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_38 = const()[name = tensor("op_38"), val = tensor([1, 1])]; tensor var_44 = const()[name = tensor("op_44"), val = tensor([1, 1])]; tensor var_49 = const()[name = tensor("op_49"), val = tensor(1)]; tensor var_54_pad_type_0 = const()[name = tensor("op_54_pad_type_0"), val = tensor("custom")]; tensor var_54_pad_0 = const()[name = tensor("op_54_pad_0"), val = tensor([0, 0, 1, 1])]; tensor var_29_to_fp16 = const()[name = tensor("op_29_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64)))]; tensor var_35_to_fp16 = const()[name = tensor("op_35_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(245888)))]; tensor var_54_cast_fp16 = conv(bias = var_35_to_fp16, dilations = var_44, groups = var_49, pad = var_54_pad_0, pad_type = var_54_pad_type_0, strides = var_38, weight = var_29_to_fp16, x = melspectrogram_features)[name = tensor("op_54_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_54_cast_fp16)[name = tensor("hidden_states_1_cast_fp16")]; tensor var_78 = const()[name = tensor("op_78"), val = tensor([2, 2])]; tensor var_84 = const()[name = tensor("op_84"), val = tensor([1, 1])]; tensor var_89 = const()[name = tensor("op_89"), val = tensor(1)]; tensor var_94_pad_type_0 = const()[name = tensor("op_94_pad_type_0"), val = tensor("custom")]; tensor var_94_pad_0 = const()[name = tensor("op_94_pad_0"), val = tensor([0, 0, 1, 1])]; tensor var_69_to_fp16 = const()[name = tensor("op_69_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(246976)))]; tensor var_75_to_fp16 = const()[name = tensor("op_75_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1819904)))]; tensor var_94_cast_fp16 = conv(bias = var_75_to_fp16, dilations = var_84, groups = var_89, pad = var_94_pad_0, pad_type = var_94_pad_type_0, strides = var_78, weight = var_69_to_fp16, x = hidden_states_1_cast_fp16)[name = tensor("op_94_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_94_cast_fp16)[name = tensor("hidden_states_3_cast_fp16")]; tensor var_112_to_fp16 = const()[name = tensor("op_112_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1820992)))]; tensor inputs_1_cast_fp16 = add(x = hidden_states_3_cast_fp16, y = var_112_to_fp16)[name = tensor("inputs_1_cast_fp16")]; tensor var_122 = const()[name = tensor("op_122"), val = tensor(3)]; tensor var_135 = const()[name = tensor("op_135"), val = tensor(1)]; tensor var_136 = const()[name = tensor("op_136"), val = tensor(true)]; tensor var_146 = const()[name = tensor("op_146"), val = tensor([1])]; tensor channels_mean_1_cast_fp16 = reduce_mean(axes = var_146, keep_dims = var_136, 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_150 = const()[name = tensor("op_150"), val = tensor([1])]; tensor var_151_cast_fp16 = reduce_mean(axes = var_150, keep_dims = var_136, x = zero_mean_sq_1_cast_fp16)[name = tensor("op_151_cast_fp16")]; tensor var_152_to_fp16 = const()[name = tensor("op_152_to_fp16"), val = tensor(0x1.5p-17)]; tensor var_153_cast_fp16 = add(x = var_151_cast_fp16, y = var_152_to_fp16)[name = tensor("op_153_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_153_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(3357056)))]; 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(3358144)))]; 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(3359232)))]; 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(3360320)))]; 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_168 = const()[name = tensor("op_168"), val = tensor([1, 1])]; tensor var_170 = const()[name = tensor("op_170"), 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(3361408)))]; 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(3885760)))]; tensor query_1_cast_fp16 = conv(bias = layers_0_self_attn_q_proj_bias_to_fp16, dilations = var_170, groups = var_135, pad = query_1_pad_0, pad_type = query_1_pad_type_0, strides = var_168, weight = layers_0_self_attn_q_proj_weight_to_fp16, x = obj_1_cast_fp16)[name = tensor("query_1_cast_fp16")]; tensor var_174 = const()[name = tensor("op_174"), val = tensor([1, 1])]; tensor var_176 = const()[name = tensor("op_176"), 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(3886848)))]; tensor key_1_cast_fp16 = conv(dilations = var_176, groups = var_135, pad = key_1_pad_0, pad_type = key_1_pad_type_0, strides = var_174, weight = layers_0_self_attn_k_proj_weight_to_fp16, x = obj_1_cast_fp16)[name = tensor("key_1_cast_fp16")]; tensor var_181 = const()[name = tensor("op_181"), val = tensor([1, 1])]; tensor var_183 = const()[name = tensor("op_183"), 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(4411200)))]; 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(4935552)))]; tensor value_1_cast_fp16 = conv(bias = layers_0_self_attn_v_proj_bias_to_fp16, dilations = var_183, groups = var_135, pad = value_1_pad_0, pad_type = value_1_pad_type_0, strides = var_181, weight = layers_0_self_attn_v_proj_weight_to_fp16, x = obj_1_cast_fp16)[name = tensor("value_1_cast_fp16")]; tensor var_190_begin_0 = const()[name = tensor("op_190_begin_0"), val = tensor([0, 0, 0, 0])]; tensor var_190_end_0 = const()[name = tensor("op_190_end_0"), val = tensor([1, 64, 1, 1500])]; tensor var_190_end_mask_0 = const()[name = tensor("op_190_end_mask_0"), val = tensor([true, false, true, true])]; tensor var_190_cast_fp16 = slice_by_index(begin = var_190_begin_0, end = var_190_end_0, end_mask = var_190_end_mask_0, x = query_1_cast_fp16)[name = tensor("op_190_cast_fp16")]; tensor var_194_begin_0 = const()[name = tensor("op_194_begin_0"), val = tensor([0, 64, 0, 0])]; tensor var_194_end_0 = const()[name = tensor("op_194_end_0"), val = tensor([1, 128, 1, 1500])]; tensor var_194_end_mask_0 = const()[name = tensor("op_194_end_mask_0"), val = tensor([true, false, true, true])]; tensor var_194_cast_fp16 = slice_by_index(begin = var_194_begin_0, end = var_194_end_0, end_mask = var_194_end_mask_0, x = query_1_cast_fp16)[name = tensor("op_194_cast_fp16")]; tensor var_198_begin_0 = const()[name = tensor("op_198_begin_0"), val = tensor([0, 128, 0, 0])]; tensor var_198_end_0 = const()[name = tensor("op_198_end_0"), val = tensor([1, 192, 1, 1500])]; tensor var_198_end_mask_0 = const()[name = tensor("op_198_end_mask_0"), val = tensor([true, false, true, true])]; tensor var_198_cast_fp16 = slice_by_index(begin = var_198_begin_0, end = var_198_end_0, end_mask = var_198_end_mask_0, x = query_1_cast_fp16)[name = tensor("op_198_cast_fp16")]; tensor var_202_begin_0 = const()[name = tensor("op_202_begin_0"), val = tensor([0, 192, 0, 0])]; tensor var_202_end_0 = const()[name = tensor("op_202_end_0"), val = tensor([1, 256, 1, 1500])]; tensor var_202_end_mask_0 = const()[name = tensor("op_202_end_mask_0"), val = tensor([true, false, true, true])]; tensor var_202_cast_fp16 = slice_by_index(begin = var_202_begin_0, end = var_202_end_0, end_mask = var_202_end_mask_0, x = query_1_cast_fp16)[name = tensor("op_202_cast_fp16")]; tensor var_206_begin_0 = const()[name = tensor("op_206_begin_0"), val = tensor([0, 256, 0, 0])]; tensor var_206_end_0 = const()[name = tensor("op_206_end_0"), val = tensor([1, 320, 1, 1500])]; tensor var_206_end_mask_0 = const()[name = tensor("op_206_end_mask_0"), val = tensor([true, false, true, true])]; tensor var_206_cast_fp16 = slice_by_index(begin = var_206_begin_0, end = var_206_end_0, end_mask = var_206_end_mask_0, x = query_1_cast_fp16)[name = tensor("op_206_cast_fp16")]; tensor var_210_begin_0 = const()[name = tensor("op_210_begin_0"), val = tensor([0, 320, 0, 0])]; tensor var_210_end_0 = const()[name = tensor("op_210_end_0"), val = tensor([1, 384, 1, 1500])]; tensor var_210_end_mask_0 = const()[name = tensor("op_210_end_mask_0"), val = tensor([true, false, true, true])]; tensor var_210_cast_fp16 = slice_by_index(begin = var_210_begin_0, end = var_210_end_0, end_mask = var_210_end_mask_0, x = query_1_cast_fp16)[name = tensor("op_210_cast_fp16")]; tensor var_214_begin_0 = const()[name = tensor("op_214_begin_0"), val = tensor([0, 384, 0, 0])]; tensor var_214_end_0 = const()[name = tensor("op_214_end_0"), val = tensor([1, 448, 1, 1500])]; tensor var_214_end_mask_0 = const()[name = tensor("op_214_end_mask_0"), val = tensor([true, false, true, true])]; tensor var_214_cast_fp16 = slice_by_index(begin = var_214_begin_0, end = var_214_end_0, end_mask = var_214_end_mask_0, x = query_1_cast_fp16)[name = tensor("op_214_cast_fp16")]; tensor var_218_begin_0 = const()[name = tensor("op_218_begin_0"), val = tensor([0, 448, 0, 0])]; tensor var_218_end_0 = const()[name = tensor("op_218_end_0"), val = tensor([1, 512, 1, 1500])]; tensor var_218_end_mask_0 = const()[name = tensor("op_218_end_mask_0"), val = tensor([true, false, true, true])]; tensor var_218_cast_fp16 = slice_by_index(begin = var_218_begin_0, end = var_218_end_0, end_mask = var_218_end_mask_0, x = query_1_cast_fp16)[name = tensor("op_218_cast_fp16")]; tensor var_227_begin_0 = const()[name = tensor("op_227_begin_0"), val = tensor([0, 0, 0, 0])]; tensor var_227_end_0 = const()[name = tensor("op_227_end_0"), val = tensor([1, 64, 1, 375])]; 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_190_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, 375])]; tensor var_234_end_0 = const()[name = tensor("op_234_end_0"), val = tensor([1, 64, 1, 750])]; 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_190_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, 750])]; tensor var_241_end_0 = const()[name = tensor("op_241_end_0"), val = tensor([1, 64, 1, 1125])]; 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_190_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, 1125])]; tensor var_248_end_0 = const()[name = tensor("op_248_end_0"), val = tensor([1, 64, 1, 1500])]; 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_190_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, 0])]; tensor var_255_end_0 = const()[name = tensor("op_255_end_0"), val = tensor([1, 64, 1, 375])]; 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_194_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, 375])]; tensor var_262_end_0 = const()[name = tensor("op_262_end_0"), val = tensor([1, 64, 1, 750])]; 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_194_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, 750])]; tensor var_269_end_0 = const()[name = tensor("op_269_end_0"), val = tensor([1, 64, 1, 1125])]; 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_194_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, 1125])]; tensor var_276_end_0 = const()[name = tensor("op_276_end_0"), val = tensor([1, 64, 1, 1500])]; 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_194_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, 0])]; tensor var_283_end_0 = const()[name = tensor("op_283_end_0"), val = tensor([1, 64, 1, 375])]; 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_198_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, 375])]; tensor var_290_end_0 = const()[name = tensor("op_290_end_0"), val = tensor([1, 64, 1, 750])]; 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_198_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, 750])]; tensor var_297_end_0 = const()[name = tensor("op_297_end_0"), val = tensor([1, 64, 1, 1125])]; 