_chunk_69_cast_fp16")]; + tensor var_1002_equation_0 = const()[name = tensor("op_1002_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor var_1002_cast_fp16 = einsum(equation = var_1002_equation_0, values = (var_920_cast_fp16, var_823_cast_fp16))[name = tensor("op_1002_cast_fp16")]; + tensor var_1003_to_fp16 = const()[name = tensor("op_1003_to_fp16"), val = tensor(0x1p-3)]; + tensor aw_chunk_71_cast_fp16 = mul(x = var_1002_cast_fp16, y = var_1003_to_fp16)[name = tensor("aw_chunk_71_cast_fp16")]; + tensor var_1006_equation_0 = const()[name = tensor("op_1006_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor var_1006_cast_fp16 = einsum(equation = var_1006_equation_0, values = (var_924_cast_fp16, var_830_cast_fp16))[name = tensor("op_1006_cast_fp16")]; + tensor var_1007_to_fp16 = const()[name = tensor("op_1007_to_fp16"), val = tensor(0x1p-3)]; + tensor aw_chunk_73_cast_fp16 = mul(x = var_1006_cast_fp16, y = var_1007_to_fp16)[name = tensor("aw_chunk_73_cast_fp16")]; + tensor var_1010_equation_0 = const()[name = tensor("op_1010_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor var_1010_cast_fp16 = einsum(equation = var_1010_equation_0, values = (var_924_cast_fp16, var_837_cast_fp16))[name = tensor("op_1010_cast_fp16")]; + tensor var_1011_to_fp16 = const()[name = tensor("op_1011_to_fp16"), val = tensor(0x1p-3)]; + tensor aw_chunk_75_cast_fp16 = mul(x = var_1010_cast_fp16, y = var_1011_to_fp16)[name = tensor("aw_chunk_75_cast_fp16")]; + tensor var_1014_equation_0 = const()[name = tensor("op_1014_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor var_1014_cast_fp16 = einsum(equation = var_1014_equation_0, values = (var_924_cast_fp16, var_844_cast_fp16))[name = tensor("op_1014_cast_fp16")]; + tensor var_1015_to_fp16 = const()[name = tensor("op_1015_to_fp16"), val = tensor(0x1p-3)]; + tensor aw_chunk_77_cast_fp16 = mul(x = var_1014_cast_fp16, y = var_1015_to_fp16)[name = tensor("aw_chunk_77_cast_fp16")]; + tensor var_1018_equation_0 = const()[name = tensor("op_1018_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor var_1018_cast_fp16 = einsum(equation = var_1018_equation_0, values = (var_924_cast_fp16, var_851_cast_fp16))[name = tensor("op_1018_cast_fp16")]; + tensor var_1019_to_fp16 = const()[name = tensor("op_1019_to_fp16"), val = tensor(0x1p-3)]; + tensor aw_chunk_79_cast_fp16 = mul(x = var_1018_cast_fp16, y = var_1019_to_fp16)[name = tensor("aw_chunk_79_cast_fp16")]; + tensor var_1022_equation_0 = const()[name = tensor("op_1022_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor var_1022_cast_fp16 = einsum(equation = var_1022_equation_0, values = (var_928_cast_fp16, var_858_cast_fp16))[name = tensor("op_1022_cast_fp16")]; + tensor var_1023_to_fp16 = const()[name = tensor("op_1023_to_fp16"), val = tensor(0x1p-3)]; + tensor aw_chunk_81_cast_fp16 = mul(x = var_1022_cast_fp16, y = var_1023_to_fp16)[name = tensor("aw_chunk_81_cast_fp16")]; + tensor var_1026_equation_0 = const()[name = tensor("op_1026_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor var_1026_cast_fp16 = einsum(equation = var_1026_equation_0, values = (var_928_cast_fp16, var_865_cast_fp16))[name = tensor("op_1026_cast_fp16")]; + tensor var_1027_to_fp16 = const()[name = tensor("op_1027_to_fp16"), val = tensor(0x1p-3)]; + tensor aw_chunk_83_cast_fp16 = mul(x = var_1026_cast_fp16, y = var_1027_to_fp16)[name = tensor("aw_chunk_83_cast_fp16")]; + tensor var_1030_equation_0 = const()[name = tensor("op_1030_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor var_1030_cast_fp16 = einsum(equation = var_1030_equation_0, values = (var_928_cast_fp16, var_872_cast_fp16))[name = tensor("op_1030_cast_fp16")]; + tensor var_1031_to_fp16 = const()[name = tensor("op_1031_to_fp16"), val = tensor(0x1p-3)]; + tensor aw_chunk_85_cast_fp16 = mul(x = var_1030_cast_fp16, y = var_1031_to_fp16)[name = tensor("aw_chunk_85_cast_fp16")]; + tensor var_1034_equation_0 = const()[name = tensor("op_1034_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor var_1034_cast_fp16 = einsum(equation = var_1034_equation_0, values = (var_928_cast_fp16, var_879_cast_fp16))[name = tensor("op_1034_cast_fp16")]; + tensor var_1035_to_fp16 = const()[name = tensor("op_1035_to_fp16"), val = tensor(0x1p-3)]; + tensor aw_chunk_87_cast_fp16 = mul(x = var_1034_cast_fp16, y = var_1035_to_fp16)[name = tensor("aw_chunk_87_cast_fp16")]; + tensor var_1038_equation_0 = const()[name = tensor("op_1038_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor var_1038_cast_fp16 = einsum(equation = var_1038_equation_0, values = (var_932_cast_fp16, var_886_cast_fp16))[name = tensor("op_1038_cast_fp16")]; + tensor var_1039_to_fp16 = const()[name = tensor("op_1039_to_fp16"), val = tensor(0x1p-3)]; + tensor aw_chunk_89_cast_fp16 = mul(x = var_1038_cast_fp16, y = var_1039_to_fp16)[name = tensor("aw_chunk_89_cast_fp16")]; + tensor var_1042_equation_0 = const()[name = tensor("op_1042_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor var_1042_cast_fp16 = einsum(equation = var_1042_equation_0, values = (var_932_cast_fp16, var_893_cast_fp16))[name = tensor("op_1042_cast_fp16")]; + tensor var_1043_to_fp16 = const()[name = tensor("op_1043_to_fp16"), val = tensor(0x1p-3)]; + tensor aw_chunk_91_cast_fp16 = mul(x = var_1042_cast_fp16, y = var_1043_to_fp16)[name = tensor("aw_chunk_91_cast_fp16")]; + tensor var_1046_equation_0 = const()[name = tensor("op_1046_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor var_1046_cast_fp16 = einsum(equation = var_1046_equation_0, values = (var_932_cast_fp16, var_900_cast_fp16))[name = tensor("op_1046_cast_fp16")]; + tensor var_1047_to_fp16 = const()[name = tensor("op_1047_to_fp16"), val = tensor(0x1p-3)]; + tensor aw_chunk_93_cast_fp16 = mul(x = var_1046_cast_fp16, y = var_1047_to_fp16)[name = tensor("aw_chunk_93_cast_fp16")]; + tensor var_1050_equation_0 = const()[name = tensor("op_1050_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor var_1050_cast_fp16 = einsum(equation = var_1050_equation_0, values = (var_932_cast_fp16, var_907_cast_fp16))[name = tensor("op_1050_cast_fp16")]; + tensor var_1051_to_fp16 = const()[name = tensor("op_1051_to_fp16"), val = tensor(0x1p-3)]; + tensor aw_chunk_95_cast_fp16 = mul(x = var_1050_cast_fp16, y = var_1051_to_fp16)[name = tensor("aw_chunk_95_cast_fp16")]; + tensor var_1053_cast_fp16 = softmax(axis = var_662, x = aw_chunk_49_cast_fp16)[name = tensor("op_1053_cast_fp16")]; + tensor var_1054_cast_fp16 = softmax(axis = var_662, x = aw_chunk_51_cast_fp16)[name = tensor("op_1054_cast_fp16")]; + tensor var_1055_cast_fp16 = softmax(axis = var_662, x = aw_chunk_53_cast_fp16)[name = tensor("op_1055_cast_fp16")]; + tensor var_1056_cast_fp16 = softmax(axis = var_662, x = aw_chunk_55_cast_fp16)[name = tensor("op_1056_cast_fp16")]; + tensor var_1057_cast_fp16 = softmax(axis = var_662, x = aw_chunk_57_cast_fp16)[name = tensor("op_1057_cast_fp16")]; + tensor var_1058_cast_fp16 = softmax(axis = var_662, x = aw_chunk_59_cast_fp16)[name = tensor("op_1058_cast_fp16")]; + tensor var_1059_cast_fp16 = softmax(axis = var_662, x = aw_chunk_61_cast_fp16)[name = tensor("op_1059_cast_fp16")]; + tensor var_1060_cast_fp16 = softmax(axis = var_662, x = aw_chunk_63_cast_fp16)[name = tensor("op_1060_cast_fp16")]; + tensor var_1061_cast_fp16 = softmax(axis = var_662, x = aw_chunk_65_cast_fp16)[name = tensor("op_1061_cast_fp16")]; + tensor var_1062_cast_fp16 = softmax(axis = var_662, x = aw_chunk_67_cast_fp16)[name = tensor("op_1062_cast_fp16")]; + tensor var_1063_cast_fp16 = softmax(axis = var_662, x = aw_chunk_69_cast_fp16)[name = tensor("op_1063_cast_fp16")]; + tensor var_1064_cast_fp16 = softmax(axis = var_662, x = aw_chunk_71_cast_fp16)[name = tensor("op_1064_cast_fp16")]; + tensor var_1065_cast_fp16 = softmax(axis = var_662, x = aw_chunk_73_cast_fp16)[name = tensor("op_1065_cast_fp16")]; + tensor var_1066_cast_fp16 = softmax(axis = var_662, x = aw_chunk_75_cast_fp16)[name = tensor("op_1066_cast_fp16")]; + tensor var_1067_cast_fp16 = softmax(axis = var_662, x = aw_chunk_77_cast_fp16)[name = tensor("op_1067_cast_fp16")]; + tensor var_1068_cast_fp16 = softmax(axis = var_662, x = aw_chunk_79_cast_fp16)[name = tensor("op_1068_cast_fp16")]; + tensor var_1069_cast_fp16 = softmax(axis = var_662, x = aw_chunk_81_cast_fp16)[name = tensor("op_1069_cast_fp16")]; + tensor var_1070_cast_fp16 = softmax(axis = var_662, x = aw_chunk_83_cast_fp16)[name = tensor("op_1070_cast_fp16")]; + tensor var_1071_cast_fp16 = softmax(axis = var_662, x = aw_chunk_85_cast_fp16)[name = tensor("op_1071_cast_fp16")]; + tensor var_1072_cast_fp16 = softmax(axis = var_662, x = aw_chunk_87_cast_fp16)[name = tensor("op_1072_cast_fp16")]; + tensor var_1073_cast_fp16 = softmax(axis = var_662, x = aw_chunk_89_cast_fp16)[name = tensor("op_1073_cast_fp16")]; + tensor var_1074_cast_fp16 = softmax(axis = var_662, x = aw_chunk_91_cast_fp16)[name = tensor("op_1074_cast_fp16")]; + tensor var_1075_cast_fp16 = softmax(axis = var_662, x = aw_chunk_93_cast_fp16)[name = tensor("op_1075_cast_fp16")]; + tensor var_1076_cast_fp16 = softmax(axis = var_662, x = aw_chunk_95_cast_fp16)[name = tensor("op_1076_cast_fp16")]; + tensor var_1078_equation_0 = const()[name = tensor("op_1078_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1078_cast_fp16 = einsum(equation = var_1078_equation_0, values = (var_934_cast_fp16, var_1053_cast_fp16))[name = tensor("op_1078_cast_fp16")]; + tensor var_1080_equation_0 = const()[name = tensor("op_1080_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1080_cast_fp16 = einsum(equation = var_1080_equation_0, values = (var_934_cast_fp16, var_1054_cast_fp16))[name = tensor("op_1080_cast_fp16")]; + tensor var_1082_equation_0 = const()[name = tensor("op_1082_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1082_cast_fp16 = einsum(equation = var_1082_equation_0, values = (var_934_cast_fp16, var_1055_cast_fp16))[name = tensor("op_1082_cast_fp16")]; + tensor var_1084_equation_0 = const()[name = tensor("op_1084_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1084_cast_fp16 = einsum(equation = var_1084_equation_0, values = (var_934_cast_fp16, var_1056_cast_fp16))[name = tensor("op_1084_cast_fp16")]; + tensor var_1086_equation_0 = const()[name = tensor("op_1086_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1086_cast_fp16 = einsum(equation = var_1086_equation_0, values = (var_938_cast_fp16, var_1057_cast_fp16))[name = tensor("op_1086_cast_fp16")]; + tensor var_1088_equation_0 = const()[name = tensor("op_1088_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1088_cast_fp16 = einsum(equation = var_1088_equation_0, values = (var_938_cast_fp16, var_1058_cast_fp16))[name = tensor("op_1088_cast_fp16")]; + tensor var_1090_equation_0 = const()[name = tensor("op_1090_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1090_cast_fp16 = einsum(equation = var_1090_equation_0, values = (var_938_cast_fp16, var_1059_cast_fp16))[name = tensor("op_1090_cast_fp16")]; + tensor var_1092_equation_0 = const()[name = tensor("op_1092_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1092_cast_fp16 = einsum(equation = var_1092_equation_0, values = (var_938_cast_fp16, var_1060_cast_fp16))[name = tensor("op_1092_cast_fp16")]; + tensor var_1094_equation_0 = const()[name = tensor("op_1094_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1094_cast_fp16 = einsum(equation = var_1094_equation_0, values = (var_942_cast_fp16, var_1061_cast_fp16))[name = tensor("op_1094_cast_fp16")]; + tensor var_1096_equation_0 = const()[name = tensor("op_1096_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1096_cast_fp16 = einsum(equation = var_1096_equation_0, values = (var_942_cast_fp16, var_1062_cast_fp16))[name = tensor("op_1096_cast_fp16")]; + tensor var_1098_equation_0 = const()[name = tensor("op_1098_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1098_cast_fp16 = einsum(equation = var_1098_equation_0, values = (var_942_cast_fp16, var_1063_cast_fp16))[name = tensor("op_1098_cast_fp16")]; + tensor var_1100_equation_0 = const()[name = tensor("op_1100_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1100_cast_fp16 = einsum(equation = var_1100_equation_0, values = (var_942_cast_fp16, var_1064_cast_fp16))[name = tensor("op_1100_cast_fp16")]; + tensor var_1102_equation_0 = const()[name = tensor("op_1102_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1102_cast_fp16 = einsum(equation = var_1102_equation_0, values = (var_946_cast_fp16, var_1065_cast_fp16))[name = tensor("op_1102_cast_fp16")]; + tensor var_1104_equation_0 = const()[name = tensor("op_1104_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1104_cast_fp16 = einsum(equation = var_1104_equation_0, values = (var_946_cast_fp16, var_1066_cast_fp16))[name = tensor("op_1104_cast_fp16")]; + tensor var_1106_equation_0 = const()[name = tensor("op_1106_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1106_cast_fp16 = einsum(equation = var_1106_equation_0, values = (var_946_cast_fp16, var_1067_cast_fp16))[name = tensor("op_1106_cast_fp16")]; + tensor var_1108_equation_0 = const()[name = tensor("op_1108_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1108_cast_fp16 = einsum(equation = var_1108_equation_0, values = (var_946_cast_fp16, var_1068_cast_fp16))[name = tensor("op_1108_cast_fp16")]; + tensor var_1110_equation_0 = const()[name = tensor("op_1110_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1110_cast_fp16 = einsum(equation = var_1110_equation_0, values = (var_950_cast_fp16, var_1069_cast_fp16))[name = tensor("op_1110_cast_fp16")]; + tensor var_1112_equation_0 = const()[name = tensor("op_1112_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1112_cast_fp16 = einsum(equation = var_1112_equation_0, values = (var_950_cast_fp16, var_1070_cast_fp16))[name = tensor("op_1112_cast_fp16")]; + tensor var_1114_equation_0 = const()[name = tensor("op_1114_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1114_cast_fp16 = einsum(equation = var_1114_equation_0, values = (var_950_cast_fp16, var_1071_cast_fp16))[name = tensor("op_1114_cast_fp16")]; + tensor var_1116_equation_0 = const()[name = tensor("op_1116_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1116_cast_fp16 = einsum(equation = var_1116_equation_0, values = (var_950_cast_fp16, var_1072_cast_fp16))[name = tensor("op_1116_cast_fp16")]; + tensor var_1118_equation_0 = const()[name = tensor("op_1118_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1118_cast_fp16 = einsum(equation = var_1118_equation_0, values = (var_954_cast_fp16, var_1073_cast_fp16))[name = tensor("op_1118_cast_fp16")]; + tensor var_1120_equation_0 = const()[name = tensor("op_1120_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1120_cast_fp16 = einsum(equation = var_1120_equation_0, values = (var_954_cast_fp16, var_1074_cast_fp16))[name = tensor("op_1120_cast_fp16")]; + tensor var_1122_equation_0 = const()[name = tensor("op_1122_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1122_cast_fp16 = einsum(equation = var_1122_equation_0, values = (var_954_cast_fp16, var_1075_cast_fp16))[name = tensor("op_1122_cast_fp16")]; + tensor var_1124_equation_0 = const()[name = tensor("op_1124_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1124_cast_fp16 = einsum(equation = var_1124_equation_0, values = (var_954_cast_fp16, var_1076_cast_fp16))[name = tensor("op_1124_cast_fp16")]; + tensor var_1126_interleave_0 = const()[name = tensor("op_1126_interleave_0"), val = tensor(false)]; + tensor var_1126_cast_fp16 = concat(axis = var_651, interleave = var_1126_interleave_0, values = (var_1078_cast_fp16, var_1080_cast_fp16, var_1082_cast_fp16, var_1084_cast_fp16))[name = tensor("op_1126_cast_fp16")]; + tensor var_1128_interleave_0 = const()[name = tensor("op_1128_interleave_0"), val = tensor(false)]; + tensor var_1128_cast_fp16 = concat(axis = var_651, interleave = var_1128_interleave_0, values = (var_1086_cast_fp16, var_1088_cast_fp16, var_1090_cast_fp16, var_1092_cast_fp16))[name = tensor("op_1128_cast_fp16")]; + tensor var_1130_interleave_0 = const()[name = tensor("op_1130_interleave_0"), val = tensor(false)]; + tensor var_1130_cast_fp16 = concat(axis = var_651, interleave = var_1130_interleave_0, values = (var_1094_cast_fp16, var_1096_cast_fp16, var_1098_cast_fp16, var_1100_cast_fp16))[name = tensor("op_1130_cast_fp16")]; + tensor var_1132_interleave_0 = const()[name = tensor("op_1132_interleave_0"), val = tensor(false)]; + tensor var_1132_cast_fp16 = concat(axis = var_651, interleave = var_1132_interleave_0, values = (var_1102_cast_fp16, var_1104_cast_fp16, var_1106_cast_fp16, var_1108_cast_fp16))[name = tensor("op_1132_cast_fp16")]; + tensor var_1134_interleave_0 = const()[name = tensor("op_1134_interleave_0"), val = tensor(false)]; + tensor var_1134_cast_fp16 = concat(axis = var_651, interleave = var_1134_interleave_0, values = (var_1110_cast_fp16, var_1112_cast_fp16, var_1114_cast_fp16, var_1116_cast_fp16))[name = tensor("op_1134_cast_fp16")]; + tensor var_1136_interleave_0 = const()[name = tensor("op_1136_interleave_0"), val = tensor(false)]; + tensor var_1136_cast_fp16 = concat(axis = var_651, interleave = var_1136_interleave_0, values = (var_1118_cast_fp16, var_1120_cast_fp16, var_1122_cast_fp16, var_1124_cast_fp16))[name = tensor("op_1136_cast_fp16")]; + tensor input_9_interleave_0 = const()[name = tensor("input_9_interleave_0"), val = tensor(false)]; + tensor input_9_cast_fp16 = concat(axis = var_662, interleave = input_9_interleave_0, values = (var_1126_cast_fp16, var_1128_cast_fp16, var_1130_cast_fp16, var_1132_cast_fp16, var_1134_cast_fp16, var_1136_cast_fp16))[name = tensor("input_9_cast_fp16")]; + tensor var_1141 = const()[name = tensor("op_1141"), val = tensor([1, 1])]; + tensor var_1143 = const()[name = tensor("op_1143"), val = tensor([1, 1])]; + tensor obj_7_pad_type_0 = const()[name = tensor("obj_7_pad_type_0"), val = tensor("custom")]; + tensor obj_7_pad_0 = const()[name = tensor("obj_7_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_1_self_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_1_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6662016)))]; + tensor layers_1_self_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_1_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6956992)))]; + tensor obj_7_cast_fp16 = conv(bias = layers_1_self_attn_o_proj_bias_to_fp16, dilations = var_1143, groups = var_662, pad = obj_7_pad_0, pad_type = obj_7_pad_type_0, strides = var_1141, weight = layers_1_self_attn_o_proj_weight_to_fp16, x = input_9_cast_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_1149 = const()[name = tensor("op_1149"), val = tensor([1])]; + tensor channels_mean_7_cast_fp16 = reduce_mean(axes = var_1149, keep_dims = var_663, 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_1153 = const()[name = tensor("op_1153"), val = tensor([1])]; + tensor var_1154_cast_fp16 = reduce_mean(axes = var_1153, keep_dims = var_663, x = zero_mean_sq_7_cast_fp16)[name = tensor("op_1154_cast_fp16")]; + tensor var_1155_to_fp16 = const()[name = tensor("op_1155_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_1156_cast_fp16 = add(x = var_1154_cast_fp16, y = var_1155_to_fp16)[name = tensor("op_1156_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_1156_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 input_11_gamma_0_to_fp16 = const()[name = tensor("input_11_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6957824)))]; + tensor input_11_beta_0_to_fp16 = const()[name = tensor("input_11_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6958656)))]; + tensor input_11_epsilon_0_to_fp16 = const()[name = tensor("input_11_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_11_cast_fp16 = batch_norm(beta = input_11_beta_0_to_fp16, epsilon = input_11_epsilon_0_to_fp16, gamma = input_11_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("input_11_cast_fp16")]; + tensor var_1167 = const()[name = tensor("op_1167"), val = tensor([1, 1])]; + tensor var_1169 = const()[name = tensor("op_1169"), val = tensor([1, 1])]; + tensor input_13_pad_type_0 = const()[name = tensor("input_13_pad_type_0"), val = tensor("custom")]; + tensor input_13_pad_0 = const()[name = tensor("input_13_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_1_fc1_weight_to_fp16 = const()[name = tensor("layers_1_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6959488)))]; + tensor layers_1_fc1_bias_to_fp16 = const()[name = tensor("layers_1_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8139200)))]; + tensor input_13_cast_fp16 = conv(bias = layers_1_fc1_bias_to_fp16, dilations = var_1169, groups = var_662, pad = input_13_pad_0, pad_type = input_13_pad_type_0, strides = var_1167, weight = layers_1_fc1_weight_to_fp16, x = input_11_cast_fp16)[name = tensor("input_13_cast_fp16")]; + tensor input_15_mode_0 = const()[name = tensor("input_15_mode_0"), val = tensor("EXACT")]; + tensor input_15_cast_fp16 = gelu(mode = input_15_mode_0, x = input_13_cast_fp16)[name = tensor("input_15_cast_fp16")]; + tensor var_1175 = const()[name = tensor("op_1175"), val = tensor([1, 1])]; + tensor var_1177 = const()[name = tensor("op_1177"), val = tensor([1, 1])]; + tensor hidden_states_7_pad_type_0 = const()[name = tensor("hidden_states_7_pad_type_0"), val = tensor("custom")]; + tensor hidden_states_7_pad_0 = const()[name = tensor("hidden_states_7_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_1_fc2_weight_to_fp16 = const()[name = tensor("layers_1_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8142336)))]; + tensor layers_1_fc2_bias_to_fp16 = const()[name = tensor("layers_1_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9322048)))]; + tensor hidden_states_7_cast_fp16 = conv(bias = layers_1_fc2_bias_to_fp16, dilations = var_1177, groups = var_662, pad = hidden_states_7_pad_0, pad_type = hidden_states_7_pad_type_0, strides = var_1175, weight = layers_1_fc2_weight_to_fp16, x = input_15_cast_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_1184 = const()[name = tensor("op_1184"), val = tensor(3)]; + tensor var_1195 = const()[name = tensor("op_1195"), val = tensor(1)]; + tensor var_1196 = const()[name = tensor("op_1196"), val = tensor(true)]; + tensor var_1206 = const()[name = tensor("op_1206"), val = tensor([1])]; + tensor channels_mean_9_cast_fp16 = reduce_mean(axes = var_1206, keep_dims = var_1196, 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_1210 = const()[name = tensor("op_1210"), val = tensor([1])]; + tensor var_1211_cast_fp16 = reduce_mean(axes = var_1210, keep_dims = var_1196, x = zero_mean_sq_9_cast_fp16)[name = tensor("op_1211_cast_fp16")]; + tensor var_1212_to_fp16 = const()[name = tensor("op_1212_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_1213_cast_fp16 = add(x = var_1211_cast_fp16, y = var_1212_to_fp16)[name = tensor("op_1213_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_1213_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(9322880)))]; + 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(9323712)))]; + 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 var_1228 = const()[name = tensor("op_1228"), val = tensor([1, 1])]; + tensor var_1230 = const()[name = tensor("op_1230"), val = tensor([1, 1])]; + tensor query_5_pad_type_0 = const()[name = tensor("query_5_pad_type_0"), val = tensor("custom")]; + tensor query_5_pad_0 = const()[name = tensor("query_5_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_2_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_2_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9324544)))]; + tensor layers_2_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_2_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9619520)))]; + tensor query_5_cast_fp16 = conv(bias = layers_2_self_attn_q_proj_bias_to_fp16, dilations = var_1230, groups = var_1195, pad = query_5_pad_0, pad_type = query_5_pad_type_0, strides = var_1228, weight = layers_2_self_attn_q_proj_weight_to_fp16, x = obj_9_cast_fp16)[name = tensor("query_5_cast_fp16")]; + tensor var_1234 = const()[name = tensor("op_1234"), val = tensor([1, 1])]; + tensor var_1236 = const()[name = tensor("op_1236"), val = tensor([1, 1])]; + tensor key_5_pad_type_0 = const()[name = tensor("key_5_pad_type_0"), val = tensor("custom")]; + tensor key_5_pad_0 = const()[name = tensor("key_5_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_2_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_2_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9620352)))]; + tensor key_5_cast_fp16 = conv(dilations = var_1236, groups = var_1195, pad = key_5_pad_0, pad_type = key_5_pad_type_0, strides = var_1234, weight = layers_2_self_attn_k_proj_weight_to_fp16, x = obj_9_cast_fp16)[name = tensor("key_5_cast_fp16")]; + tensor var_1241 = const()[name = tensor("op_1241"), val = tensor([1, 1])]; + tensor var_1243 = const()[name = tensor("op_1243"), val = tensor([1, 1])]; + tensor value_5_pad_type_0 = const()[name = tensor("value_5_pad_type_0"), val = tensor("custom")]; + tensor value_5_pad_0 = const()[name = tensor("value_5_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_2_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_2_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9915328)))]; + tensor layers_2_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_2_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10210304)))]; + tensor value_5_cast_fp16 = conv(bias = layers_2_self_attn_v_proj_bias_to_fp16, dilations = var_1243, groups = var_1195, pad = value_5_pad_0, pad_type = value_5_pad_type_0, strides = var_1241, weight = layers_2_self_attn_v_proj_weight_to_fp16, x = obj_9_cast_fp16)[name = tensor("value_5_cast_fp16")]; + tensor var_1250_begin_0 = const()[name = tensor("op_1250_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1250_end_0 = const()[name = tensor("op_1250_end_0"), val = tensor([1, 64, 1, 1500])]; + tensor var_1250_end_mask_0 = const()[name = tensor("op_1250_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1250_cast_fp16 = slice_by_index(begin = var_1250_begin_0, end = var_1250_end_0, end_mask = var_1250_end_mask_0, x = query_5_cast_fp16)[name = tensor("op_1250_cast_fp16")]; + tensor var_1254_begin_0 = const()[name = tensor("op_1254_begin_0"), val = tensor([0, 64, 0, 0])]; + tensor var_1254_end_0 = const()[name = tensor("op_1254_end_0"), val = tensor([1, 128, 1, 1500])]; + tensor var_1254_end_mask_0 = const()[name = tensor("op_1254_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1254_cast_fp16 = slice_by_index(begin = var_1254_begin_0, end = var_1254_end_0, end_mask = var_1254_end_mask_0, x = query_5_cast_fp16)[name = tensor("op_1254_cast_fp16")]; + tensor var_1258_begin_0 = const()[name = tensor("op_1258_begin_0"), val = tensor([0, 128, 0, 0])]; + tensor var_1258_end_0 = const()[name = tensor("op_1258_end_0"), val = tensor([1, 192, 1, 1500])]; + tensor var_1258_end_mask_0 = const()[name = tensor("op_1258_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1258_cast_fp16 = slice_by_index(begin = var_1258_begin_0, end = var_1258_end_0, end_mask = var_1258_end_mask_0, x = query_5_cast_fp16)[name = tensor("op_1258_cast_fp16")]; + tensor var_1262_begin_0 = const()[name = tensor("op_1262_begin_0"), val = tensor([0, 192, 0, 0])]; + tensor var_1262_end_0 = const()[name = tensor("op_1262_end_0"), val = tensor([1, 256, 1, 1500])]; + tensor var_1262_end_mask_0 = const()[name = tensor("op_1262_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1262_cast_fp16 = slice_by_index(begin = var_1262_begin_0, end = var_1262_end_0, end_mask = var_1262_end_mask_0, x = query_5_cast_fp16)[name = tensor("op_1262_cast_fp16")]; + tensor var_1266_begin_0 = const()[name = tensor("op_1266_begin_0"), val = tensor([0, 256, 0, 0])]; + tensor var_1266_end_0 = const()[name = tensor("op_1266_end_0"), val = tensor([1, 320, 1, 1500])]; + tensor var_1266_end_mask_0 = const()[name = tensor("op_1266_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1266_cast_fp16 = slice_by_index(begin = var_1266_begin_0, end = var_1266_end_0, end_mask = var_1266_end_mask_0, x = query_5_cast_fp16)[name = tensor("op_1266_cast_fp16")]; + tensor var_1270_begin_0 = const()[name = tensor("op_1270_begin_0"), val = tensor([0, 320, 0, 0])]; + tensor var_1270_end_0 = const()[name = tensor("op_1270_end_0"), val = tensor([1, 384, 1, 1500])]; + tensor var_1270_end_mask_0 = const()[name = tensor("op_1270_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1270_cast_fp16 = slice_by_index(begin = var_1270_begin_0, end = var_1270_end_0, end_mask = var_1270_end_mask_0, x = query_5_cast_fp16)[name = tensor("op_1270_cast_fp16")]; + tensor var_1279_begin_0 = const()[name = tensor("op_1279_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1279_end_0 = const()[name = tensor("op_1279_end_0"), val = tensor([1, 64, 1, 375])]; + tensor var_1279_end_mask_0 = const()[name = tensor("op_1279_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_1279_cast_fp16 = slice_by_index(begin = var_1279_begin_0, end = var_1279_end_0, end_mask = var_1279_end_mask_0, x = var_1250_cast_fp16)[name = tensor("op_1279_cast_fp16")]; + tensor var_1286_begin_0 = const()[name = tensor("op_1286_begin_0"), val = tensor([0, 0, 0, 375])]; + tensor var_1286_end_0 = const()[name = tensor("op_1286_end_0"), val = tensor([1, 64, 1, 750])]; + tensor var_1286_end_mask_0 = const()[name = tensor("op_1286_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_1286_cast_fp16 = slice_by_index(begin = var_1286_begin_0, end = var_1286_end_0, end_mask = var_1286_end_mask_0, x = var_1250_cast_fp16)[name = tensor("op_1286_cast_fp16")]; + tensor var_1293_begin_0 = const()[name = tensor("op_1293_begin_0"), val = tensor([0, 0, 0, 750])]; + tensor var_1293_end_0 = const()[name = tensor("op_1293_end_0"), val = tensor([1, 64, 1, 1125])]; + tensor var_1293_end_mask_0 = const()[name = tensor("op_1293_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_1293_cast_fp16 = slice_by_index(begin = var_1293_begin_0, end = var_1293_end_0, end_mask = var_1293_end_mask_0, x = var_1250_cast_fp16)[name = tensor("op_1293_cast_fp16")]; + tensor var_1300_begin_0 = const()[name = tensor("op_1300_begin_0"), val = tensor([0, 0, 0, 1125])]; + tensor var_1300_end_0 = const()[name = tensor("op_1300_end_0"), val = tensor([1, 64, 1, 1500])]; + tensor var_1300_end_mask_0 = const()[name = tensor("op_1300_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_1300_cast_fp16 = slice_by_index(begin = var_1300_begin_0, end = var_1300_end_0, end_mask = var_1300_end_mask_0, x = var_1250_cast_fp16)[name = tensor("op_1300_cast_fp16")]; + tensor var_1307_begin_0 = const()[name = tensor("op_1307_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1307_end_0 = const()[name = tensor("op_1307_end_0"), val = tensor([1, 64, 1, 375])]; + tensor var_1307_end_mask_0 = const()[name = tensor("op_1307_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_1307_cast_fp16 = slice_by_index(begin = var_1307_begin_0, end = var_1307_end_0, end_mask = var_1307_end_mask_0, x = var_1254_cast_fp16)[name = tensor("op_1307_cast_fp16")]; + tensor var_1314_begin_0 = const()[name = tensor("op_1314_begin_0"), val = tensor([0, 0, 0, 375])]; + tensor var_1314_end_0 = const()[name = tensor("op_1314_end_0"), val = tensor([1, 64, 1, 750])]; + tensor var_1314_end_mask_0 = const()[name = tensor("op_1314_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_1314_cast_fp16 = slice_by_index(begin = var_1314_begin_0, end = var_1314_end_0, end_mask = var_1314_end_mask_0, x = var_1254_cast_fp16)[name = tensor("op_1314_cast_fp16")]; + tensor var_1321_begin_0 = const()[name = tensor("op_1321_begin_0"), val = tensor([0, 0, 0, 750])]; + tensor var_1321_end_0 = const()[name = tensor("op_1321_end_0"), val = tensor([1, 64, 1, 1125])]; + tensor var_1321_end_mask_0 = const()[name = tensor("op_1321_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_1321_cast_fp16 = slice_by_index(begin = var_1321_begin_0, end = var_1321_end_0, end_mask = var_1321_end_mask_0, x = var_1254_cast_fp16)[name = tensor("op_1321_cast_fp16")]; + tensor var_1328_begin_0 = const()[name = tensor("op_1328_begin_0"), val = tensor([0, 0, 0, 1125])]; + tensor var_1328_end_0 = const()[name = tensor("op_1328_end_0"), val = tensor([1, 64, 1, 1500])]; + tensor var_1328_end_mask_0 = const()[name = tensor("op_1328_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_1328_cast_fp16 = slice_by_index(begin = var_1328_begin_0, end = var_1328_end_0, end_mask = var_1328_end_mask_0, x = var_1254_cast_fp16)[name = tensor("op_1328_cast_fp16")]; + tensor var_1335_begin_0 = const()[name = tensor("op_1335_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1335_end_0 = const()[name = tensor("op_1335_end_0"), val = tensor([1, 64, 1, 375])]; + tensor var_1335_end_mask_0 = const()[name = tensor("op_1335_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_1335_cast_fp16 = slice_by_index(begin = var_1335_begin_0, end = var_1335_end_0, end_mask = var_1335_end_mask_0, x = var_1258_cast_fp16)[name = tensor("op_1335_cast_fp16")]; + tensor var_1342_begin_0 = const()[name = tensor("op_1342_begin_0"), val = tensor([0, 0, 0, 375])]; + tensor var_1342_end_0 = const()[name = tensor("op_1342_end_0"), val = tensor([1, 64, 1, 750])]; + tensor var_1342_end_mask_0 = const()[name = tensor("op_1342_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_1342_cast_fp16 = slice_by_index(begin = var_1342_begin_0, end = var_1342_end_0, end_mask = var_1342_end_mask_0, x = var_1258_cast_fp16)[name = tensor("op_1342_cast_fp16")]; + tensor var_1349_begin_0 = const()[name = tensor("op_1349_begin_0"), val = tensor([0, 0, 0, 750])]; + tensor var_1349_end_0 = const()[name = tensor("op_1349_end_0"), val = tensor([1, 64, 1, 1125])]; + tensor var_1349_end_mask_0 = const()[name = tensor("op_1349_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_1349_cast_fp16 = slice_by_index(begin = var_1349_begin_0, end = var_1349_end_0, end_mask = var_1349_end_mask_0, x = var_1258_cast_fp16)[name = tensor("op_1349_cast_fp16")]; + tensor var_1356_begin_0 = const()[name = tensor("op_1356_begin_0"), val = tensor([0, 0, 0, 1125])]; + tensor var_1356_end_0 = const()[name = tensor("op_1356_end_0"), val = tensor([1, 64, 1, 1500])]; + tensor var_1356_end_mask_0 = const()[name = tensor("op_1356_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_1356_cast_fp16 = slice_by_index(begin = var_1356_begin_0, end = var_1356_end_0, end_mask = var_1356_end_mask_0, x = var_1258_cast_fp16)[name = tensor("op_1356_cast_fp16")]; + tensor var_1363_begin_0 = const()[name = tensor("op_1363_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1363_end_0 = const()[name = tensor("op_1363_end_0"), val = tensor([1, 64, 1, 375])]; + tensor var_1363_end_mask_0 = const()[name = tensor("op_1363_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_1363_cast_fp16 = slice_by_index(begin = var_1363_begin_0, end = var_1363_end_0, end_mask = var_1363_end_mask_0, x = var_1262_cast_fp16)[name = tensor("op_1363_cast_fp16")]; + tensor var_1370_begin_0 = const()[name = tensor("op_1370_begin_0"), val = tensor([0, 0, 0, 375])]; + tensor var_1370_end_0 = const()[name = tensor("op_1370_end_0"), val = tensor([1, 64, 1, 750])]; + tensor var_1370_end_mask_0 = const()[name = tensor("op_1370_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_1370_cast_fp16 = slice_by_index(begin = var_1370_begin_0, end = var_1370_end_0, end_mask = var_1370_end_mask_0, x = var_1262_cast_fp16)[name = tensor("op_1370_cast_fp16")]; + tensor var_1377_begin_0 = const()[name = tensor("op_1377_begin_0"), val = tensor([0, 0, 0, 750])]; + tensor var_1377_end_0 = const()[name = tensor("op_1377_end_0"), val = tensor([1, 64, 1, 1125])]; + tensor var_1377_end_mask_0 = const()[name = tensor("op_1377_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_1377_cast_fp16 = slice_by_index(begin = var_1377_begin_0, end = var_1377_end_0, end_mask = var_1377_end_mask_0, x = var_1262_cast_fp16)[name = tensor("op_1377_cast_fp16")]; + tensor var_1384_begin_0 = const()[name = tensor("op_1384_begin_0"), val = tensor([0, 0, 0, 1125])]; + tensor var_1384_end_0 = const()[name = tensor("op_1384_end_0"), val = tensor([1, 64, 1, 1500])]; + tensor var_1384_end_mask_0 = const()[name = tensor("op_1384_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_1384_cast_fp16 = slice_by_index(begin = var_1384_begin_0, end = var_1384_end_0, end_mask = var_1384_end_mask_0, x = var_1262_cast_fp16)[name = tensor("op_1384_cast_fp16")]; + tensor var_1391_begin_0 = const()[name = tensor("op_1391_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1391_end_0 = const()[name = tensor("op_1391_end_0"), val = tensor([1, 64, 1, 375])]; + tensor var_1391_end_mask_0 = const()[name = tensor("op_1391_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_1391_cast_fp16 = slice_by_index(begin = var_1391_begin_0, end = var_1391_end_0, end_mask = var_1391_end_mask_0, x = var_1266_cast_fp16)[name = tensor("op_1391_cast_fp16")]; + tensor var_1398_begin_0 = const()[name = tensor("op_1398_begin_0"), val = tensor([0, 0, 0, 375])]; + tensor var_1398_end_0 = const()[name = tensor("op_1398_end_0"), val = tensor([1, 64, 1, 750])]; + tensor var_1398_end_mask_0 = const()[name = tensor("op_1398_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_1398_cast_fp16 = slice_by_index(begin = var_1398_begin_0, end = var_1398_end_0, end_mask = var_1398_end_mask_0, x = var_1266_cast_fp16)[name = tensor("op_1398_cast_fp16")]; + tensor var_1405_begin_0 = const()[name = tensor("op_1405_begin_0"), val = tensor([0, 0, 0, 750])]; + tensor var_1405_end_0 = const()[name = tensor("op_1405_end_0"), val = tensor([1, 64, 1, 1125])]; + tensor var_1405_end_mask_0 = const()[name = tensor("op_1405_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_1405_cast_fp16 = slice_by_index(begin = var_1405_begin_0, end = var_1405_end_0, end_mask = var_1405_end_mask_0, x = var_1266_cast_fp16)[name = tensor("op_1405_cast_fp16")]; + tensor var_1412_begin_0 = const()[name = tensor("op_1412_begin_0"), val = tensor([0, 0, 0, 1125])]; + tensor var_1412_end_0 = const()[name = tensor("op_1412_end_0"), val = tensor([1, 64, 1, 1500])]; + tensor var_1412_end_mask_0 = const()[name = tensor("op_1412_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_1412_cast_fp16 = slice_by_index(begin = var_1412_begin_0, end = var_1412_end_0, end_mask = var_1412_end_mask_0, x = var_1266_cast_fp16)[name = tensor("op_1412_cast_fp16")]; + tensor var_1419_begin_0 = const()[name = tensor("op_1419_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1419_end_0 = const()[name = tensor("op_1419_end_0"), val = tensor([1, 64, 1, 375])]; + tensor var_1419_end_mask_0 = const()[name = tensor("op_1419_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_1419_cast_fp16 = slice_by_index(begin = var_1419_begin_0, end = var_1419_end_0, end_mask = var_1419_end_mask_0, x = var_1270_cast_fp16)[name = tensor("op_1419_cast_fp16")]; + tensor var_1426_begin_0 = const()[name = tensor("op_1426_begin_0"), val = tensor([0, 0, 0, 375])]; + tensor var_1426_end_0 = const()[name = tensor("op_1426_end_0"), val = tensor([1, 64, 1, 750])]; + tensor var_1426_end_mask_0 = const()[name = tensor("op_1426_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_1426_cast_fp16 = slice_by_index(begin = var_1426_begin_0, end = var_1426_end_0, end_mask = var_1426_end_mask_0, x = var_1270_cast_fp16)[name = tensor("op_1426_cast_fp16")]; + tensor var_1433_begin_0 = const()[name = tensor("op_1433_begin_0"), val = tensor([0, 0, 0, 750])]; + tensor var_1433_end_0 = const()[name = tensor("op_1433_end_0"), val = tensor([1, 64, 1, 1125])]; + tensor var_1433_end_mask_0 = const()[name = tensor("op_1433_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_1433_cast_fp16 = slice_by_index(begin = var_1433_begin_0, end = var_1433_end_0, end_mask = var_1433_end_mask_0, x = var_1270_cast_fp16)[name = tensor("op_1433_cast_fp16")]; + tensor var_1440_begin_0 = const()[name = tensor("op_1440_begin_0"), val = tensor([0, 0, 0, 1125])]; + tensor var_1440_end_0 = const()[name = tensor("op_1440_end_0"), val = tensor([1, 64, 1, 1500])]; + tensor var_1440_end_mask_0 = const()[name = tensor("op_1440_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_1440_cast_fp16 = slice_by_index(begin = var_1440_begin_0, end = var_1440_end_0, end_mask = var_1440_end_mask_0, x = var_1270_cast_fp16)[name = tensor("op_1440_cast_fp16")]; + tensor k_5_perm_0 = const()[name = tensor("k_5_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor var_1445_begin_0 = const()[name = tensor("op_1445_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1445_end_0 = const()[name = tensor("op_1445_end_0"), val = tensor([1, 1500, 1, 64])]; + tensor var_1445_end_mask_0 = const()[name = tensor("op_1445_end_mask_0"), val = tensor([true, true, true, false])]; + tensor transpose_1 = transpose(perm = k_5_perm_0, x = key_5_cast_fp16)[name = tensor("transpose_1")]; + tensor var_1445_cast_fp16 = slice_by_index(begin = var_1445_begin_0, end = var_1445_end_0, end_mask = var_1445_end_mask_0, x = transpose_1)[name = tensor("op_1445_cast_fp16")]; + tensor var_1449_begin_0 = const()[name = tensor("op_1449_begin_0"), val = tensor([0, 0, 0, 64])]; + tensor var_1449_end_0 = const()[name = tensor("op_1449_end_0"), val = tensor([1, 1500, 1, 128])]; + tensor var_1449_end_mask_0 = const()[name = tensor("op_1449_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_1449_cast_fp16 = slice_by_index(begin = var_1449_begin_0, end = var_1449_end_0, end_mask = var_1449_end_mask_0, x = transpose_1)[name = tensor("op_1449_cast_fp16")]; + tensor var_1453_begin_0 = const()[name = tensor("op_1453_begin_0"), val = tensor([0, 0, 0, 128])]; + tensor var_1453_end_0 = const()[name = tensor("op_1453_end_0"), val = tensor([1, 1500, 1, 192])]; + tensor var_1453_end_mask_0 = const()[name = tensor("op_1453_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_1453_cast_fp16 = slice_by_index(begin = var_1453_begin_0, end = var_1453_end_0, end_mask = var_1453_end_mask_0, x = transpose_1)[name = tensor("op_1453_cast_fp16")]; + tensor var_1457_begin_0 = const()[name = tensor("op_1457_begin_0"), val = tensor([0, 0, 0, 192])]; + tensor var_1457_end_0 = const()[name = tensor("op_1457_end_0"), val = tensor([1, 1500, 1, 256])]; + tensor var_1457_end_mask_0 = const()[name = tensor("op_1457_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_1457_cast_fp16 = slice_by_index(begin = var_1457_begin_0, end = var_1457_end_0, end_mask = var_1457_end_mask_0, x = transpose_1)[name = tensor("op_1457_cast_fp16")]; + tensor var_1461_begin_0 = const()[name = tensor("op_1461_begin_0"), val = tensor([0, 0, 0, 256])]; + tensor var_1461_end_0 = const()[name = tensor("op_1461_end_0"), val = tensor([1, 1500, 1, 320])]; + tensor var_1461_end_mask_0 = const()[name = tensor("op_1461_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_1461_cast_fp16 = slice_by_index(begin = var_1461_begin_0, end = var_1461_end_0, end_mask = var_1461_end_mask_0, x = transpose_1)[name = tensor("op_1461_cast_fp16")]; + tensor var_1465_begin_0 = const()[name = tensor("op_1465_begin_0"), val = tensor([0, 0, 0, 320])]; + tensor var_1465_end_0 = const()[name = tensor("op_1465_end_0"), val = tensor([1, 1500, 1, 384])]; + tensor var_1465_end_mask_0 = const()[name = tensor("op_1465_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_1465_cast_fp16 = slice_by_index(begin = var_1465_begin_0, end = var_1465_end_0, end_mask = var_1465_end_mask_0, x = transpose_1)[name = tensor("op_1465_cast_fp16")]; + tensor var_1467_begin_0 = const()[name = tensor("op_1467_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1467_end_0 = const()[name = tensor("op_1467_end_0"), val = tensor([1, 64, 1, 1500])]; + tensor var_1467_end_mask_0 = const()[name = tensor("op_1467_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1467_cast_fp16 = slice_by_index(begin = var_1467_begin_0, end = var_1467_end_0, end_mask = var_1467_end_mask_0, x = value_5_cast_fp16)[name = tensor("op_1467_cast_fp16")]; + tensor var_1471_begin_0 = const()[name = tensor("op_1471_begin_0"), val = tensor([0, 64, 0, 0])]; + tensor var_1471_end_0 = const()[name = tensor("op_1471_end_0"), val = tensor([1, 128, 1, 1500])]; + tensor var_1471_end_mask_0 = const()[name = tensor("op_1471_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1471_cast_fp16 = slice_by_index(begin = var_1471_begin_0, end = var_1471_end_0, end_mask = var_1471_end_mask_0, x = value_5_cast_fp16)[name = tensor("op_1471_cast_fp16")]; + tensor var_1475_begin_0 = const()[name = tensor("op_1475_begin_0"), val = tensor([0, 128, 0, 0])]; + tensor var_1475_end_0 = const()[name = tensor("op_1475_end_0"), val = tensor([1, 192, 1, 1500])]; + tensor var_1475_end_mask_0 = const()[name = tensor("op_1475_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1475_cast_fp16 = slice_by_index(begin = var_1475_begin_0, end = var_1475_end_0, end_mask = var_1475_end_mask_0, x = value_5_cast_fp16)[name = tensor("op_1475_cast_fp16")]; + tensor var_1479_begin_0 = const()[name = tensor("op_1479_begin_0"), val = tensor([0, 192, 0, 0])]; + tensor var_1479_end_0 = const()[name = tensor("op_1479_end_0"), val = tensor([1, 256, 1, 1500])]; + tensor var_1479_end_mask_0 = const()[name = tensor("op_1479_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1479_cast_fp16 = slice_by_index(begin = var_1479_begin_0, end = var_1479_end_0, end_mask = var_1479_end_mask_0, x = value_5_cast_fp16)[name = tensor("op_1479_cast_fp16")]; + tensor var_1483_begin_0 = const()[name = tensor("op_1483_begin_0"), val = tensor([0, 256, 0, 0])]; + tensor var_1483_end_0 = const()[name = tensor("op_1483_end_0"), val = tensor([1, 320, 1, 1500])]; + tensor var_1483_end_mask_0 = const()[name = tensor("op_1483_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1483_cast_fp16 = slice_by_index(begin = var_1483_begin_0, end = var_1483_end_0, end_mask = var_1483_end_mask_0, x = value_5_cast_fp16)[name = tensor("op_1483_cast_fp16")]; + tensor var_1487_begin_0 = const()[name = tensor("op_1487_begin_0"), val = tensor([0, 320, 0, 0])]; + tensor var_1487_end_0 = const()[name = tensor("op_1487_end_0"), val = tensor([1, 384, 1, 1500])]; + tensor var_1487_end_mask_0 = const()[name = tensor("op_1487_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1487_cast_fp16 = slice_by_index(begin = var_1487_begin_0, end = var_1487_end_0, end_mask = var_1487_end_mask_0, x = value_5_cast_fp16)[name = tensor("op_1487_cast_fp16")]; + tensor var_1491_equation_0 = const()[name = tensor("op_1491_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor var_1491_cast_fp16 = einsum(equation = var_1491_equation_0, values = (var_1445_cast_fp16, var_1279_cast_fp16))[name = tensor("op_1491_cast_fp16")]; + tensor var_1492_to_fp16 = const()[name = tensor("op_1492_to_fp16"), val = tensor(0x1p-3)]; + tensor aw_chunk_97_cast_fp16 = mul(x = var_1491_cast_fp16, y = var_1492_to_fp16)[name = tensor("aw_chunk_97_cast_fp16")]; + tensor var_1495_equation_0 = const()[name = tensor("op_1495_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor var_1495_cast_fp16 = einsum(equation = var_1495_equation_0, values = (var_1445_cast_fp16, var_1286_cast_fp16))[name = tensor("op_1495_cast_fp16")]; + tensor var_1496_to_fp16 = const()[name = tensor("op_1496_to_fp16"), val = tensor(0x1p-3)]; + tensor aw_chunk_99_cast_fp16 = mul(x = var_1495_cast_fp16, y = var_1496_to_fp16)[name = tensor("aw_chunk_99_cast_fp16")]; + tensor var_1499_equation_0 = const()[name = tensor("op_1499_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor var_1499_cast_fp16 = einsum(equation = var_1499_equation_0, values = (var_1445_cast_fp16, var_1293_cast_fp16))[name = tensor("op_1499_cast_fp16")]; + tensor var_1500_to_fp16 = const()[name = tensor("op_1500_to_fp16"), val = tensor(0x1p-3)]; + tensor aw_chunk_101_cast_fp16 = mul(x = var_1499_cast_fp16, y = var_1500_to_fp16)[name = tensor("aw_chunk_101_cast_fp16")]; + tensor var_1503_equation_0 = const()[name = tensor("op_1503_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor var_1503_cast_fp16 = einsum(equation = var_1503_equation_0, values = (var_1445_cast_fp16, var_1300_cast_fp16))[name = tensor("op_1503_cast_fp16")]; + tensor var_1504_to_fp16 = const()[name = tensor("op_1504_to_fp16"), val = tensor(0x1p-3)]; + tensor aw_chunk_103_cast_fp16 = mul(x = var_1503_cast_fp16, y = var_1504_to_fp16)[name = tensor("aw_chunk_103_cast_fp16")]; + tensor var_1507_equation_0 = const()[name = tensor("op_1507_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor var_1507_cast_fp16 = einsum(equation = var_1507_equation_0, values = (var_1449_cast_fp16, var_1307_cast_fp16))[name = tensor("op_1507_cast_fp16")]; + tensor var_1508_to_fp16 = const()[name = tensor("op_1508_to_fp16"), val = tensor(0x1p-3)]; + tensor aw_chunk_105_cast_fp16 = mul(x = var_1507_cast_fp16, y = var_1508_to_fp16)[name = tensor("aw_chunk_105_cast_fp16")]; + tensor var_1511_equation_0 = const()[name = tensor("op_1511_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor var_1511_cast_fp16 = einsum(equation = var_1511_equation_0, values = (var_1449_cast_fp16, var_1314_cast_fp16))[name = tensor("op_1511_cast_fp16")]; + tensor var_1512_to_fp16 = const()[name = tensor("op_1512_to_fp16"), val = tensor(0x1p-3)]; + tensor aw_chunk_107_cast_fp16 = mul(x = var_1511_cast_fp16, y = var_1512_to_fp16)[name = tensor("aw_chunk_107_cast_fp16")]; + tensor var_1515_equation_0 = const()[name = tensor("op_1515_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor var_1515_cast_fp16 = einsum(equation = var_1515_equation_0, values = (var_1449_cast_fp16, var_1321_cast_fp16))[name = tensor("op_1515_cast_fp16")]; + tensor var_1516_to_fp16 = const()[name = tensor("op_1516_to_fp16"), val = tensor(0x1p-3)]; + tensor aw_chunk_109_cast_fp16 = mul(x = var_1515_cast_fp16, y = var_1516_to_fp16)[name = tensor("aw_chunk_109_cast_fp16")]; + tensor var_1519_equation_0 = const()[name = tensor("op_1519_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor var_1519_cast_fp16 = einsum(equation = var_1519_equation_0, values = (var_1449_cast_fp16, var_1328_cast_fp16))[name = tensor("op_1519_cast_fp16")]; + tensor var_1520_to_fp16 = const()[name = tensor("op_1520_to_fp16"), val = tensor(0x1p-3)]; + tensor aw_chunk_111_cast_fp16 = mul(x = var_1519_cast_fp16, y = var_1520_to_fp16)[name = tensor("aw_chunk_111_cast_fp16")]; + tensor var_1523_equation_0 = const()[name = tensor("op_1523_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor var_1523_cast_fp16 = einsum(equation = var_1523_equation_0, values = (var_1453_cast_fp16, var_1335_cast_fp16))[name = tensor("op_1523_cast_fp16")]; + tensor var_1524_to_fp16 = const()[name = tensor("op_1524_to_fp16"), val = tensor(0x1p-3)]; + tensor aw_chunk_113_cast_fp16 = mul(x = var_1523_cast_fp16, y = var_1524_to_fp16)[name = tensor("aw_chunk_113_cast_fp16")]; + tensor var_1527_equation_0 = const()[name = tensor("op_1527_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor var_1527_cast_fp16 = einsum(equation = var_1527_equation_0, values = (var_1453_cast_fp16, var_1342_cast_fp16))[name = tensor("op_1527_cast_fp16")]; + tensor var_1528_to_fp16 = const()[name = tensor("op_1528_to_fp16"), val = tensor(0x1p-3)]; + tensor aw_chunk_115_cast_fp16 = mul(x = var_1527_cast_fp16, y = var_1528_to_fp16)[name = tensor("aw_chunk_115_cast_fp16")]; + tensor var_1531_equation_0 = const()[name = tensor("op_1531_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor var_1531_cast_fp16 = einsum(equation = var_1531_equation_0, values = (var_1453_cast_fp16, var_1349_cast_fp16))[name = tensor("op_1531_cast_fp16")]; + tensor var_1532_to_fp16 = const()[name = tensor("op_1532_to_fp16"), val = tensor(0x1p-3)]; + tensor aw_chunk_117_cast_fp16 = mul(x = var_1531_cast_fp16, y = var_1532_to_fp16)[name = tensor("aw_chunk_117_cast_fp16")]; + tensor var_1535_equation_0 = const()[name = tensor("op_1535_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor var_1535_cast_fp16 = einsum(equation = var_1535_equation_0, values = (var_1453_cast_fp16, var_1356_cast_fp16))[name = tensor("op_1535_cast_fp16")]; + tensor var_1536_to_fp16 = const()[name = tensor("op_1536_to_fp16"), val = tensor(0x1p-3)]; + tensor aw_chunk_119_cast_fp16 = mul(x = var_1535_cast_fp16, y = var_1536_to_fp16)[name = tensor("aw_chunk_119_cast_fp16")]; + tensor var_1539_equation_0 = const()[name = tensor("op_1539_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor var_1539_cast_fp16 = einsum(equation = var_1539_equation_0, values = (var_1457_cast_fp16, var_1363_cast_fp16))[name = tensor("op_1539_cast_fp16")]; + tensor var_1540_to_fp16 = const()[name = tensor("op_1540_to_fp16"), val = tensor(0x1p-3)]; + tensor aw_chunk_121_cast_fp16 = mul(x = var_1539_cast_fp16, y = var_1540_to_fp16)[name = tensor("aw_chunk_121_cast_fp16")]; + tensor var_1543_equation_0 = const()[name = tensor("op_1543_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor var_1543_cast_fp16 = einsum(equation = var_1543_equation_0, values = (var_1457_cast_fp16, var_1370_cast_fp16))[name = tensor("op_1543_cast_fp16")]; + tensor var_1544_to_fp16 = const()[name = tensor("op_1544_to_fp16"), val = tensor(0x1p-3)]; + tensor aw_chunk_123_cast_fp16 = mul(x = var_1543_cast_fp16, y = var_1544_to_fp16)[name = tensor("aw_chunk_123_cast_fp16")]; + tensor var_1547_equation_0 = const()[name = tensor("op_1547_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor var_1547_cast_fp16 = einsum(equation = var_1547_equation_0, values = (var_1457_cast_fp16, var_1377_cast_fp16))[name = tensor("op_1547_cast_fp16")]; + tensor var_1548_to_fp16 = const()[name = tensor("op_1548_to_fp16"), val = tensor(0x1p-3)]; + tensor aw_chunk_125_cast_fp16 = mul(x = var_1547_cast_fp16, y = var_1548_to_fp16)[name = tensor("aw_chunk_125_cast_fp16")]; + tensor var_1551_equation_0 = const()[name = tensor("op_1551_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor var_1551_cast_fp16 = einsum(equation = var_1551_equation_0, values = (var_1457_cast_fp16, var_1384_cast_fp16))[name = tensor("op_1551_cast_fp16")]; + tensor var_1552_to_fp16 = const()[name = tensor("op_1552_to_fp16"), val = tensor(0x1p-3)]; + tensor aw_chunk_127_cast_fp16 = mul(x = var_1551_cast_fp16, y = var_1552_to_fp16)[name = tensor("aw_chunk_127_cast_fp16")]; + tensor var_1555_equation_0 = const()[name = tensor("op_1555_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor var_1555_cast_fp16 = einsum(equation = var_1555_equation_0, values = (var_1461_cast_fp16, var_1391_cast_fp16))[name = tensor("op_1555_cast_fp16")]; + tensor var_1556_to_fp16 = const()[name = tensor("op_1556_to_fp16"), val = tensor(0x1p-3)]; + tensor aw_chunk_129_cast_fp16 = mul(x = var_1555_cast_fp16, y = var_1556_to_fp16)[name = tensor("aw_chunk_129_cast_fp16")]; + tensor var_1559_equation_0 = const()[name = tensor("op_1559_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor var_1559_cast_fp16 = einsum(equation = var_1559_equation_0, values = (var_1461_cast_fp16, var_1398_cast_fp16))[name = tensor("op_1559_cast_fp16")]; + tensor var_1560_to_fp16 = const()[name = tensor("op_1560_to_fp16"), val = tensor(0x1p-3)]; + tensor aw_chunk_131_cast_fp16 = mul(x = var_1559_cast_fp16, y = var_1560_to_fp16)[name = tensor("aw_chunk_131_cast_fp16")]; + tensor var_1563_equation_0 = const()[name = tensor("op_1563_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor var_1563_cast_fp16 = einsum(equation = var_1563_equation_0, values = (var_1461_cast_fp16, var_1405_cast_fp16))[name = tensor("op_1563_cast_fp16")]; + tensor var_1564_to_fp16 = const()[name = tensor("op_1564_to_fp16"), val = tensor(0x1p-3)]; + tensor aw_chunk_133_cast_fp16 = mul(x = var_1563_cast_fp16, y = var_1564_to_fp16)[name = tensor("aw_chunk_133_cast_fp16")]; + tensor var_1567_equation_0 = const()[name = tensor("op_1567_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor var_1567_cast_fp16 = einsum(equation = var_1567_equation_0, values = (var_1461_cast_fp16, var_1412_cast_fp16))[name = tensor("op_1567_cast_fp16")]; + tensor var_1568_to_fp16 = const()[name = tensor("op_1568_to_fp16"), val = tensor(0x1p-3)]; + tensor aw_chunk_135_cast_fp16 = mul(x = var_1567_cast_fp16, y = var_1568_to_fp16)[name = tensor("aw_chunk_135_cast_fp16")]; + tensor var_1571_equation_0 = const()[name = tensor("op_1571_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor var_1571_cast_fp16 = einsum(equation = var_1571_equation_0, values = (var_1465_cast_fp16, var_1419_cast_fp16))[name = tensor("op_1571_cast_fp16")]; + tensor var_1572_to_fp16 = const()[name = tensor("op_1572_to_fp16"), val = tensor(0x1p-3)]; + tensor aw_chunk_137_cast_fp16 = mul(x = var_1571_cast_fp16, y = var_1572_to_fp16)[name = tensor("aw_chunk_137_cast_fp16")]; + tensor var_1575_equation_0 = const()[name = tensor("op_1575_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor var_1575_cast_fp16 = einsum(equation = var_1575_equation_0, values = (var_1465_cast_fp16, var_1426_cast_fp16))[name = tensor("op_1575_cast_fp16")]; + tensor var_1576_to_fp16 = const()[name = tensor("op_1576_to_fp16"), val = tensor(0x1p-3)]; + tensor aw_chunk_139_cast_fp16 = mul(x = var_1575_cast_fp16, y = var_1576_to_fp16)[name = tensor("aw_chunk_139_cast_fp16")]; + tensor var_1579_equation_0 = const()[name = tensor("op_1579_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor var_1579_cast_fp16 = einsum(equation = var_1579_equation_0, values = (var_1465_cast_fp16, var_1433_cast_fp16))[name = tensor("op_1579_cast_fp16")]; + tensor var_1580_to_fp16 = const()[name = tensor("op_1580_to_fp16"), val = tensor(0x1p-3)]; + tensor aw_chunk_141_cast_fp16 = mul(x = var_1579_cast_fp16, y = var_1580_to_fp16)[name = tensor("aw_chunk_141_cast_fp16")]; + tensor var_1583_equation_0 = const()[name = tensor("op_1583_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor var_1583_cast_fp16 = einsum(equation = var_1583_equation_0, values = (var_1465_cast_fp16, var_1440_cast_fp16))[name = tensor("op_1583_cast_fp16")]; + tensor var_1584_to_fp16 = const()[name = tensor("op_1584_to_fp16"), val = tensor(0x1p-3)]; + tensor aw_chunk_143_cast_fp16 = mul(x = var_1583_cast_fp16, y = var_1584_to_fp16)[name = tensor("aw_chunk_143_cast_fp16")]; + tensor var_1586_cast_fp16 = softmax(axis = var_1195, x = aw_chunk_97_cast_fp16)[name = tensor("op_1586_cast_fp16")]; + tensor var_1587_cast_fp16 = softmax(axis = var_1195, x = aw_chunk_99_cast_fp16)[name = tensor("op_1587_cast_fp16")]; + tensor var_1588_cast_fp16 = softmax(axis = var_1195, x = aw_chunk_101_cast_fp16)[name = tensor("op_1588_cast_fp16")]; + tensor var_1589_cast_fp16 = softmax(axis = var_1195, x = aw_chunk_103_cast_fp16)[name = tensor("op_1589_cast_fp16")]; + tensor var_1590_cast_fp16 = softmax(axis = var_1195, x = aw_chunk_105_cast_fp16)[name = tensor("op_1590_cast_fp16")]; + tensor var_1591_cast_fp16 = softmax(axis = var_1195, x = aw_chunk_107_cast_fp16)[name = tensor("op_1591_cast_fp16")]; + tensor var_1592_cast_fp16 = softmax(axis = var_1195, x = aw_chunk_109_cast_fp16)[name = tensor("op_1592_cast_fp16")]; + tensor var_1593_cast_fp16 = softmax(axis = var_1195, x = aw_chunk_111_cast_fp16)[name = tensor("op_1593_cast_fp16")]; + tensor var_1594_cast_fp16 = softmax(axis = var_1195, x = aw_chunk_113_cast_fp16)[name = tensor("op_1594_cast_fp16")]; + tensor var_1595_cast_fp16 = softmax(axis = var_1195, x = aw_chunk_115_cast_fp16)[name = tensor("op_1595_cast_fp16")]; + tensor var_1596_cast_fp16 = softmax(axis = var_1195, x = aw_chunk_117_cast_fp16)[name = tensor("op_1596_cast_fp16")]; + tensor var_1597_cast_fp16 = softmax(axis = var_1195, x = aw_chunk_119_cast_fp16)[name = tensor("op_1597_cast_fp16")]; + tensor var_1598_cast_fp16 = softmax(axis = var_1195, x = aw_chunk_121_cast_fp16)[name = tensor("op_1598_cast_fp16")]; + tensor var_1599_cast_fp16 = softmax(axis = var_1195, x = aw_chunk_123_cast_fp16)[name = tensor("op_1599_cast_fp16")]; + tensor var_1600_cast_fp16 = softmax(axis = var_1195, x = aw_chunk_125_cast_fp16)[name = tensor("op_1600_cast_fp16")]; + tensor var_1601_cast_fp16 = softmax(axis = var_1195, x = aw_chunk_127_cast_fp16)[name = tensor("op_1601_cast_fp16")]; + tensor var_1602_cast_fp16 = softmax(axis = var_1195, x = aw_chunk_129_cast_fp16)[name = tensor("op_1602_cast_fp16")]; + tensor var_1603_cast_fp16 = softmax(axis = var_1195, x = aw_chunk_131_cast_fp16)[name = tensor("op_1603_cast_fp16")]; + tensor var_1604_cast_fp16 = softmax(axis = var_1195, x = aw_chunk_133_cast_fp16)[name = tensor("op_1604_cast_fp16")]; + tensor var_1605_cast_fp16 = softmax(axis = var_1195, x = aw_chunk_135_cast_fp16)[name = tensor("op_1605_cast_fp16")]; + tensor var_1606_cast_fp16 = softmax(axis = var_1195, x = aw_chunk_137_cast_fp16)[name = tensor("op_1606_cast_fp16")]; + tensor var_1607_cast_fp16 = softmax(axis = var_1195, x = aw_chunk_139_cast_fp16)[name = tensor("op_1607_cast_fp16")]; + tensor var_1608_cast_fp16 = softmax(axis = var_1195, x = aw_chunk_141_cast_fp16)[name = tensor("op_1608_cast_fp16")]; + tensor var_1609_cast_fp16 = softmax(axis = var_1195, x = aw_chunk_143_cast_fp16)[name = tensor("op_1609_cast_fp16")]; + tensor var_1611_equation_0 = const()[name = tensor("op_1611_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1611_cast_fp16 = einsum(equation = var_1611_equation_0, values = (var_1467_cast_fp16, var_1586_cast_fp16))[name = tensor("op_1611_cast_fp16")]; + tensor var_1613_equation_0 = const()[name = tensor("op_1613_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1613_cast_fp16 = einsum(equation = var_1613_equation_0, values = (var_1467_cast_fp16, var_1587_cast_fp16))[name = tensor("op_1613_cast_fp16")]; + tensor var_1615_equation_0 = const()[name = tensor("op_1615_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1615_cast_fp16 = einsum(equation = var_1615_equation_0, values = (var_1467_cast_fp16, var_1588_cast_fp16))[name = tensor("op_1615_cast_fp16")]; + tensor var_1617_equation_0 = const()[name = tensor("op_1617_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1617_cast_fp16 = einsum(equation = var_1617_equation_0, values = (var_1467_cast_fp16, var_1589_cast_fp16))[name = tensor("op_1617_cast_fp16")]; + tensor var_1619_equation_0 = const()[name = tensor("op_1619_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1619_cast_fp16 = einsum(equation = var_1619_equation_0, values = (var_1471_cast_fp16, var_1590_cast_fp16))[name = tensor("op_1619_cast_fp16")]; + tensor var_1621_equation_0 = const()[name = tensor("op_1621_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1621_cast_fp16 = einsum(equation = var_1621_equation_0, values = (var_1471_cast_fp16, var_1591_cast_fp16))[name = tensor("op_1621_cast_fp16")]; + tensor var_1623_equation_0 = const()[name = tensor("op_1623_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1623_cast_fp16 = einsum(equation = var_1623_equation_0, values = (var_1471_cast_fp16, var_1592_cast_fp16))[name = tensor("op_1623_cast_fp16")]; + tensor var_1625_equation_0 = const()[name = tensor("op_1625_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1625_cast_fp16 = einsum(equation = var_1625_equation_0, values = (var_1471_cast_fp16, var_1593_cast_fp16))[name = tensor("op_1625_cast_fp16")]; + tensor var_1627_equation_0 = const()[name = tensor("op_1627_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1627_cast_fp16 = einsum(equation = var_1627_equation_0, values = (var_1475_cast_fp16, var_1594_cast_fp16))[name = tensor("op_1627_cast_fp16")]; + tensor var_1629_equation_0 = const()[name = tensor("op_1629_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1629_cast_fp16 = einsum(equation = var_1629_equation_0, values = (var_1475_cast_fp16, var_1595_cast_fp16))[name = tensor("op_1629_cast_fp16")]; + tensor var_1631_equation_0 = const()[name = tensor("op_1631_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1631_cast_fp16 = einsum(equation = var_1631_equation_0, values = (var_1475_cast_fp16, var_1596_cast_fp16))[name = tensor("op_1631_cast_fp16")]; + tensor var_1633_equation_0 = const()[name = tensor("op_1633_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1633_cast_fp16 = einsum(equation = var_1633_equation_0, values = (var_1475_cast_fp16, var_1597_cast_fp16))[name = tensor("op_1633_cast_fp16")]; + tensor var_1635_equation_0 = const()[name = tensor("op_1635_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1635_cast_fp16 = einsum(equation = var_1635_equation_0, values = (var_1479_cast_fp16, var_1598_cast_fp16))[name = tensor("op_1635_cast_fp16")]; + tensor var_1637_equation_0 = const()[name = tensor("op_1637_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1637_cast_fp16 = einsum(equation = var_1637_equation_0, values = (var_1479_cast_fp16, var_1599_cast_fp16))[name = tensor("op_1637_cast_fp16")]; + tensor var_1639_equation_0 = const()[name = tensor("op_1639_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1639_cast_fp16 = einsum(equation = var_1639_equation_0, values = (var_1479_cast_fp16, var_1600_cast_fp16))[name = tensor("op_1639_cast_fp16")]; + tensor var_1641_equation_0 = const()[name = tensor("op_1641_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1641_cast_fp16 = einsum(equation = var_1641_equation_0, values = (var_1479_cast_fp16, var_1601_cast_fp16))[name = tensor("op_1641_cast_fp16")]; + tensor var_1643_equation_0 = const()[name = tensor("op_1643_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1643_cast_fp16 = einsum(equation = var_1643_equation_0, values = (var_1483_cast_fp16, var_1602_cast_fp16))[name = tensor("op_1643_cast_fp16")]; + tensor var_1645_equation_0 = const()[name = tensor("op_1645_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1645_cast_fp16 = einsum(equation = var_1645_equation_0, values = (var_1483_cast_fp16, var_1603_cast_fp16))[name = tensor("op_1645_cast_fp16")]; + tensor var_1647_equation_0 = const()[name = tensor("op_1647_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1647_cast_fp16 = einsum(equation = var_1647_equation_0, values = (var_1483_cast_fp16, var_1604_cast_fp16))[name = tensor("op_1647_cast_fp16")]; + tensor var_1649_equation_0 = const()[name = tensor("op_1649_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1649_cast_fp16 = einsum(equation = var_1649_equation_0, values = (var_1483_cast_fp16, var_1605_cast_fp16))[name = tensor("op_1649_cast_fp16")]; + tensor var_1651_equation_0 = const()[name = tensor("op_1651_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1651_cast_fp16 = einsum(equation = var_1651_equation_0, values = (var_1487_cast_fp16, var_1606_cast_fp16))[name = tensor("op_1651_cast_fp16")]; + tensor var_1653_equation_0 = const()[name = tensor("op_1653_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1653_cast_fp16 = einsum(equation = var_1653_equation_0, values = (var_1487_cast_fp16, var_1607_cast_fp16))[name = tensor("op_1653_cast_fp16")]; + tensor var_1655_equation_0 = const()[name = tensor("op_1655_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1655_cast_fp16 = einsum(equation = var_1655_equation_0, values = (var_1487_cast_fp16, var_1608_cast_fp16))[name = tensor("op_1655_cast_fp16")]; + tensor var_1657_equation_0 = const()[name = tensor("op_1657_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1657_cast_fp16 = einsum(equation = var_1657_equation_0, values = (var_1487_cast_fp16, var_1609_cast_fp16))[name = tensor("op_1657_cast_fp16")]; + tensor var_1659_interleave_0 = const()[name = tensor("op_1659_interleave_0"), val = tensor(false)]; + tensor var_1659_cast_fp16 = concat(axis = var_1184, interleave = var_1659_interleave_0, values = (var_1611_cast_fp16, var_1613_cast_fp16, var_1615_cast_fp16, var_1617_cast_fp16))[name = tensor("op_1659_cast_fp16")]; + tensor var_1661_interleave_0 = const()[name = tensor("op_1661_interleave_0"), val = tensor(false)]; + tensor var_1661_cast_fp16 = concat(axis = var_1184, interleave = var_1661_interleave_0, values = (var_1619_cast_fp16, var_1621_cast_fp16, var_1623_cast_fp16, var_1625_cast_fp16))[name = tensor("op_1661_cast_fp16")]; + tensor var_1663_interleave_0 = const()[name = tensor("op_1663_interleave_0"), val = tensor(false)]; + tensor var_1663_cast_fp16 = concat(axis = var_1184, interleave = var_1663_interleave_0, values = (var_1627_cast_fp16, var_1629_cast_fp16, var_1631_cast_fp16, var_1633_cast_fp16))[name = tensor("op_1663_cast_fp16")]; + tensor var_1665_interleave_0 = const()[name = tensor("op_1665_interleave_0"), val = tensor(false)]; + tensor var_1665_cast_fp16 = concat(axis = var_1184, interleave = var_1665_interleave_0, values = (var_1635_cast_fp16, var_1637_cast_fp16, var_1639_cast_fp16, var_1641_cast_fp16))[name = tensor("op_1665_cast_fp16")]; + tensor var_1667_interleave_0 = const()[name = tensor("op_1667_interleave_0"), val = tensor(false)]; + tensor var_1667_cast_fp16 = concat(axis = var_1184, interleave = var_1667_interleave_0, values = (var_1643_cast_fp16, var_1645_cast_fp16, var_1647_cast_fp16, var_1649_cast_fp16))[name = tensor("op_1667_cast_fp16")]; + tensor var_1669_interleave_0 = const()[name = tensor("op_1669_interleave_0"), val = tensor(false)]; + tensor var_1669_cast_fp16 = concat(axis = var_1184, interleave = var_1669_interleave_0, values = (var_1651_cast_fp16, var_1653_cast_fp16, var_1655_cast_fp16, var_1657_cast_fp16))[name = tensor("op_1669_cast_fp16")]; + tensor input_17_interleave_0 = const()[name = tensor("input_17_interleave_0"), val = tensor(false)]; + tensor input_17_cast_fp16 = concat(axis = var_1195, interleave = input_17_interleave_0, values = (var_1659_cast_fp16, var_1661_cast_fp16, var_1663_cast_fp16, var_1665_cast_fp16, var_1667_cast_fp16, var_1669_cast_fp16))[name = tensor("input_17_cast_fp16")]; + tensor var_1674 = const()[name = tensor("op_1674"), val = tensor([1, 1])]; + tensor var_1676 = const()[name = tensor("op_1676"), val = tensor([1, 1])]; + tensor obj_11_pad_type_0 = const()[name = tensor("obj_11_pad_type_0"), val = tensor("custom")]; + tensor obj_11_pad_0 = const()[name = tensor("obj_11_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_2_self_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_2_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10211136)))]; + tensor layers_2_self_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_2_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10506112)))]; + tensor obj_11_cast_fp16 = conv(bias = layers_2_self_attn_o_proj_bias_to_fp16, dilations = var_1676, groups = var_1195, pad = obj_11_pad_0, pad_type = obj_11_pad_type_0, strides = var_1674, weight = layers_2_self_attn_o_proj_weight_to_fp16, x = input_17_cast_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_1682 = const()[name = tensor("op_1682"), val = tensor([1])]; + tensor channels_mean_11_cast_fp16 = reduce_mean(axes = var_1682, keep_dims = var_1196, 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_1686 = const()[name = tensor("op_1686"), val = tensor([1])]; + tensor var_1687_cast_fp16 = reduce_mean(axes = var_1686, keep_dims = var_1196, x = zero_mean_sq_11_cast_fp16)[name = tensor("op_1687_cast_fp16")]; + tensor var_1688_to_fp16 = const()[name = tensor("op_1688_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_1689_cast_fp16 = add(x = var_1687_cast_fp16, y = var_1688_to_fp16)[name = tensor("op_1689_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_1689_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 input_19_gamma_0_to_fp16 = const()[name = tensor("input_19_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10506944)))]; + tensor input_19_beta_0_to_fp16 = const()[name = tensor("input_19_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10507776)))]; + tensor input_19_epsilon_0_to_fp16 = const()[name = tensor("input_19_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_19_cast_fp16 = batch_norm(beta = input_19_beta_0_to_fp16, epsilon = input_19_epsilon_0_to_fp16, gamma = input_19_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("input_19_cast_fp16")]; + tensor var_1700 = const()[name = tensor("op_1700"), val = tensor([1, 1])]; + tensor var_1702 = const()[name = tensor("op_1702"), val = tensor([1, 1])]; + tensor input_21_pad_type_0 = const()[name = tensor("input_21_pad_type_0"), val = tensor("custom")]; + tensor input_21_pad_0 = const()[name = tensor("input_21_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_2_fc1_weight_to_fp16 = const()[name = tensor("layers_2_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10508608)))]; + tensor layers_2_fc1_bias_to_fp16 = const()[name = tensor("layers_2_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11688320)))]; + tensor input_21_cast_fp16 = conv(bias = layers_2_fc1_bias_to_fp16, dilations = var_1702, groups = var_1195, pad = input_21_pad_0, pad_type = input_21_pad_type_0, strides = var_1700, weight = layers_2_fc1_weight_to_fp16, x = input_19_cast_fp16)[name = tensor("input_21_cast_fp16")]; + tensor input_23_mode_0 = const()[name = tensor("input_23_mode_0"), val = tensor("EXACT")]; + tensor input_23_cast_fp16 = gelu(mode = input_23_mode_0, x = input_21_cast_fp16)[name = tensor("input_23_cast_fp16")]; + tensor var_1708 = const()[name = tensor("op_1708"), val = tensor([1, 1])]; + tensor var_1710 = const()[name = tensor("op_1710"), val = tensor([1, 1])]; + tensor hidden_states_9_pad_type_0 = const()[name = tensor("hidden_states_9_pad_type_0"), val = tensor("custom")]; + tensor hidden_states_9_pad_0 = const()[name = tensor("hidden_states_9_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_2_fc2_weight_to_fp16 = const()[name = tensor("layers_2_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11691456)))]; + tensor layers_2_fc2_bias_to_fp16 = const()[name = tensor("layers_2_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12871168)))]; + tensor hidden_states_9_cast_fp16 = conv(bias = layers_2_fc2_bias_to_fp16, dilations = var_1710, groups = var_1195, pad = hidden_states_9_pad_0, pad_type = hidden_states_9_pad_type_0, strides = var_1708, weight = layers_2_fc2_weight_to_fp16, x = input_23_cast_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_1717 = const()[name = tensor("op_1717"), val = tensor(3)]; + tensor var_1728 = const()[name = tensor("op_1728"), val = tensor(1)]; + tensor var_1729 = const()[name = tensor("op_1729"), val = tensor(true)]; + tensor var_1739 = const()[name = tensor("op_1739"), val = tensor([1])]; + tensor channels_mean_13_cast_fp16 = reduce_mean(axes = var_1739, keep_dims = var_1729, 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_1743 = const()[name = tensor("op_1743"), val = tensor([1])]; + tensor var_1744_cast_fp16 = reduce_mean(axes = var_1743, keep_dims = var_1729, x = zero_mean_sq_13_cast_fp16)[name = tensor("op_1744_cast_fp16")]; + tensor var_1745_to_fp16 = const()[name = tensor("op_1745_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_1746_cast_fp16 = add(x = var_1744_cast_fp16, y = var_1745_to_fp16)[name = tensor("op_1746_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_1746_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(12872000)))]; + 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(12872832)))]; + 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 var_1761 = const()[name = tensor("op_1761"), val = tensor([1, 1])]; + tensor var_1763 = const()[name = tensor("op_1763"), val = tensor([1, 1])]; + tensor query_pad_type_0 = const()[name = tensor("query_pad_type_0"), val = tensor("custom")]; + tensor query_pad_0 = const()[name = tensor("query_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_3_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_3_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12873664)))]; + tensor layers_3_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_3_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13168640)))]; + tensor query_cast_fp16 = conv(bias = layers_3_self_attn_q_proj_bias_to_fp16, dilations = var_1763, groups = var_1728, pad = query_pad_0, pad_type = query_pad_type_0, strides = var_1761, weight = layers_3_self_attn_q_proj_weight_to_fp16, x = obj_13_cast_fp16)[name = tensor("query_cast_fp16")]; + tensor var_1767 = const()[name = tensor("op_1767"), val = tensor([1, 1])]; + tensor var_1769 = const()[name = tensor("op_1769"), val = tensor([1, 1])]; + tensor key_pad_type_0 = const()[name = tensor("key_pad_type_0"), val = tensor("custom")]; + tensor key_pad_0 = const()[name = tensor("key_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_3_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_3_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13169472)))]; + tensor key_cast_fp16 = conv(dilations = var_1769, groups = var_1728, pad = key_pad_0, pad_type = key_pad_type_0, strides = var_1767, weight = layers_3_self_attn_k_proj_weight_to_fp16, x = obj_13_cast_fp16)[name = tensor("key_cast_fp16")]; + tensor var_1774 = const()[name = tensor("op_1774"), val = tensor([1, 1])]; + tensor var_1776 = const()[name = tensor("op_1776"), val = tensor([1, 1])]; + tensor value_pad_type_0 = const()[name = tensor("value_pad_type_0"), val = tensor("custom")]; + tensor value_pad_0 = const()[name = tensor("value_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_3_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_3_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13464448)))]; + tensor layers_3_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_3_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13759424)))]; + tensor value_cast_fp16 = conv(bias = layers_3_self_attn_v_proj_bias_to_fp16, dilations = var_1776, groups = var_1728, pad = value_pad_0, pad_type = value_pad_type_0, strides = var_1774, weight = layers_3_self_attn_v_proj_weight_to_fp16, x = obj_13_cast_fp16)[name = tensor("value_cast_fp16")]; + tensor var_1783_begin_0 = const()[name = tensor("op_1783_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1783_end_0 = const()[name = tensor("op_1783_end_0"), val = tensor([1, 64, 1, 1500])]; + tensor var_1783_end_mask_0 = const()[name = tensor("op_1783_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1783_cast_fp16 = slice_by_index(begin = var_1783_begin_0, end = var_1783_end_0, end_mask = var_1783_end_mask_0, x = query_cast_fp16)[name = tensor("op_1783_cast_fp16")]; + tensor var_1787_begin_0 = const()[name = tensor("op_1787_begin_0"), val = tensor([0, 64, 0, 0])]; + tensor var_1787_end_0 = const()[name = tensor("op_1787_end_0"), val = tensor([1, 128, 1, 1500])]; + tensor var_1787_end_mask_0 = const()[name = tensor("op_1787_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1787_cast_fp16 = slice_by_index(begin = var_1787_begin_0, end = var_1787_end_0, end_mask = var_1787_end_mask_0, x = query_cast_fp16)[name = tensor("op_1787_cast_fp16")]; + tensor var_1791_begin_0 = const()[name = tensor("op_1791_begin_0"), val = tensor([0, 128, 0, 0])]; + tensor var_1791_end_0 = const()[name = tensor("op_1791_end_0"), val = tensor([1, 192, 1, 1500])]; + tensor var_1791_end_mask_0 = const()[name = tensor("op_1791_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1791_cast_fp16 = slice_by_index(begin = var_1791_begin_0, end = var_1791_end_0, end_mask = var_1791_end_mask_0, x = query_cast_fp16)[name = tensor("op_1791_cast_fp16")]; + tensor var_1795_begin_0 = const()[name = tensor("op_1795_begin_0"), val = tensor([0, 192, 0, 0])]; + tensor var_1795_end_0 = const()[name = tensor("op_1795_end_0"), val = tensor([1, 256, 1, 1500])]; + tensor var_1795_end_mask_0 = const()[name = tensor("op_1795_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1795_cast_fp16 = slice_by_index(begin = var_1795_begin_0, end = var_1795_end_0, end_mask = var_1795_end_mask_0, x = query_cast_fp16)[name = tensor("op_1795_cast_fp16")]; + tensor var_1799_begin_0 = const()[name = tensor("op_1799_begin_0"), val = tensor([0, 256, 0, 0])]; + tensor var_1799_end_0 = const()[name = tensor("op_1799_end_0"), val = tensor([1, 320, 1, 1500])]; + tensor var_1799_end_mask_0 = const()[name = tensor("op_1799_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1799_cast_fp16 = slice_by_index(begin = var_1799_begin_0, end = var_1799_end_0, end_mask = var_1799_end_mask_0, x = query_cast_fp16)[name = tensor("op_1799_cast_fp16")]; + tensor var_1803_begin_0 = const()[name = tensor("op_1803_begin_0"), val = tensor([0, 320, 0, 0])]; + tensor var_1803_end_0 = const()[name = tensor("op_1803_end_0"), val = tensor([1, 384, 1, 1500])]; + tensor var_1803_end_mask_0 = const()[name = tensor("op_1803_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1803_cast_fp16 = slice_by_index(begin = var_1803_begin_0, end = var_1803_end_0, end_mask = var_1803_end_mask_0, x = query_cast_fp16)[name = tensor("op_1803_cast_fp16")]; + tensor var_1812_begin_0 = const()[name = tensor("op_1812_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1812_end_0 = const()[name = tensor("op_1812_end_0"), val = tensor([1, 64, 1, 375])]; + tensor var_1812_end_mask_0 = const()[name = tensor("op_1812_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_1812_cast_fp16 = slice_by_index(begin = var_1812_begin_0, end = var_1812_end_0, end_mask = var_1812_end_mask_0, x = var_1783_cast_fp16)[name = tensor("op_1812_cast_fp16")]; + tensor var_1819_begin_0 = const()[name = tensor("op_1819_begin_0"), val = tensor([0, 0, 0, 375])]; + tensor var_1819_end_0 = const()[name = tensor("op_1819_end_0"), val = tensor([1, 64, 1, 750])]; + tensor var_1819_end_mask_0 = const()[name = tensor("op_1819_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_1819_cast_fp16 = slice_by_index(begin = var_1819_begin_0, end = var_1819_end_0, end_mask = var_1819_end_mask_0, x = var_1783_cast_fp16)[name = tensor("op_1819_cast_fp16")]; + tensor var_1826_begin_0 = const()[name = tensor("op_1826_begin_0"), val = tensor([0, 0, 0, 750])]; + tensor var_1826_end_0 = const()[name = tensor("op_1826_end_0"), val = tensor([1, 64, 1, 1125])]; + tensor var_1826_end_mask_0 = const()[name = tensor("op_1826_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_1826_cast_fp16 = slice_by_index(begin = var_1826_begin_0, end = var_1826_end_0, end_mask = var_1826_end_mask_0, x = var_1783_cast_fp16)[name = tensor("op_1826_cast_fp16")]; + tensor var_1833_begin_0 = const()[name = tensor("op_1833_begin_0"), val = tensor([0, 0, 0, 1125])]; + tensor var_1833_end_0 = const()[name = tensor("op_1833_end_0"), val = tensor([1, 64, 1, 1500])]; + tensor var_1833_end_mask_0 = const()[name = tensor("op_1833_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_1833_cast_fp16 = slice_by_index(begin = var_1833_begin_0, end = var_1833_end_0, end_mask = var_1833_end_mask_0, x = var_1783_cast_fp16)[name = tensor("op_1833_cast_fp16")]; + tensor var_1840_begin_0 = const()[name = tensor("op_1840_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1840_end_0 = const()[name = tensor("op_1840_end_0"), val = tensor([1, 64, 1, 375])]; + tensor var_1840_end_mask_0 = const()[name = tensor("op_1840_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_1840_cast_fp16 = slice_by_index(begin = var_1840_begin_0, end = var_1840_end_0, end_mask = var_1840_end_mask_0, x = var_1787_cast_fp16)[name = tensor("op_1840_cast_fp16")]; + tensor var_1847_begin_0 = const()[name = tensor("op_1847_begin_0"), val = tensor([0, 0, 0, 375])]; + tensor var_1847_end_0 = const()[name = tensor("op_1847_end_0"), val = tensor([1, 64, 1, 750])]; + tensor var_1847_end_mask_0 = const()[name = tensor("op_1847_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_1847_cast_fp16 = slice_by_index(begin = var_1847_begin_0, end = var_1847_end_0, end_mask = var_1847_end_mask_0, x = var_1787_cast_fp16)[name = tensor("op_1847_cast_fp16")]; + tensor var_1854_begin_0 = const()[name = tensor("op_1854_begin_0"), val = tensor([0, 0, 0, 750])]; + tensor var_1854_end_0 = const()[name = tensor("op_1854_end_0"), val = tensor([1, 64, 1, 1125])]; + tensor var_1854_end_mask_0 = const()[name = tensor("op_1854_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_1854_cast_fp16 = slice_by_index(begin = var_1854_begin_0, end = var_1854_end_0, end_mask = var_1854_end_mask_0, x = var_1787_cast_fp16)[name = tensor("op_1854_cast_fp16")]; + tensor var_1861_begin_0 = const()[name = tensor("op_1861_begin_0"), val = tensor([0, 0, 0, 1125])]; + tensor var_1861_end_0 = const()[name = tensor("op_1861_end_0"), val = tensor([1, 64, 1, 1500])]; + tensor var_1861_end_mask_0 = const()[name = tensor("op_1861_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_1861_cast_fp16 = slice_by_index(begin = var_1861_begin_0, end = var_1861_end_0, end_mask = var_1861_end_mask_0, x = var_1787_cast_fp16)[name = tensor("op_1861_cast_fp16")]; + tensor var_1868_begin_0 = const()[name = tensor("op_1868_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1868_end_0 = const()[name = tensor("op_1868_end_0"), val = tensor([1, 64, 1, 375])]; + tensor var_1868_end_mask_0 = const()[name = tensor("op_1868_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_1868_cast_fp16 = slice_by_index(begin = var_1868_begin_0, end = var_1868_end_0, end_mask = var_1868_end_mask_0, x = var_1791_cast_fp16)[name = tensor("op_1868_cast_fp16")]; + tensor var_1875_begin_0 = const()[name = tensor("op_1875_begin_0"), val = tensor([0, 0, 0, 375])]; + tensor var_1875_end_0 = const()[name = tensor("op_1875_end_0"), val = tensor([1, 64, 1, 750])]; + tensor var_1875_end_mask_0 = const()[name = tensor("op_1875_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_1875_cast_fp16 = slice_by_index(begin = var_1875_begin_0, end = var_1875_end_0, end_mask = var_1875_end_mask_0, x = var_1791_cast_fp16)[name = tensor("op_1875_cast_fp16")]; + tensor var_1882_begin_0 = const()[name = tensor("op_1882_begin_0"), val = tensor([0, 0, 0, 750])]; + tensor var_1882_end_0 = const()[name = tensor("op_1882_end_0"), val = tensor([1, 64, 1, 1125])]; + tensor var_1882_end_mask_0 = const()[name = tensor("op_1882_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_1882_cast_fp16 = slice_by_index(begin = var_1882_begin_0, end = var_1882_end_0, end_mask = var_1882_end_mask_0, x = var_1791_cast_fp16)[name = tensor("op_1882_cast_fp16")]; + tensor var_1889_begin_0 = const()[name = tensor("op_1889_begin_0"), val = tensor([0, 0, 0, 1125])]; + tensor var_1889_end_0 = const()[name = tensor("op_1889_end_0"), val = tensor([1, 64, 1, 1500])]; + tensor var_1889_end_mask_0 = const()[name = tensor("op_1889_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_1889_cast_fp16 = slice_by_index(begin = var_1889_begin_0, end = var_1889_end_0, end_mask = var_1889_end_mask_0, x = var_1791_cast_fp16)[name = tensor("op_1889_cast_fp16")]; + tensor var_1896_begin_0 = const()[name = tensor("op_1896_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1896_end_0 = const()[name = tensor("op_1896_end_0"), val = tensor([1, 64, 1, 375])]; + tensor var_1896_end_mask_0 = const()[name = tensor("op_1896_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_1896_cast_fp16 = slice_by_index(begin = var_1896_begin_0, end = var_1896_end_0, end_mask = var_1896_end_mask_0, x = var_1795_cast_fp16)[name = tensor("op_1896_cast_fp16")]; + tensor var_1903_begin_0 = const()[name = tensor("op_1903_begin_0"), val = tensor([0, 0, 0, 375])]; + tensor var_1903_end_0 = const()[name = tensor("op_1903_end_0"), val = tensor([1, 64, 1, 750])]; + tensor var_1903_end_mask_0 = const()[name = tensor("op_1903_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_1903_cast_fp16 = slice_by_index(begin = var_1903_begin_0, end = var_1903_end_0, end_mask = var_1903_end_mask_0, x = var_1795_cast_fp16)[name = tensor("op_1903_cast_fp16")]; + tensor var_1910_begin_0 = const()[name = tensor("op_1910_begin_0"), val = tensor([0, 0, 0, 750])]; + tensor var_1910_end_0 = const()[name = tensor("op_1910_end_0"), val = tensor([1, 64, 1, 1125])]; + tensor var_1910_end_mask_0 = const()[name = tensor("op_1910_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_1910_cast_fp16 = slice_by_index(begin = var_1910_begin_0, end = var_1910_end_0, end_mask = var_1910_end_mask_0, x = var_1795_cast_fp16)[name = tensor("op_1910_cast_fp16")]; + tensor var_1917_begin_0 = const()[name = tensor("op_1917_begin_0"), val = tensor([0, 0, 0, 1125])]; + tensor var_1917_end_0 = const()[name = tensor("op_1917_end_0"), val = tensor([1, 64, 1, 1500])]; + tensor var_1917_end_mask_0 = const()[name = tensor("op_1917_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_1917_cast_fp16 = slice_by_index(begin = var_1917_begin_0, end = var_1917_end_0, end_mask = var_1917_end_mask_0, x = var_1795_cast_fp16)[name = tensor("op_1917_cast_fp16")]; + tensor var_1924_begin_0 = const()[name = tensor("op_1924_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1924_end_0 = const()[name = tensor("op_1924_end_0"), val = tensor([1, 64, 1, 375])]; + tensor var_1924_end_mask_0 = const()[name = tensor("op_1924_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_1924_cast_fp16 = slice_by_index(begin = var_1924_begin_0, end = var_1924_end_0, end_mask = var_1924_end_mask_0, x = var_1799_cast_fp16)[name = tensor("op_1924_cast_fp16")]; + tensor var_1931_begin_0 = const()[name = tensor("op_1931_begin_0"), val = tensor([0, 0, 0, 375])]; + tensor var_1931_end_0 = const()[name = tensor("op_1931_end_0"), val = tensor([1, 64, 1, 750])]; + tensor var_1931_end_mask_0 = const()[name = tensor("op_1931_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_1931_cast_fp16 = slice_by_index(begin = var_1931_begin_0, end = var_1931_end_0, end_mask = var_1931_end_mask_0, x = var_1799_cast_fp16)[name = tensor("op_1931_cast_fp16")]; + tensor var_1938_begin_0 = const()[name = tensor("op_1938_begin_0"), val = tensor([0, 0, 0, 750])]; + tensor var_1938_end_0 = const()[name = tensor("op_1938_end_0"), val = tensor([1, 64, 1, 1125])]; + tensor var_1938_end_mask_0 = const()[name = tensor("op_1938_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_1938_cast_fp16 = slice_by_index(begin = var_1938_begin_0, end = var_1938_end_0, end_mask = var_1938_end_mask_0, x = var_1799_cast_fp16)[name = tensor("op_1938_cast_fp16")]; + tensor var_1945_begin_0 = const()[name = tensor("op_1945_begin_0"), val = tensor([0, 0, 0, 1125])]; + tensor var_1945_end_0 = const()[name = tensor("op_1945_end_0"), val = tensor([1, 64, 1, 1500])]; + tensor var_1945_end_mask_0 = const()[name = tensor("op_1945_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_1945_cast_fp16 = slice_by_index(begin = var_1945_begin_0, end = var_1945_end_0, end_mask = var_1945_end_mask_0, x = var_1799_cast_fp16)[name = tensor("op_1945_cast_fp16")]; + tensor var_1952_begin_0 = const()[name = tensor("op_1952_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1952_end_0 = const()[name = tensor("op_1952_end_0"), val = tensor([1, 64, 1, 375])]; + tensor var_1952_end_mask_0 = const()[name = tensor("op_1952_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_1952_cast_fp16 = slice_by_index(begin = var_1952_begin_0, end = var_1952_end_0, end_mask = var_1952_end_mask_0, x = var_1803_cast_fp16)[name = tensor("op_1952_cast_fp16")]; + tensor var_1959_begin_0 = const()[name = tensor("op_1959_begin_0"), val = tensor([0, 0, 0, 375])]; + tensor var_1959_end_0 = const()[name = tensor("op_1959_end_0"), val = tensor([1, 64, 1, 750])]; + tensor var_1959_end_mask_0 = const()[name = tensor("op_1959_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_1959_cast_fp16 = slice_by_index(begin = var_1959_begin_0, end = var_1959_end_0, end_mask = var_1959_end_mask_0, x = var_1803_cast_fp16)[name = tensor("op_1959_cast_fp16")]; + tensor var_1966_begin_0 = const()[name = tensor("op_1966_begin_0"), val = tensor([0, 0, 0, 750])]; + tensor var_1966_end_0 = const()[name = tensor("op_1966_end_0"), val = tensor([1, 64, 1, 1125])]; + tensor var_1966_end_mask_0 = const()[name = tensor("op_1966_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_1966_cast_fp16 = slice_by_index(begin = var_1966_begin_0, end = var_1966_end_0, end_mask = var_1966_end_mask_0, x = var_1803_cast_fp16)[name = tensor("op_1966_cast_fp16")]; + tensor var_1973_begin_0 = const()[name = tensor("op_1973_begin_0"), val = tensor([0, 0, 0, 1125])]; + tensor var_1973_end_0 = const()[name = tensor("op_1973_end_0"), val = tensor([1, 64, 1, 1500])]; + tensor var_1973_end_mask_0 = const()[name = tensor("op_1973_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_1973_cast_fp16 = slice_by_index(begin = var_1973_begin_0, end = var_1973_end_0, end_mask = var_1973_end_mask_0, x = var_1803_cast_fp16)[name = tensor("op_1973_cast_fp16")]; + tensor k_perm_0 = const()[name = tensor("k_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor var_1978_begin_0 = const()[name = tensor("op_1978_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1978_end_0 = const()[name = tensor("op_1978_end_0"), val = tensor([1, 1500, 1, 64])]; + tensor var_1978_end_mask_0 = const()[name = tensor("op_1978_end_mask_0"), val = tensor([true, true, true, false])]; + tensor transpose_0 = transpose(perm = k_perm_0, x = key_cast_fp16)[name = tensor("transpose_0")]; + tensor var_1978_cast_fp16 = slice_by_index(begin = var_1978_begin_0, end = var_1978_end_0, end_mask = var_1978_end_mask_0, x = transpose_0)[name = tensor("op_1978_cast_fp16")]; + tensor var_1982_begin_0 = const()[name = tensor("op_1982_begin_0"), val = tensor([0, 0, 0, 64])]; + tensor var_1982_end_0 = const()[name = tensor("op_1982_end_0"), val = tensor([1, 1500, 1, 128])]; + tensor var_1982_end_mask_0 = const()[name = tensor("op_1982_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_1982_cast_fp16 = slice_by_index(begin = var_1982_begin_0, end = var_1982_end_0, end_mask = var_1982_end_mask_0, x = transpose_0)[name = tensor("op_1982_cast_fp16")]; + tensor var_1986_begin_0 = const()[name = tensor("op_1986_begin_0"), val = tensor([0, 0, 0, 128])]; + tensor var_1986_end_0 = const()[name = tensor("op_1986_end_0"), val = tensor([1, 1500, 1, 192])]; + tensor var_1986_end_mask_0 = const()[name = tensor("op_1986_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_1986_cast_fp16 = slice_by_index(begin = var_1986_begin_0, end = var_1986_end_0, end_mask = var_1986_end_mask_0, x = transpose_0)[name = tensor("op_1986_cast_fp16")]; + tensor var_1990_begin_0 = const()[name = tensor("op_1990_begin_0"), val = tensor([0, 0, 0, 192])]; + tensor var_1990_end_0 = const()[name = tensor("op_1990_end_0"), val = tensor([1, 1500, 1, 256])]; + tensor var_1990_end_mask_0 = const()[name = tensor("op_1990_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_1990_cast_fp16 = slice_by_index(begin = var_1990_begin_0, end = var_1990_end_0, end_mask = var_1990_end_mask_0, x = transpose_0)[name = tensor("op_1990_cast_fp16")]; + tensor var_1994_begin_0 = const()[name = tensor("op_1994_begin_0"), val = tensor([0, 0, 0, 256])]; + tensor var_1994_end_0 = const()[name = tensor("op_1994_end_0"), val = tensor([1, 1500, 1, 320])]; + tensor var_1994_end_mask_0 = const()[name = tensor("op_1994_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_1994_cast_fp16 = slice_by_index(begin = var_1994_begin_0, end = var_1994_end_0, end_mask = var_1994_end_mask_0, x = transpose_0)[name = tensor("op_1994_cast_fp16")]; + tensor var_1998_begin_0 = const()[name = tensor("op_1998_begin_0"), val = tensor([0, 0, 0, 320])]; + tensor var_1998_end_0 = const()[name = tensor("op_1998_end_0"), val = tensor([1, 1500, 1, 384])]; + tensor var_1998_end_mask_0 = const()[name = tensor("op_1998_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_1998_cast_fp16 = slice_by_index(begin = var_1998_begin_0, end = var_1998_end_0, end_mask = var_1998_end_mask_0, x = transpose_0)[name = tensor("op_1998_cast_fp16")]; + tensor var_2000_begin_0 = const()[name = tensor("op_2000_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_2000_end_0 = const()[name = tensor("op_2000_end_0"), val = tensor([1, 64, 1, 1500])]; + tensor var_2000_end_mask_0 = const()[name = tensor("op_2000_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_2000_cast_fp16 = slice_by_index(begin = var_2000_begin_0, end = var_2000_end_0, end_mask = var_2000_end_mask_0, x = value_cast_fp16)[name = tensor("op_2000_cast_fp16")]; + tensor var_2004_begin_0 = const()[name = tensor("op_2004_begin_0"), val = tensor([0, 64, 0, 0])]; + tensor var_2004_end_0 = const()[name = tensor("op_2004_end_0"), val = tensor([1, 128, 1, 1500])]; + tensor var_2004_end_mask_0 = const()[name = tensor("op_2004_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_2004_cast_fp16 = slice_by_index(begin = var_2004_begin_0, end = var_2004_end_0, end_mask = var_2004_end_mask_0, x = value_cast_fp16)[name = tensor("op_2004_cast_fp16")]; + tensor var_2008_begin_0 = const()[name = tensor("op_2008_begin_0"), val = tensor([0, 128, 0, 0])]; + tensor var_2008_end_0 = const()[name = tensor("op_2008_end_0"), val = tensor([1, 192, 1, 1500])]; + tensor var_2008_end_mask_0 = const()[name = tensor("op_2008_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_2008_cast_fp16 = slice_by_index(begin = var_2008_begin_0, end = var_2008_end_0, end_mask = var_2008_end_mask_0, x = value_cast_fp16)[name = tensor("op_2008_cast_fp16")]; + tensor var_2012_begin_0 = const()[name = tensor("op_2012_begin_0"), val = tensor([0, 192, 0, 0])]; + tensor var_2012_end_0 = const()[name = tensor("op_2012_end_0"), val = tensor([1, 256, 1, 1500])]; + tensor var_2012_end_mask_0 = const()[name = tensor("op_2012_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_2012_cast_fp16 = slice_by_index(begin = var_2012_begin_0, end = var_2012_end_0, end_mask = var_2012_end_mask_0, x = value_cast_fp16)[name = tensor("op_2012_cast_fp16")]; + tensor var_2016_begin_0 = const()[name = tensor("op_2016_begin_0"), val = tensor([0, 256, 0, 0])]; + tensor var_2016_end_0 = const()[name = tensor("op_2016_end_0"), val = tensor([1, 320, 1, 1500])]; + tensor var_2016_end_mask_0 = const()[name = tensor("op_2016_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_2016_cast_fp16 = slice_by_index(begin = var_2016_begin_0, end = var_2016_end_0, end_mask = var_2016_end_mask_0, x = value_cast_fp16)[name = tensor("op_2016_cast_fp16")]; + tensor var_2020_begin_0 = const()[name = tensor("op_2020_begin_0"), val = tensor([0, 320, 0, 0])]; + tensor var_2020_end_0 = const()[name = tensor("op_2020_end_0"), val = tensor([1, 384, 1, 1500])]; + tensor var_2020_end_mask_0 = const()[name = tensor("op_2020_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_2020_cast_fp16 = slice_by_index(begin = var_2020_begin_0, end = var_2020_end_0, end_mask = var_2020_end_mask_0, x = value_cast_fp16)[name = tensor("op_2020_cast_fp16")]; + tensor var_2024_equation_0 = const()[name = tensor("op_2024_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor var_2024_cast_fp16 = einsum(equation = var_2024_equation_0, values = (var_1978_cast_fp16, var_1812_cast_fp16))[name = tensor("op_2024_cast_fp16")]; + tensor var_2025_to_fp16 = const()[name = tensor("op_2025_to_fp16"), val = tensor(0x1p-3)]; + tensor aw_chunk_145_cast_fp16 = mul(x = var_2024_cast_fp16, y = var_2025_to_fp16)[name = tensor("aw_chunk_145_cast_fp16")]; + tensor var_2028_equation_0 = const()[name = tensor("op_2028_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor var_2028_cast_fp16 = einsum(equation = var_2028_equation_0, values = (var_1978_cast_fp16, var_1819_cast_fp16))[name = tensor("op_2028_cast_fp16")]; + tensor var_2029_to_fp16 = const()[name = tensor("op_2029_to_fp16"), val = tensor(0x1p-3)]; + tensor aw_chunk_147_cast_fp16 = mul(x = var_2028_cast_fp16, y = var_2029_to_fp16)[name = tensor("aw_chunk_147_cast_fp16")]; + tensor var_2032_equation_0 = const()[name = tensor("op_2032_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor var_2032_cast_fp16 = einsum(equation = var_2032_equation_0, values = (var_1978_cast_fp16, var_1826_cast_fp16))[name = tensor("op_2032_cast_fp16")]; + tensor var_2033_to_fp16 = const()[name = tensor("op_2033_to_fp16"), val = tensor(0x1p-3)]; + tensor aw_chunk_149_cast_fp16 = mul(x = var_2032_cast_fp16, y = var_2033_to_fp16)[name = tensor("aw_chunk_149_cast_fp16")]; + tensor var_2036_equation_0 = const()[name = tensor("op_2036_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor var_2036_cast_fp16 = einsum(equation = var_2036_equation_0, values = (var_1978_cast_fp16, var_1833_cast_fp16))[name = tensor("op_2036_cast_fp16")]; + tensor var_2037_to_fp16 = const()[name = tensor("op_2037_to_fp16"), val = tensor(0x1p-3)]; + tensor aw_chunk_151_cast_fp16 = mul(x = var_2036_cast_fp16, y = var_2037_to_fp16)[name = tensor("aw_chunk_151_cast_fp16")]; + tensor var_2040_equation_0 = const()[name = tensor("op_2040_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor var_2040_cast_fp16 = einsum(equation = var_2040_equation_0, values = (var_1982_cast_fp16, var_1840_cast_fp16))[name = tensor("op_2040_cast_fp16")]; + tensor var_2041_to_fp16 = const()[name = tensor("op_2041_to_fp16"), val = tensor(0x1p-3)]; + tensor aw_chunk_153_cast_fp16 = mul(x = var_2040_cast_fp16, y = var_2041_to_fp16)[name = tensor("aw_chunk_153_cast_fp16")]; + tensor var_2044_equation_0 = const()[name = tensor("op_2044_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor var_2044_cast_fp16 = einsum(equation = var_2044_equation_0, values = (var_1982_cast_fp16, var_1847_cast_fp16))[name = tensor("op_2044_cast_fp16")]; + tensor var_2045_to_fp16 = const()[name = tensor("op_2045_to_fp16"), val = tensor(0x1p-3)]; + tensor aw_chunk_155_cast_fp16 = mul(x = var_2044_cast_fp16, y = var_2045_to_fp16)[name = tensor("aw_chunk_155_cast_fp16")]; + tensor var_2048_equation_0 = const()[name = tensor("op_2048_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor var_2048_cast_fp16 = einsum(equation = var_2048_equation_0, values = (var_1982_cast_fp16, var_1854_cast_fp16))[name = tensor("op_2048_cast_fp16")]; + tensor var_2049_to_fp16 = const()[name = tensor("op_2049_to_fp16"), val = tensor(0x1p-3)]; + tensor aw_chunk_157_cast_fp16 = mul(x = var_2048_cast_fp16, y = var_2049_to_fp16)[name = tensor("aw_chunk_157_cast_fp16")]; + tensor var_2052_equation_0 = const()[name = tensor("op_2052_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor var_2052_cast_fp16 = einsum(equation = var_2052_equation_0, values = (var_1982_cast_fp16, var_1861_cast_fp16))[name = tensor("op_2052_cast_fp16")]; + tensor var_2053_to_fp16 = const()[name = tensor("op_2053_to_fp16"), val = tensor(0x1p-3)]; + tensor aw_chunk_159_cast_fp16 = mul(x = var_2052_cast_fp16, y = var_2053_to_fp16)[name = tensor("aw_chunk_159_cast_fp16")]; + tensor var_2056_equation_0 = const()[name = tensor("op_2056_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor var_2056_cast_fp16 = einsum(equation = var_2056_equation_0, values = (var_1986_cast_fp16, var_1868_cast_fp16))[name = tensor("op_2056_cast_fp16")]; + tensor var_2057_to_fp16 = const()[name = tensor("op_2057_to_fp16"), val = tensor(0x1p-3)]; + tensor aw_chunk_161_cast_fp16 = mul(x = var_2056_cast_fp16, y = var_2057_to_fp16)[name = tensor("aw_chunk_161_cast_fp16")]; + tensor var_2060_equation_0 = const()[name = tensor("op_2060_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor var_2060_cast_fp16 = einsum(equation = var_2060_equation_0, values = (var_1986_cast_fp16, var_1875_cast_fp16))[name = tensor("op_2060_cast_fp16")]; + tensor var_2061_to_fp16 = const()[name = tensor("op_2061_to_fp16"), val = tensor(0x1p-3)]; + tensor aw_chunk_163_cast_fp16 = mul(x = var_2060_cast_fp16, y = var_2061_to_fp16)[name = tensor("aw_chunk_163_cast_fp16")]; + tensor var_2064_equation_0 = const()[name = tensor("op_2064_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor var_2064_cast_fp16 = einsum(equation = var_2064_equation_0, values = (var_1986_cast_fp16, var_1882_cast_fp16))[name = tensor("op_2064_cast_fp16")]; + tensor var_2065_to_fp16 = const()[name = tensor("op_2065_to_fp16"), val = tensor(0x1p-3)]; + tensor aw_chunk_165_cast_fp16 = mul(x = var_2064_cast_fp16, y = var_2065_to_fp16)[name = tensor("aw_chunk_165_cast_fp16")]; + tensor var_2068_equation_0 = const()[name = tensor("op_2068_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor var_2068_cast_fp16 = einsum(equation = var_2068_equation_0, values = (var_1986_cast_fp16, var_1889_cast_fp16))[name = tensor("op_2068_cast_fp16")]; + tensor var_2069_to_fp16 = const()[name = tensor("op_2069_to_fp16"), val = tensor(0x1p-3)]; + tensor aw_chunk_167_cast_fp16 = mul(x = var_2068_cast_fp16, y = var_2069_to_fp16)[name = tensor("aw_chunk_167_cast_fp16")]; + tensor var_2072_equation_0 = const()[name = tensor("op_2072_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor var_2072_cast_fp16 = einsum(equation = var_2072_equation_0, values = (var_1990_cast_fp16, var_1896_cast_fp16))[name = tensor("op_2072_cast_fp16")]; + tensor var_2073_to_fp16 = const()[name = tensor("op_2073_to_fp16"), val = tensor(0x1p-3)]; + tensor aw_chunk_169_cast_fp16 = mul(x = var_2072_cast_fp16, y = var_2073_to_fp16)[name = tensor("aw_chunk_169_cast_fp16")]; + tensor var_2076_equation_0 = const()[name = tensor("op_2076_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor var_2076_cast_fp16 = einsum(equation = var_2076_equation_0, values = (var_1990_cast_fp16, var_1903_cast_fp16))[name = tensor("op_2076_cast_fp16")]; + tensor var_2077_to_fp16 = const()[name = tensor("op_2077_to_fp16"), val = tensor(0x1p-3)]; + tensor aw_chunk_171_cast_fp16 = mul(x = var_2076_cast_fp16, y = var_2077_to_fp16)[name = tensor("aw_chunk_171_cast_fp16")]; + tensor var_2080_equation_0 = const()[name = tensor("op_2080_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor var_2080_cast_fp16 = einsum(equation = var_2080_equation_0, values = (var_1990_cast_fp16, var_1910_cast_fp16))[name = tensor("op_2080_cast_fp16")]; + tensor var_2081_to_fp16 = const()[name = tensor("op_2081_to_fp16"), val = tensor(0x1p-3)]; + tensor aw_chunk_173_cast_fp16 = mul(x = var_2080_cast_fp16, y = var_2081_to_fp16)[name = tensor("aw_chunk_173_cast_fp16")]; + tensor var_2084_equation_0 = const()[name = tensor("op_2084_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor var_2084_cast_fp16 = einsum(equation = var_2084_equation_0, values = (var_1990_cast_fp16, var_1917_cast_fp16))[name = tensor("op_2084_cast_fp16")]; + tensor var_2085_to_fp16 = const()[name = tensor("op_2085_to_fp16"), val = tensor(0x1p-3)]; + tensor aw_chunk_175_cast_fp16 = mul(x = var_2084_cast_fp16, y = var_2085_to_fp16)[name = tensor("aw_chunk_175_cast_fp16")]; + tensor var_2088_equation_0 = const()[name = tensor("op_2088_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor var_2088_cast_fp16 = einsum(equation = var_2088_equation_0, values = (var_1994_cast_fp16, var_1924_cast_fp16))[name = tensor("op_2088_cast_fp16")]; + tensor var_2089_to_fp16 = const()[name = tensor("op_2089_to_fp16"), val = tensor(0x1p-3)]; + tensor aw_chunk_177_cast_fp16 = mul(x = var_2088_cast_fp16, y = var_2089_to_fp16)[name = tensor("aw_chunk_177_cast_fp16")]; + tensor var_2092_equation_0 = const()[name = tensor("op_2092_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor var_2092_cast_fp16 = einsum(equation = var_2092_equation_0, values = (var_1994_cast_fp16, var_1931_cast_fp16))[name = tensor("op_2092_cast_fp16")]; + tensor var_2093_to_fp16 = const()[name = tensor("op_2093_to_fp16"), val = tensor(0x1p-3)]; + tensor aw_chunk_179_cast_fp16 = mul(x = var_2092_cast_fp16, y = var_2093_to_fp16)[name = tensor("aw_chunk_179_cast_fp16")]; + tensor var_2096_equation_0 = const()[name = tensor("op_2096_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor var_2096_cast_fp16 = einsum(equation = var_2096_equation_0, values = (var_1994_cast_fp16, var_1938_cast_fp16))[name = tensor("op_2096_cast_fp16")]; + tensor var_2097_to_fp16 = const()[name = tensor("op_2097_to_fp16"), val = tensor(0x1p-3)]; + tensor aw_chunk_181_cast_fp16 = mul(x = var_2096_cast_fp16, y = var_2097_to_fp16)[name = tensor("aw_chunk_181_cast_fp16")]; + tensor var_2100_equation_0 = const()[name = tensor("op_2100_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor var_2100_cast_fp16 = einsum(equation = var_2100_equation_0, values = (var_1994_cast_fp16, var_1945_cast_fp16))[name = tensor("op_2100_cast_fp16")]; + tensor var_2101_to_fp16 = const()[name = tensor("op_2101_to_fp16"), val = tensor(0x1p-3)]; + tensor aw_chunk_183_cast_fp16 = mul(x = var_2100_cast_fp16, y = var_2101_to_fp16)[name = tensor("aw_chunk_183_cast_fp16")]; + tensor var_2104_equation_0 = const()[name = tensor("op_2104_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor var_2104_cast_fp16 = einsum(equation = var_2104_equation_0, values = (var_1998_cast_fp16, var_1952_cast_fp16))[name = tensor("op_2104_cast_fp16")]; + tensor var_2105_to_fp16 = const()[name = tensor("op_2105_to_fp16"), val = tensor(0x1p-3)]; + tensor aw_chunk_185_cast_fp16 = mul(x = var_2104_cast_fp16, y = var_2105_to_fp16)[name = tensor("aw_chunk_185_cast_fp16")]; + tensor var_2108_equation_0 = const()[name = tensor("op_2108_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor var_2108_cast_fp16 = einsum(equation = var_2108_equation_0, values = (var_1998_cast_fp16, var_1959_cast_fp16))[name = tensor("op_2108_cast_fp16")]; + tensor var_2109_to_fp16 = const()[name = tensor("op_2109_to_fp16"), val = tensor(0x1p-3)]; + tensor aw_chunk_187_cast_fp16 = mul(x = var_2108_cast_fp16, y = var_2109_to_fp16)[name = tensor("aw_chunk_187_cast_fp16")]; + tensor var_2112_equation_0 = const()[name = tensor("op_2112_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor var_2112_cast_fp16 = einsum(equation = var_2112_equation_0, values = (var_1998_cast_fp16, var_1966_cast_fp16))[name = tensor("op_2112_cast_fp16")]; + tensor var_2113_to_fp16 = const()[name = tensor("op_2113_to_fp16"), val = tensor(0x1p-3)]; + tensor aw_chunk_189_cast_fp16 = mul(x = var_2112_cast_fp16, y = var_2113_to_fp16)[name = tensor("aw_chunk_189_cast_fp16")]; + tensor var_2116_equation_0 = const()[name = tensor("op_2116_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor var_2116_cast_fp16 = einsum(equation = var_2116_equation_0, values = (var_1998_cast_fp16, var_1973_cast_fp16))[name = tensor("op_2116_cast_fp16")]; + tensor var_2117_to_fp16 = const()[name = tensor("op_2117_to_fp16"), val = tensor(0x1p-3)]; + tensor aw_chunk_cast_fp16 = mul(x = var_2116_cast_fp16, y = var_2117_to_fp16)[name = tensor("aw_chunk_cast_fp16")]; + tensor var_2119_cast_fp16 = softmax(axis = var_1728, x = aw_chunk_145_cast_fp16)[name = tensor("op_2119_cast_fp16")]; + tensor var_2120_cast_fp16 = softmax(axis = var_1728, x = aw_chunk_147_cast_fp16)[name = tensor("op_2120_cast_fp16")]; + tensor var_2121_cast_fp16 = softmax(axis = var_1728, x = aw_chunk_149_cast_fp16)[name = tensor("op_2121_cast_fp16")]; + tensor var_2122_cast_fp16 = softmax(axis = var_1728, x = aw_chunk_151_cast_fp16)[name = tensor("op_2122_cast_fp16")]; + tensor var_2123_cast_fp16 = softmax(axis = var_1728, x = aw_chunk_153_cast_fp16)[name = tensor("op_2123_cast_fp16")]; + tensor var_2124_cast_fp16 = softmax(axis = var_1728, x = aw_chunk_155_cast_fp16)[name = tensor("op_2124_cast_fp16")]; + tensor var_2125_cast_fp16 = softmax(axis = var_1728, x = aw_chunk_157_cast_fp16)[name = tensor("op_2125_cast_fp16")]; + tensor var_2126_cast_fp16 = softmax(axis = var_1728, x = aw_chunk_159_cast_fp16)[name = tensor("op_2126_cast_fp16")]; + tensor var_2127_cast_fp16 = softmax(axis = var_1728, x = aw_chunk_161_cast_fp16)[name = tensor("op_2127_cast_fp16")]; + tensor var_2128_cast_fp16 = softmax(axis = var_1728, x = aw_chunk_163_cast_fp16)[name = tensor("op_2128_cast_fp16")]; + tensor var_2129_cast_fp16 = softmax(axis = var_1728, x = aw_chunk_165_cast_fp16)[name = tensor("op_2129_cast_fp16")]; + tensor var_2130_cast_fp16 = softmax(axis = var_1728, x = aw_chunk_167_cast_fp16)[name = tensor("op_2130_cast_fp16")]; + tensor var_2131_cast_fp16 = softmax(axis = var_1728, x = aw_chunk_169_cast_fp16)[name = tensor("op_2131_cast_fp16")]; + tensor var_2132_cast_fp16 = softmax(axis = var_1728, x = aw_chunk_171_cast_fp16)[name = tensor("op_2132_cast_fp16")]; + tensor var_2133_cast_fp16 = softmax(axis = var_1728, x = aw_chunk_173_cast_fp16)[name = tensor("op_2133_cast_fp16")]; + tensor var_2134_cast_fp16 = softmax(axis = var_1728, x = aw_chunk_175_cast_fp16)[name = tensor("op_2134_cast_fp16")]; + tensor var_2135_cast_fp16 = softmax(axis = var_1728, x = aw_chunk_177_cast_fp16)[name = tensor("op_2135_cast_fp16")]; + tensor var_2136_cast_fp16 = softmax(axis = var_1728, x = aw_chunk_179_cast_fp16)[name = tensor("op_2136_cast_fp16")]; + tensor var_2137_cast_fp16 = softmax(axis = var_1728, x = aw_chunk_181_cast_fp16)[name = tensor("op_2137_cast_fp16")]; + tensor var_2138_cast_fp16 = softmax(axis = var_1728, x = aw_chunk_183_cast_fp16)[name = tensor("op_2138_cast_fp16")]; + tensor var_2139_cast_fp16 = softmax(axis = var_1728, x = aw_chunk_185_cast_fp16)[name = tensor("op_2139_cast_fp16")]; + tensor var_2140_cast_fp16 = softmax(axis = var_1728, x = aw_chunk_187_cast_fp16)[name = tensor("op_2140_cast_fp16")]; + tensor var_2141_cast_fp16 = softmax(axis = var_1728, x = aw_chunk_189_cast_fp16)[name = tensor("op_2141_cast_fp16")]; + tensor var_2142_cast_fp16 = softmax(axis = var_1728, x = aw_chunk_cast_fp16)[name = tensor("op_2142_cast_fp16")]; + tensor var_2144_equation_0 = const()[name = tensor("op_2144_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2144_cast_fp16 = einsum(equation = var_2144_equation_0, values = (var_2000_cast_fp16, var_2119_cast_fp16))[name = tensor("op_2144_cast_fp16")]; + tensor var_2146_equation_0 = const()[name = tensor("op_2146_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2146_cast_fp16 = einsum(equation = var_2146_equation_0, values = (var_2000_cast_fp16, var_2120_cast_fp16))[name = tensor("op_2146_cast_fp16")]; + tensor var_2148_equation_0 = const()[name = tensor("op_2148_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2148_cast_fp16 = einsum(equation = var_2148_equation_0, values = (var_2000_cast_fp16, var_2121_cast_fp16))[name = tensor("op_2148_cast_fp16")]; + tensor var_2150_equation_0 = const()[name = tensor("op_2150_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2150_cast_fp16 = einsum(equation = var_2150_equation_0, values = (var_2000_cast_fp16, var_2122_cast_fp16))[name = tensor("op_2150_cast_fp16")]; + tensor var_2152_equation_0 = const()[name = tensor("op_2152_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2152_cast_fp16 = einsum(equation = var_2152_equation_0, values = (var_2004_cast_fp16, var_2123_cast_fp16))[name = tensor("op_2152_cast_fp16")]; + tensor var_2154_equation_0 = const()[name = tensor("op_2154_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2154_cast_fp16 = einsum(equation = var_2154_equation_0, values = (var_2004_cast_fp16, var_2124_cast_fp16))[name = tensor("op_2154_cast_fp16")]; + tensor var_2156_equation_0 = const()[name = tensor("op_2156_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2156_cast_fp16 = einsum(equation = var_2156_equation_0, values = (var_2004_cast_fp16, var_2125_cast_fp16))[name = tensor("op_2156_cast_fp16")]; + tensor var_2158_equation_0 = const()[name = tensor("op_2158_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2158_cast_fp16 = einsum(equation = var_2158_equation_0, values = (var_2004_cast_fp16, var_2126_cast_fp16))[name = tensor("op_2158_cast_fp16")]; + tensor var_2160_equation_0 = const()[name = tensor("op_2160_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2160_cast_fp16 = einsum(equation = var_2160_equation_0, values = (var_2008_cast_fp16, var_2127_cast_fp16))[name = tensor("op_2160_cast_fp16")]; + tensor var_2162_equation_0 = const()[name = tensor("op_2162_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2162_cast_fp16 = einsum(equation = var_2162_equation_0, values = (var_2008_cast_fp16, var_2128_cast_fp16))[name = tensor("op_2162_cast_fp16")]; + tensor var_2164_equation_0 = const()[name = tensor("op_2164_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2164_cast_fp16 = einsum(equation = var_2164_equation_0, values = (var_2008_cast_fp16, var_2129_cast_fp16))[name = tensor("op_2164_cast_fp16")]; + tensor var_2166_equation_0 = const()[name = tensor("op_2166_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2166_cast_fp16 = einsum(equation = var_2166_equation_0, values = (var_2008_cast_fp16, var_2130_cast_fp16))[name = tensor("op_2166_cast_fp16")]; + tensor var_2168_equation_0 = const()[name = tensor("op_2168_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2168_cast_fp16 = einsum(equation = var_2168_equation_0, values = (var_2012_cast_fp16, var_2131_cast_fp16))[name = tensor("op_2168_cast_fp16")]; + tensor var_2170_equation_0 = const()[name = tensor("op_2170_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2170_cast_fp16 = einsum(equation = var_2170_equation_0, values = (var_2012_cast_fp16, var_2132_cast_fp16))[name = tensor("op_2170_cast_fp16")]; + tensor var_2172_equation_0 = const()[name = tensor("op_2172_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2172_cast_fp16 = einsum(equation = var_2172_equation_0, values = (var_2012_cast_fp16, var_2133_cast_fp16))[name = tensor("op_2172_cast_fp16")]; + tensor var_2174_equation_0 = const()[name = tensor("op_2174_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2174_cast_fp16 = einsum(equation = var_2174_equation_0, values = (var_2012_cast_fp16, var_2134_cast_fp16))[name = tensor("op_2174_cast_fp16")]; + tensor var_2176_equation_0 = const()[name = tensor("op_2176_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2176_cast_fp16 = einsum(equation = var_2176_equation_0, values = (var_2016_cast_fp16, var_2135_cast_fp16))[name = tensor("op_2176_cast_fp16")]; + tensor var_2178_equation_0 = const()[name = tensor("op_2178_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2178_cast_fp16 = einsum(equation = var_2178_equation_0, values = (var_2016_cast_fp16, var_2136_cast_fp16))[name = tensor("op_2178_cast_fp16")]; + tensor var_2180_equation_0 = const()[name = tensor("op_2180_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2180_cast_fp16 = einsum(equation = var_2180_equation_0, values = (var_2016_cast_fp16, var_2137_cast_fp16))[name = tensor("op_2180_cast_fp16")]; + tensor var_2182_equation_0 = const()[name = tensor("op_2182_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2182_cast_fp16 = einsum(equation = var_2182_equation_0, values = (var_2016_cast_fp16, var_2138_cast_fp16))[name = tensor("op_2182_cast_fp16")]; + tensor var_2184_equation_0 = const()[name = tensor("op_2184_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2184_cast_fp16 = einsum(equation = var_2184_equation_0, values = (var_2020_cast_fp16, var_2139_cast_fp16))[name = tensor("op_2184_cast_fp16")]; + tensor var_2186_equation_0 = const()[name = tensor("op_2186_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2186_cast_fp16 = einsum(equation = var_2186_equation_0, values = (var_2020_cast_fp16, var_2140_cast_fp16))[name = tensor("op_2186_cast_fp16")]; + tensor var_2188_equation_0 = const()[name = tensor("op_2188_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2188_cast_fp16 = einsum(equation = var_2188_equation_0, values = (var_2020_cast_fp16, var_2141_cast_fp16))[name = tensor("op_2188_cast_fp16")]; + tensor var_2190_equation_0 = const()[name = tensor("op_2190_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2190_cast_fp16 = einsum(equation = var_2190_equation_0, values = (var_2020_cast_fp16, var_2142_cast_fp16))[name = tensor("op_2190_cast_fp16")]; + tensor var_2192_interleave_0 = const()[name = tensor("op_2192_interleave_0"), val = tensor(false)]; + tensor var_2192_cast_fp16 = concat(axis = var_1717, interleave = var_2192_interleave_0, values = (var_2144_cast_fp16, var_2146_cast_fp16, var_2148_cast_fp16, var_2150_cast_fp16))[name = tensor("op_2192_cast_fp16")]; + tensor var_2194_interleave_0 = const()[name = tensor("op_2194_interleave_0"), val = tensor(false)]; + tensor var_2194_cast_fp16 = concat(axis = var_1717, interleave = var_2194_interleave_0, values = (var_2152_cast_fp16, var_2154_cast_fp16, var_2156_cast_fp16, var_2158_cast_fp16))[name = tensor("op_2194_cast_fp16")]; + tensor var_2196_interleave_0 = const()[name = tensor("op_2196_interleave_0"), val = tensor(false)]; + tensor var_2196_cast_fp16 = concat(axis = var_1717, interleave = var_2196_interleave_0, values = (var_2160_cast_fp16, var_2162_cast_fp16, var_2164_cast_fp16, var_2166_cast_fp16))[name = tensor("op_2196_cast_fp16")]; + tensor var_2198_interleave_0 = const()[name = tensor("op_2198_interleave_0"), val = tensor(false)]; + tensor var_2198_cast_fp16 = concat(axis = var_1717, interleave = var_2198_interleave_0, values = (var_2168_cast_fp16, var_2170_cast_fp16, var_2172_cast_fp16, var_2174_cast_fp16))[name = tensor("op_2198_cast_fp16")]; + tensor var_2200_interleave_0 = const()[name = tensor("op_2200_interleave_0"), val = tensor(false)]; + tensor var_2200_cast_fp16 = concat(axis = var_1717, interleave = var_2200_interleave_0, values = (var_2176_cast_fp16, var_2178_cast_fp16, var_2180_cast_fp16, var_2182_cast_fp16))[name = tensor("op_2200_cast_fp16")]; + tensor var_2202_interleave_0 = const()[name = tensor("op_2202_interleave_0"), val = tensor(false)]; + tensor var_2202_cast_fp16 = concat(axis = var_1717, interleave = var_2202_interleave_0, values = (var_2184_cast_fp16, var_2186_cast_fp16, var_2188_cast_fp16, var_2190_cast_fp16))[name = tensor("op_2202_cast_fp16")]; + tensor input_25_interleave_0 = const()[name = tensor("input_25_interleave_0"), val = tensor(false)]; + tensor input_25_cast_fp16 = concat(axis = var_1728, interleave = input_25_interleave_0, values = (var_2192_cast_fp16, var_2194_cast_fp16, var_2196_cast_fp16, var_2198_cast_fp16, var_2200_cast_fp16, var_2202_cast_fp16))[name = tensor("input_25_cast_fp16")]; + tensor var_2207 = const()[name = tensor("op_2207"), val = tensor([1, 1])]; + tensor var_2209 = const()[name = tensor("op_2209"), val = tensor([1, 1])]; + tensor obj_pad_type_0 = const()[name = tensor("obj_pad_type_0"), val = tensor("custom")]; + tensor obj_pad_0 = const()[name = tensor("obj_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_3_self_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_3_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13760256)))]; + tensor layers_3_self_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_3_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14055232)))]; + tensor obj_cast_fp16 = conv(bias = layers_3_self_attn_o_proj_bias_to_fp16, dilations = var_2209, groups = var_1728, pad = obj_pad_0, pad_type = obj_pad_type_0, strides = var_2207, weight = layers_3_self_attn_o_proj_weight_to_fp16, x = input_25_cast_fp16)[name = tensor("obj_cast_fp16")]; + tensor inputs_15_cast_fp16 = add(x = inputs_13_cast_fp16, y = obj_cast_fp16)[name = tensor("inputs_15_cast_fp16")]; + tensor var_2215 = const()[name = tensor("op_2215"), val = tensor([1])]; + tensor channels_mean_15_cast_fp16 = reduce_mean(axes = var_2215, keep_dims = var_1729, 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_2219 = const()[name = tensor("op_2219"), val = tensor([1])]; + tensor var_2220_cast_fp16 = reduce_mean(axes = var_2219, keep_dims = var_1729, x = zero_mean_sq_15_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_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_2222_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 input_27_gamma_0_to_fp16 = const()[name = tensor("input_27_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14056064)))]; + tensor input_27_beta_0_to_fp16 = const()[name = tensor("input_27_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14056896)))]; + tensor input_27_epsilon_0_to_fp16 = const()[name = tensor("input_27_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_27_cast_fp16 = batch_norm(beta = input_27_beta_0_to_fp16, epsilon = input_27_epsilon_0_to_fp16, gamma = input_27_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("input_27_cast_fp16")]; + tensor var_2233 = const()[name = tensor("op_2233"), val = tensor([1, 1])]; + tensor var_2235 = const()[name = tensor("op_2235"), val = tensor([1, 1])]; + tensor input_29_pad_type_0 = const()[name = tensor("input_29_pad_type_0"), val = tensor("custom")]; + tensor input_29_pad_0 = const()[name = tensor("input_29_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_3_fc1_weight_to_fp16 = const()[name = tensor("layers_3_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14057728)))]; + tensor layers_3_fc1_bias_to_fp16 = const()[name = tensor("layers_3_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15237440)))]; + tensor input_29_cast_fp16 = conv(bias = layers_3_fc1_bias_to_fp16, dilations = var_2235, groups = var_1728, pad = input_29_pad_0, pad_type = input_29_pad_type_0, strides = var_2233, weight = layers_3_fc1_weight_to_fp16, x = input_27_cast_fp16)[name = tensor("input_29_cast_fp16")]; + tensor input_mode_0 = const()[name = tensor("input_mode_0"), val = tensor("EXACT")]; + tensor input_cast_fp16 = gelu(mode = input_mode_0, x = input_29_cast_fp16)[name = tensor("input_cast_fp16")]; + tensor var_2241 = const()[name = tensor("op_2241"), val = tensor([1, 1])]; + tensor var_2243 = const()[name = tensor("op_2243"), val = tensor([1, 1])]; + tensor hidden_states_pad_type_0 = const()[name = tensor("hidden_states_pad_type_0"), val = tensor("custom")]; + tensor hidden_states_pad_0 = const()[name = tensor("hidden_states_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_3_fc2_weight_to_fp16 = const()[name = tensor("layers_3_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15240576)))]; + tensor layers_3_fc2_bias_to_fp16 = const()[name = tensor("layers_3_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16420288)))]; + tensor hidden_states_cast_fp16 = conv(bias = layers_3_fc2_bias_to_fp16, dilations = var_2243, groups = var_1728, pad = hidden_states_pad_0, pad_type = hidden_states_pad_type_0, strides = var_2241, weight = layers_3_fc2_weight_to_fp16, x = input_cast_fp16)[name = tensor("hidden_states_cast_fp16")]; + tensor inputs_cast_fp16 = add(x = inputs_15_cast_fp16, y = hidden_states_cast_fp16)[name = tensor("inputs_cast_fp16")]; + tensor var_2249 = const()[name = tensor("op_2249"), val = tensor(true)]; + tensor var_2253 = const()[name = tensor("op_2253"), val = tensor([1])]; + tensor channels_mean_cast_fp16 = reduce_mean(axes = var_2253, keep_dims = var_2249, 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_2257 = const()[name = tensor("op_2257"), val = tensor([1])]; + tensor var_2258_cast_fp16 = reduce_mean(axes = var_2257, keep_dims = var_2249, x = zero_mean_sq_cast_fp16)[name = tensor("op_2258_cast_fp16")]; + tensor var_2259_to_fp16 = const()[name = tensor("op_2259_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_2260_cast_fp16 = add(x = var_2258_cast_fp16, y = var_2259_to_fp16)[name = tensor("op_2260_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_2260_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(16421120)))]; + 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(16421952)))]; + 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