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_198_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, 1125])]; tensor var_304_end_0 = const()[name = tensor("op_304_end_0"), val = tensor([1, 64, 1, 1500])]; 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_198_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, 0])]; tensor var_311_end_0 = const()[name = tensor("op_311_end_0"), val = tensor([1, 64, 1, 375])]; 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_202_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, 375])]; tensor var_318_end_0 = const()[name = tensor("op_318_end_0"), val = tensor([1, 64, 1, 750])]; 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_202_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, 750])]; tensor var_325_end_0 = const()[name = tensor("op_325_end_0"), val = tensor([1, 64, 1, 1125])]; 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_202_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, 1125])]; tensor var_332_end_0 = const()[name = tensor("op_332_end_0"), val = tensor([1, 64, 1, 1500])]; 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_202_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, 0])]; tensor var_339_end_0 = const()[name = tensor("op_339_end_0"), val = tensor([1, 64, 1, 375])]; 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_206_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, 375])]; tensor var_346_end_0 = const()[name = tensor("op_346_end_0"), val = tensor([1, 64, 1, 750])]; 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_206_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, 750])]; tensor var_353_end_0 = const()[name = tensor("op_353_end_0"), val = tensor([1, 64, 1, 1125])]; 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_206_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, 1125])]; tensor var_360_end_0 = const()[name = tensor("op_360_end_0"), val = tensor([1, 64, 1, 1500])]; 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_206_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, 0])]; tensor var_367_end_0 = const()[name = tensor("op_367_end_0"), val = tensor([1, 64, 1, 375])]; 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_210_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, 375])]; tensor var_374_end_0 = const()[name = tensor("op_374_end_0"), val = tensor([1, 64, 1, 750])]; 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_210_cast_fp16)[name = tensor("op_374_cast_fp16")]; tensor var_381_begin_0 = const()[name = tensor("op_381_begin_0"), val = tensor([0, 0, 0, 750])]; tensor var_381_end_0 = const()[name = tensor("op_381_end_0"), val = tensor([1, 64, 1, 1125])]; tensor var_381_end_mask_0 = const()[name = tensor("op_381_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_381_cast_fp16 = slice_by_index(begin = var_381_begin_0, end = var_381_end_0, end_mask = var_381_end_mask_0, x = var_210_cast_fp16)[name = tensor("op_381_cast_fp16")]; tensor var_388_begin_0 = const()[name = tensor("op_388_begin_0"), val = tensor([0, 0, 0, 1125])]; tensor var_388_end_0 = const()[name = tensor("op_388_end_0"), val = tensor([1, 64, 1, 1500])]; tensor var_388_end_mask_0 = const()[name = tensor("op_388_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_388_cast_fp16 = slice_by_index(begin = var_388_begin_0, end = var_388_end_0, end_mask = var_388_end_mask_0, x = var_210_cast_fp16)[name = tensor("op_388_cast_fp16")]; tensor var_395_begin_0 = const()[name = tensor("op_395_begin_0"), val = tensor([0, 0, 0, 0])]; tensor var_395_end_0 = const()[name = tensor("op_395_end_0"), val = tensor([1, 64, 1, 375])]; 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 = var_214_cast_fp16)[name = tensor("op_395_cast_fp16")]; tensor var_402_begin_0 = const()[name = tensor("op_402_begin_0"), val = tensor([0, 0, 0, 375])]; tensor var_402_end_0 = const()[name = tensor("op_402_end_0"), val = tensor([1, 64, 1, 750])]; tensor var_402_end_mask_0 = const()[name = tensor("op_402_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_402_cast_fp16 = slice_by_index(begin = var_402_begin_0, end = var_402_end_0, end_mask = var_402_end_mask_0, x = var_214_cast_fp16)[name = tensor("op_402_cast_fp16")]; tensor var_409_begin_0 = const()[name = tensor("op_409_begin_0"), val = tensor([0, 0, 0, 750])]; tensor var_409_end_0 = const()[name = tensor("op_409_end_0"), val = tensor([1, 64, 1, 1125])]; tensor var_409_end_mask_0 = const()[name = tensor("op_409_end_mask_0"), val = tensor([true, true, true, false])]; 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 = var_214_cast_fp16)[name = tensor("op_409_cast_fp16")]; tensor var_416_begin_0 = const()[name = tensor("op_416_begin_0"), val = tensor([0, 0, 0, 1125])]; tensor var_416_end_0 = const()[name = tensor("op_416_end_0"), val = tensor([1, 64, 1, 1500])]; tensor var_416_end_mask_0 = const()[name = tensor("op_416_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_416_cast_fp16 = slice_by_index(begin = var_416_begin_0, end = var_416_end_0, end_mask = var_416_end_mask_0, x = var_214_cast_fp16)[name = tensor("op_416_cast_fp16")]; tensor var_423_begin_0 = const()[name = tensor("op_423_begin_0"), val = tensor([0, 0, 0, 0])]; tensor var_423_end_0 = const()[name = tensor("op_423_end_0"), val = tensor([1, 64, 1, 375])]; tensor var_423_end_mask_0 = const()[name = tensor("op_423_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_423_cast_fp16 = slice_by_index(begin = var_423_begin_0, end = var_423_end_0, end_mask = var_423_end_mask_0, x = var_218_cast_fp16)[name = tensor("op_423_cast_fp16")]; tensor var_430_begin_0 = const()[name = tensor("op_430_begin_0"), val = tensor([0, 0, 0, 375])]; tensor var_430_end_0 = const()[name = tensor("op_430_end_0"), val = tensor([1, 64, 1, 750])]; tensor var_430_end_mask_0 = const()[name = tensor("op_430_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_430_cast_fp16 = slice_by_index(begin = var_430_begin_0, end = var_430_end_0, end_mask = var_430_end_mask_0, x = var_218_cast_fp16)[name = tensor("op_430_cast_fp16")]; tensor var_437_begin_0 = const()[name = tensor("op_437_begin_0"), val = tensor([0, 0, 0, 750])]; tensor var_437_end_0 = const()[name = tensor("op_437_end_0"), val = tensor([1, 64, 1, 1125])]; tensor var_437_end_mask_0 = const()[name = tensor("op_437_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_437_cast_fp16 = slice_by_index(begin = var_437_begin_0, end = var_437_end_0, end_mask = var_437_end_mask_0, x = var_218_cast_fp16)[name = tensor("op_437_cast_fp16")]; tensor var_444_begin_0 = const()[name = tensor("op_444_begin_0"), val = tensor([0, 0, 0, 1125])]; tensor var_444_end_0 = const()[name = tensor("op_444_end_0"), val = tensor([1, 64, 1, 1500])]; tensor var_444_end_mask_0 = const()[name = tensor("op_444_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_444_cast_fp16 = slice_by_index(begin = var_444_begin_0, end = var_444_end_0, end_mask = var_444_end_mask_0, x = var_218_cast_fp16)[name = tensor("op_444_cast_fp16")]; tensor k_1_perm_0 = const()[name = tensor("k_1_perm_0"), val = tensor([0, 3, 2, 1])]; tensor var_449_begin_0 = const()[name = tensor("op_449_begin_0"), val = tensor([0, 0, 0, 0])]; tensor var_449_end_0 = const()[name = tensor("op_449_end_0"), val = tensor([1, 1500, 1, 64])]; tensor var_449_end_mask_0 = const()[name = tensor("op_449_end_mask_0"), val = tensor([true, true, true, false])]; tensor transpose_5 = transpose(perm = k_1_perm_0, x = key_1_cast_fp16)[name = tensor("transpose_5")]; tensor var_449_cast_fp16 = slice_by_index(begin = var_449_begin_0, end = var_449_end_0, end_mask = var_449_end_mask_0, x = transpose_5)[name = tensor("op_449_cast_fp16")]; tensor var_453_begin_0 = const()[name = tensor("op_453_begin_0"), val = tensor([0, 0, 0, 64])]; tensor var_453_end_0 = const()[name = tensor("op_453_end_0"), val = tensor([1, 1500, 1, 128])]; tensor var_453_end_mask_0 = const()[name = tensor("op_453_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_453_cast_fp16 = slice_by_index(begin = var_453_begin_0, end = var_453_end_0, end_mask = var_453_end_mask_0, x = transpose_5)[name = tensor("op_453_cast_fp16")]; tensor var_457_begin_0 = const()[name = tensor("op_457_begin_0"), val = tensor([0, 0, 0, 128])]; tensor var_457_end_0 = const()[name = tensor("op_457_end_0"), val = tensor([1, 1500, 1, 192])]; tensor var_457_end_mask_0 = const()[name = tensor("op_457_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_457_cast_fp16 = slice_by_index(begin = var_457_begin_0, end = var_457_end_0, end_mask = var_457_end_mask_0, x = transpose_5)[name = tensor("op_457_cast_fp16")]; tensor var_461_begin_0 = const()[name = tensor("op_461_begin_0"), val = tensor([0, 0, 0, 192])]; tensor var_461_end_0 = const()[name = tensor("op_461_end_0"), val = tensor([1, 1500, 1, 256])]; tensor var_461_end_mask_0 = const()[name = tensor("op_461_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_461_cast_fp16 = slice_by_index(begin = var_461_begin_0, end = var_461_end_0, end_mask = var_461_end_mask_0, x = transpose_5)[name = tensor("op_461_cast_fp16")]; tensor var_465_begin_0 = const()[name = tensor("op_465_begin_0"), val = tensor([0, 0, 0, 256])]; tensor var_465_end_0 = const()[name = tensor("op_465_end_0"), val = tensor([1, 1500, 1, 320])]; tensor var_465_end_mask_0 = const()[name = tensor("op_465_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_465_cast_fp16 = slice_by_index(begin = var_465_begin_0, end = var_465_end_0, end_mask = var_465_end_mask_0, x = transpose_5)[name = tensor("op_465_cast_fp16")]; tensor var_469_begin_0 = const()[name = tensor("op_469_begin_0"), val = tensor([0, 0, 0, 320])]; tensor var_469_end_0 = const()[name = tensor("op_469_end_0"), val = tensor([1, 1500, 1, 384])]; tensor var_469_end_mask_0 = const()[name = tensor("op_469_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_469_cast_fp16 = slice_by_index(begin = var_469_begin_0, end = var_469_end_0, end_mask = var_469_end_mask_0, x = transpose_5)[name = tensor("op_469_cast_fp16")]; tensor var_473_begin_0 = const()[name = tensor("op_473_begin_0"), val = tensor([0, 0, 0, 384])]; tensor var_473_end_0 = const()[name = tensor("op_473_end_0"), val = tensor([1, 1500, 1, 448])]; tensor var_473_end_mask_0 = const()[name = tensor("op_473_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_473_cast_fp16 = slice_by_index(begin = var_473_begin_0, end = var_473_end_0, end_mask = var_473_end_mask_0, x = transpose_5)[name = tensor("op_473_cast_fp16")]; tensor var_477_begin_0 = const()[name = tensor("op_477_begin_0"), val = tensor([0, 0, 0, 448])]; tensor var_477_end_0 = const()[name = tensor("op_477_end_0"), val = tensor([1, 1500, 1, 512])]; tensor var_477_end_mask_0 = const()[name = tensor("op_477_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_477_cast_fp16 = slice_by_index(begin = var_477_begin_0, end = var_477_end_0, end_mask = var_477_end_mask_0, x = transpose_5)[name = tensor("op_477_cast_fp16")]; tensor var_479_begin_0 = const()[name = tensor("op_479_begin_0"), val = tensor([0, 0, 0, 0])]; tensor var_479_end_0 = const()[name = tensor("op_479_end_0"), val = tensor([1, 64, 1, 1500])]; tensor var_479_end_mask_0 = const()[name = tensor("op_479_end_mask_0"), val = tensor([true, false, true, true])]; tensor var_479_cast_fp16 = slice_by_index(begin = var_479_begin_0, end = var_479_end_0, end_mask = var_479_end_mask_0, x = value_1_cast_fp16)[name = tensor("op_479_cast_fp16")]; tensor var_483_begin_0 = const()[name = tensor("op_483_begin_0"), val = tensor([0, 64, 0, 0])]; tensor var_483_end_0 = const()[name = tensor("op_483_end_0"), val = tensor([1, 128, 1, 1500])]; tensor var_483_end_mask_0 = const()[name = tensor("op_483_end_mask_0"), val = tensor([true, false, true, true])]; tensor var_483_cast_fp16 = slice_by_index(begin = var_483_begin_0, end = var_483_end_0, end_mask = var_483_end_mask_0, x = value_1_cast_fp16)[name = tensor("op_483_cast_fp16")]; tensor var_487_begin_0 = const()[name = tensor("op_487_begin_0"), val = tensor([0, 128, 0, 0])]; tensor var_487_end_0 = const()[name = tensor("op_487_end_0"), val = tensor([1, 192, 1, 1500])]; tensor var_487_end_mask_0 = const()[name = tensor("op_487_end_mask_0"), val = tensor([true, false, true, true])]; tensor var_487_cast_fp16 = slice_by_index(begin = var_487_begin_0, end = var_487_end_0, end_mask = var_487_end_mask_0, x = value_1_cast_fp16)[name = tensor("op_487_cast_fp16")]; tensor var_491_begin_0 = const()[name = tensor("op_491_begin_0"), val = tensor([0, 192, 0, 0])]; tensor var_491_end_0 = const()[name = tensor("op_491_end_0"), val = tensor([1, 256, 1, 1500])]; tensor var_491_end_mask_0 = const()[name = tensor("op_491_end_mask_0"), val = tensor([true, false, true, true])]; tensor var_491_cast_fp16 = slice_by_index(begin = var_491_begin_0, end = var_491_end_0, end_mask = var_491_end_mask_0, x = value_1_cast_fp16)[name = tensor("op_491_cast_fp16")]; tensor var_495_begin_0 = const()[name = tensor("op_495_begin_0"), val = tensor([0, 256, 0, 0])]; tensor var_495_end_0 = const()[name = tensor("op_495_end_0"), val = tensor([1, 320, 1, 1500])]; tensor var_495_end_mask_0 = const()[name = tensor("op_495_end_mask_0"), val = tensor([true, false, true, true])]; tensor var_495_cast_fp16 = slice_by_index(begin = var_495_begin_0, end = var_495_end_0, end_mask = var_495_end_mask_0, x = value_1_cast_fp16)[name = tensor("op_495_cast_fp16")]; tensor var_499_begin_0 = const()[name = tensor("op_499_begin_0"), val = tensor([0, 320, 0, 0])]; tensor var_499_end_0 = const()[name = tensor("op_499_end_0"), val = tensor([1, 384, 1, 1500])]; tensor var_499_end_mask_0 = const()[name = tensor("op_499_end_mask_0"), val = tensor([true, false, true, true])]; tensor var_499_cast_fp16 = slice_by_index(begin = var_499_begin_0, end = var_499_end_0, end_mask = var_499_end_mask_0, x = value_1_cast_fp16)[name = tensor("op_499_cast_fp16")]; tensor var_503_begin_0 = const()[name = tensor("op_503_begin_0"), val = tensor([0, 384, 0, 0])]; tensor var_503_end_0 = const()[name = tensor("op_503_end_0"), val = tensor([1, 448, 1, 1500])]; tensor var_503_end_mask_0 = const()[name = tensor("op_503_end_mask_0"), val = tensor([true, false, true, true])]; tensor var_503_cast_fp16 = slice_by_index(begin = var_503_begin_0, end = var_503_end_0, end_mask = var_503_end_mask_0, x = value_1_cast_fp16)[name = tensor("op_503_cast_fp16")]; tensor var_507_begin_0 = const()[name = tensor("op_507_begin_0"), val = tensor([0, 448, 0, 0])]; tensor var_507_end_0 = const()[name = tensor("op_507_end_0"), val = tensor([1, 512, 1, 1500])]; tensor var_507_end_mask_0 = const()[name = tensor("op_507_end_mask_0"), val = tensor([true, false, true, true])]; tensor var_507_cast_fp16 = slice_by_index(begin = var_507_begin_0, end = var_507_end_0, end_mask = var_507_end_mask_0, x = value_1_cast_fp16)[name = tensor("op_507_cast_fp16")]; tensor var_511_equation_0 = const()[name = tensor("op_511_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_511_cast_fp16 = einsum(equation = var_511_equation_0, values = (var_449_cast_fp16, var_227_cast_fp16))[name = tensor("op_511_cast_fp16")]; tensor var_512_to_fp16 = const()[name = tensor("op_512_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_1_cast_fp16 = mul(x = var_511_cast_fp16, y = var_512_to_fp16)[name = tensor("aw_chunk_1_cast_fp16")]; tensor var_515_equation_0 = const()[name = tensor("op_515_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_515_cast_fp16 = einsum(equation = var_515_equation_0, values = (var_449_cast_fp16, var_234_cast_fp16))[name = tensor("op_515_cast_fp16")]; tensor var_516_to_fp16 = const()[name = tensor("op_516_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_3_cast_fp16 = mul(x = var_515_cast_fp16, y = var_516_to_fp16)[name = tensor("aw_chunk_3_cast_fp16")]; tensor var_519_equation_0 = const()[name = tensor("op_519_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_519_cast_fp16 = einsum(equation = var_519_equation_0, values = (var_449_cast_fp16, var_241_cast_fp16))[name = tensor("op_519_cast_fp16")]; tensor var_520_to_fp16 = const()[name = tensor("op_520_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_5_cast_fp16 = mul(x = var_519_cast_fp16, y = var_520_to_fp16)[name = tensor("aw_chunk_5_cast_fp16")]; tensor var_523_equation_0 = const()[name = tensor("op_523_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_523_cast_fp16 = einsum(equation = var_523_equation_0, values = (var_449_cast_fp16, var_248_cast_fp16))[name = tensor("op_523_cast_fp16")]; tensor var_524_to_fp16 = const()[name = tensor("op_524_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_7_cast_fp16 = mul(x = var_523_cast_fp16, y = var_524_to_fp16)[name = tensor("aw_chunk_7_cast_fp16")]; tensor var_527_equation_0 = const()[name = tensor("op_527_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_527_cast_fp16 = einsum(equation = var_527_equation_0, values = (var_453_cast_fp16, var_255_cast_fp16))[name = tensor("op_527_cast_fp16")]; tensor var_528_to_fp16 = const()[name = tensor("op_528_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_9_cast_fp16 = mul(x = var_527_cast_fp16, y = var_528_to_fp16)[name = tensor("aw_chunk_9_cast_fp16")]; tensor var_531_equation_0 = const()[name = tensor("op_531_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_531_cast_fp16 = einsum(equation = var_531_equation_0, values = (var_453_cast_fp16, var_262_cast_fp16))[name = tensor("op_531_cast_fp16")]; tensor var_532_to_fp16 = const()[name = tensor("op_532_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_11_cast_fp16 = mul(x = var_531_cast_fp16, y = var_532_to_fp16)[name = tensor("aw_chunk_11_cast_fp16")]; tensor var_535_equation_0 = const()[name = tensor("op_535_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_535_cast_fp16 = einsum(equation = var_535_equation_0, values = (var_453_cast_fp16, var_269_cast_fp16))[name = tensor("op_535_cast_fp16")]; tensor var_536_to_fp16 = const()[name = tensor("op_536_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_13_cast_fp16 = mul(x = var_535_cast_fp16, y = var_536_to_fp16)[name = tensor("aw_chunk_13_cast_fp16")]; tensor var_539_equation_0 = const()[name = tensor("op_539_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_539_cast_fp16 = einsum(equation = var_539_equation_0, values = (var_453_cast_fp16, var_276_cast_fp16))[name = tensor("op_539_cast_fp16")]; tensor var_540_to_fp16 = const()[name = tensor("op_540_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_15_cast_fp16 = mul(x = var_539_cast_fp16, y = var_540_to_fp16)[name = tensor("aw_chunk_15_cast_fp16")]; tensor var_543_equation_0 = const()[name = tensor("op_543_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_543_cast_fp16 = einsum(equation = var_543_equation_0, values = (var_457_cast_fp16, var_283_cast_fp16))[name = tensor("op_543_cast_fp16")]; tensor var_544_to_fp16 = const()[name = tensor("op_544_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_17_cast_fp16 = mul(x = var_543_cast_fp16, y = var_544_to_fp16)[name = tensor("aw_chunk_17_cast_fp16")]; tensor var_547_equation_0 = const()[name = tensor("op_547_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_547_cast_fp16 = einsum(equation = var_547_equation_0, values = (var_457_cast_fp16, var_290_cast_fp16))[name = tensor("op_547_cast_fp16")]; tensor var_548_to_fp16 = const()[name = tensor("op_548_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_19_cast_fp16 = mul(x = var_547_cast_fp16, y = var_548_to_fp16)[name = tensor("aw_chunk_19_cast_fp16")]; tensor var_551_equation_0 = const()[name = tensor("op_551_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_551_cast_fp16 = einsum(equation = var_551_equation_0, values = (var_457_cast_fp16, var_297_cast_fp16))[name = tensor("op_551_cast_fp16")]; tensor var_552_to_fp16 = const()[name = tensor("op_552_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_21_cast_fp16 = mul(x = var_551_cast_fp16, y = var_552_to_fp16)[name = tensor("aw_chunk_21_cast_fp16")]; tensor var_555_equation_0 = const()[name = tensor("op_555_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_555_cast_fp16 = einsum(equation = var_555_equation_0, values = (var_457_cast_fp16, var_304_cast_fp16))[name = tensor("op_555_cast_fp16")]; tensor var_556_to_fp16 = const()[name = tensor("op_556_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_23_cast_fp16 = mul(x = var_555_cast_fp16, y = var_556_to_fp16)[name = tensor("aw_chunk_23_cast_fp16")]; tensor var_559_equation_0 = const()[name = tensor("op_559_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_559_cast_fp16 = einsum(equation = var_559_equation_0, values = (var_461_cast_fp16, var_311_cast_fp16))[name = tensor("op_559_cast_fp16")]; tensor var_560_to_fp16 = const()[name = tensor("op_560_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_25_cast_fp16 = mul(x = var_559_cast_fp16, y = var_560_to_fp16)[name = tensor("aw_chunk_25_cast_fp16")]; tensor var_563_equation_0 = const()[name = tensor("op_563_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_563_cast_fp16 = einsum(equation = var_563_equation_0, values = (var_461_cast_fp16, var_318_cast_fp16))[name = tensor("op_563_cast_fp16")]; tensor var_564_to_fp16 = const()[name = tensor("op_564_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_27_cast_fp16 = mul(x = var_563_cast_fp16, y = var_564_to_fp16)[name = tensor("aw_chunk_27_cast_fp16")]; tensor var_567_equation_0 = const()[name = tensor("op_567_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_567_cast_fp16 = einsum(equation = var_567_equation_0, values = (var_461_cast_fp16, var_325_cast_fp16))[name = tensor("op_567_cast_fp16")]; tensor var_568_to_fp16 = const()[name = tensor("op_568_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_29_cast_fp16 = mul(x = var_567_cast_fp16, y = var_568_to_fp16)[name = tensor("aw_chunk_29_cast_fp16")]; tensor var_571_equation_0 = const()[name = tensor("op_571_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_571_cast_fp16 = einsum(equation = var_571_equation_0, values = (var_461_cast_fp16, var_332_cast_fp16))[name = tensor("op_571_cast_fp16")]; tensor var_572_to_fp16 = const()[name = tensor("op_572_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_31_cast_fp16 = mul(x = var_571_cast_fp16, y = var_572_to_fp16)[name = tensor("aw_chunk_31_cast_fp16")]; tensor var_575_equation_0 = const()[name = tensor("op_575_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_575_cast_fp16 = einsum(equation = var_575_equation_0, values = (var_465_cast_fp16, var_339_cast_fp16))[name = tensor("op_575_cast_fp16")]; tensor var_576_to_fp16 = const()[name = tensor("op_576_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_33_cast_fp16 = mul(x = var_575_cast_fp16, y = var_576_to_fp16)[name = tensor("aw_chunk_33_cast_fp16")]; tensor var_579_equation_0 = const()[name = tensor("op_579_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_579_cast_fp16 = einsum(equation = var_579_equation_0, values = (var_465_cast_fp16, var_346_cast_fp16))[name = tensor("op_579_cast_fp16")]; tensor var_580_to_fp16 = const()[name = tensor("op_580_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_35_cast_fp16 = mul(x = var_579_cast_fp16, y = var_580_to_fp16)[name = tensor("aw_chunk_35_cast_fp16")]; tensor var_583_equation_0 = const()[name = tensor("op_583_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_583_cast_fp16 = einsum(equation = var_583_equation_0, values = (var_465_cast_fp16, var_353_cast_fp16))[name = tensor("op_583_cast_fp16")]; tensor var_584_to_fp16 = const()[name = tensor("op_584_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_37_cast_fp16 = mul(x = var_583_cast_fp16, y = var_584_to_fp16)[name = tensor("aw_chunk_37_cast_fp16")]; tensor var_587_equation_0 = const()[name = tensor("op_587_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_587_cast_fp16 = einsum(equation = var_587_equation_0, values = (var_465_cast_fp16, var_360_cast_fp16))[name = tensor("op_587_cast_fp16")]; tensor var_588_to_fp16 = const()[name = tensor("op_588_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_39_cast_fp16 = mul(x = var_587_cast_fp16, y = var_588_to_fp16)[name = tensor("aw_chunk_39_cast_fp16")]; tensor var_591_equation_0 = const()[name = tensor("op_591_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_591_cast_fp16 = einsum(equation = var_591_equation_0, values = (var_469_cast_fp16, var_367_cast_fp16))[name = tensor("op_591_cast_fp16")]; tensor var_592_to_fp16 = const()[name = tensor("op_592_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_41_cast_fp16 = mul(x = var_591_cast_fp16, y = var_592_to_fp16)[name = tensor("aw_chunk_41_cast_fp16")]; tensor var_595_equation_0 = const()[name = tensor("op_595_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_595_cast_fp16 = einsum(equation = var_595_equation_0, values = (var_469_cast_fp16, var_374_cast_fp16))[name = tensor("op_595_cast_fp16")]; tensor var_596_to_fp16 = const()[name = tensor("op_596_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_43_cast_fp16 = mul(x = var_595_cast_fp16, y = var_596_to_fp16)[name = tensor("aw_chunk_43_cast_fp16")]; tensor var_599_equation_0 = const()[name = tensor("op_599_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_599_cast_fp16 = einsum(equation = var_599_equation_0, values = (var_469_cast_fp16, var_381_cast_fp16))[name = tensor("op_599_cast_fp16")]; tensor var_600_to_fp16 = const()[name = tensor("op_600_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_45_cast_fp16 = mul(x = var_599_cast_fp16, y = var_600_to_fp16)[name = tensor("aw_chunk_45_cast_fp16")]; tensor var_603_equation_0 = const()[name = tensor("op_603_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_603_cast_fp16 = einsum(equation = var_603_equation_0, values = (var_469_cast_fp16, var_388_cast_fp16))[name = tensor("op_603_cast_fp16")]; tensor var_604_to_fp16 = const()[name = tensor("op_604_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_47_cast_fp16 = mul(x = var_603_cast_fp16, y = var_604_to_fp16)[name = tensor("aw_chunk_47_cast_fp16")]; tensor var_607_equation_0 = const()[name = tensor("op_607_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_607_cast_fp16 = einsum(equation = var_607_equation_0, values = (var_473_cast_fp16, var_395_cast_fp16))[name = tensor("op_607_cast_fp16")]; tensor var_608_to_fp16 = const()[name = tensor("op_608_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_49_cast_fp16 = mul(x = var_607_cast_fp16, y = var_608_to_fp16)[name = tensor("aw_chunk_49_cast_fp16")]; tensor var_611_equation_0 = const()[name = tensor("op_611_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_611_cast_fp16 = einsum(equation = var_611_equation_0, values = (var_473_cast_fp16, var_402_cast_fp16))[name = tensor("op_611_cast_fp16")]; tensor var_612_to_fp16 = const()[name = tensor("op_612_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_51_cast_fp16 = mul(x = var_611_cast_fp16, y = var_612_to_fp16)[name = tensor("aw_chunk_51_cast_fp16")]; tensor var_615_equation_0 = const()[name = tensor("op_615_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_615_cast_fp16 = einsum(equation = var_615_equation_0, values = (var_473_cast_fp16, var_409_cast_fp16))[name = tensor("op_615_cast_fp16")]; tensor var_616_to_fp16 = const()[name = tensor("op_616_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_53_cast_fp16 = mul(x = var_615_cast_fp16, y = var_616_to_fp16)[name = tensor("aw_chunk_53_cast_fp16")]; tensor var_619_equation_0 = const()[name = tensor("op_619_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_619_cast_fp16 = einsum(equation = var_619_equation_0, values = (var_473_cast_fp16, var_416_cast_fp16))[name = tensor("op_619_cast_fp16")]; tensor var_620_to_fp16 = const()[name = tensor("op_620_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_55_cast_fp16 = mul(x = var_619_cast_fp16, y = var_620_to_fp16)[name = tensor("aw_chunk_55_cast_fp16")]; tensor var_623_equation_0 = const()[name = tensor("op_623_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_623_cast_fp16 = einsum(equation = var_623_equation_0, values = (var_477_cast_fp16, var_423_cast_fp16))[name = tensor("op_623_cast_fp16")]; tensor var_624_to_fp16 = const()[name = tensor("op_624_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_57_cast_fp16 = mul(x = var_623_cast_fp16, y = var_624_to_fp16)[name = tensor("aw_chunk_57_cast_fp16")]; tensor var_627_equation_0 = const()[name = tensor("op_627_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_627_cast_fp16 = einsum(equation = var_627_equation_0, values = (var_477_cast_fp16, var_430_cast_fp16))[name = tensor("op_627_cast_fp16")]; tensor var_628_to_fp16 = const()[name = tensor("op_628_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_59_cast_fp16 = mul(x = var_627_cast_fp16, y = var_628_to_fp16)[name = tensor("aw_chunk_59_cast_fp16")]; tensor var_631_equation_0 = const()[name = tensor("op_631_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_631_cast_fp16 = einsum(equation = var_631_equation_0, values = (var_477_cast_fp16, var_437_cast_fp16))[name = tensor("op_631_cast_fp16")]; tensor var_632_to_fp16 = const()[name = tensor("op_632_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_61_cast_fp16 = mul(x = var_631_cast_fp16, y = var_632_to_fp16)[name = tensor("aw_chunk_61_cast_fp16")]; tensor var_635_equation_0 = const()[name = tensor("op_635_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_635_cast_fp16 = einsum(equation = var_635_equation_0, values = (var_477_cast_fp16, var_444_cast_fp16))[name = tensor("op_635_cast_fp16")]; tensor var_636_to_fp16 = const()[name = tensor("op_636_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_63_cast_fp16 = mul(x = var_635_cast_fp16, y = var_636_to_fp16)[name = tensor("aw_chunk_63_cast_fp16")]; tensor var_638_cast_fp16 = softmax(axis = var_135, x = aw_chunk_1_cast_fp16)[name = tensor("op_638_cast_fp16")]; tensor var_639_cast_fp16 = softmax(axis = var_135, x = aw_chunk_3_cast_fp16)[name = tensor("op_639_cast_fp16")]; tensor var_640_cast_fp16 = softmax(axis = var_135, x = aw_chunk_5_cast_fp16)[name = tensor("op_640_cast_fp16")]; tensor var_641_cast_fp16 = softmax(axis = var_135, x = aw_chunk_7_cast_fp16)[name = tensor("op_641_cast_fp16")]; tensor var_642_cast_fp16 = softmax(axis = var_135, x = aw_chunk_9_cast_fp16)[name = tensor("op_642_cast_fp16")]; tensor var_643_cast_fp16 = softmax(axis = var_135, x = aw_chunk_11_cast_fp16)[name = tensor("op_643_cast_fp16")]; tensor var_644_cast_fp16 = softmax(axis = var_135, x = aw_chunk_13_cast_fp16)[name = tensor("op_644_cast_fp16")]; tensor var_645_cast_fp16 = softmax(axis = var_135, x = aw_chunk_15_cast_fp16)[name = tensor("op_645_cast_fp16")]; tensor var_646_cast_fp16 = softmax(axis = var_135, x = aw_chunk_17_cast_fp16)[name = tensor("op_646_cast_fp16")]; tensor var_647_cast_fp16 = softmax(axis = var_135, x = aw_chunk_19_cast_fp16)[name = tensor("op_647_cast_fp16")]; tensor var_648_cast_fp16 = softmax(axis = var_135, x = aw_chunk_21_cast_fp16)[name = tensor("op_648_cast_fp16")]; tensor var_649_cast_fp16 = softmax(axis = var_135, x = aw_chunk_23_cast_fp16)[name = tensor("op_649_cast_fp16")]; tensor var_650_cast_fp16 = softmax(axis = var_135, x = aw_chunk_25_cast_fp16)[name = tensor("op_650_cast_fp16")]; tensor var_651_cast_fp16 = softmax(axis = var_135, x = aw_chunk_27_cast_fp16)[name = tensor("op_651_cast_fp16")]; tensor var_652_cast_fp16 = softmax(axis = var_135, x = aw_chunk_29_cast_fp16)[name = tensor("op_652_cast_fp16")]; tensor var_653_cast_fp16 = softmax(axis = var_135, x = aw_chunk_31_cast_fp16)[name = tensor("op_653_cast_fp16")]; tensor var_654_cast_fp16 = softmax(axis = var_135, x = aw_chunk_33_cast_fp16)[name = tensor("op_654_cast_fp16")]; tensor var_655_cast_fp16 = softmax(axis = var_135, x = aw_chunk_35_cast_fp16)[name = tensor("op_655_cast_fp16")]; tensor var_656_cast_fp16 = softmax(axis = var_135, x = aw_chunk_37_cast_fp16)[name = tensor("op_656_cast_fp16")]; tensor var_657_cast_fp16 = softmax(axis = var_135, x = aw_chunk_39_cast_fp16)[name = tensor("op_657_cast_fp16")]; tensor var_658_cast_fp16 = softmax(axis = var_135, x = aw_chunk_41_cast_fp16)[name = tensor("op_658_cast_fp16")]; tensor var_659_cast_fp16 = softmax(axis = var_135, x = aw_chunk_43_cast_fp16)[name = tensor("op_659_cast_fp16")]; tensor var_660_cast_fp16 = softmax(axis = var_135, x = aw_chunk_45_cast_fp16)[name = tensor("op_660_cast_fp16")]; tensor var_661_cast_fp16 = softmax(axis = var_135, x = aw_chunk_47_cast_fp16)[name = tensor("op_661_cast_fp16")]; tensor var_662_cast_fp16 = softmax(axis = var_135, x = aw_chunk_49_cast_fp16)[name = tensor("op_662_cast_fp16")]; tensor var_663_cast_fp16 = softmax(axis = var_135, x = aw_chunk_51_cast_fp16)[name = tensor("op_663_cast_fp16")]; tensor var_664_cast_fp16 = softmax(axis = var_135, x = aw_chunk_53_cast_fp16)[name = tensor("op_664_cast_fp16")]; tensor var_665_cast_fp16 = softmax(axis = var_135, x = aw_chunk_55_cast_fp16)[name = tensor("op_665_cast_fp16")]; tensor var_666_cast_fp16 = softmax(axis = var_135, x = aw_chunk_57_cast_fp16)[name = tensor("op_666_cast_fp16")]; tensor var_667_cast_fp16 = softmax(axis = var_135, x = aw_chunk_59_cast_fp16)[name = tensor("op_667_cast_fp16")]; tensor var_668_cast_fp16 = softmax(axis = var_135, x = aw_chunk_61_cast_fp16)[name = tensor("op_668_cast_fp16")]; tensor var_669_cast_fp16 = softmax(axis = var_135, x = aw_chunk_63_cast_fp16)[name = tensor("op_669_cast_fp16")]; tensor var_671_equation_0 = const()[name = tensor("op_671_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_671_cast_fp16 = einsum(equation = var_671_equation_0, values = (var_479_cast_fp16, var_638_cast_fp16))[name = tensor("op_671_cast_fp16")]; tensor var_673_equation_0 = const()[name = tensor("op_673_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_673_cast_fp16 = einsum(equation = var_673_equation_0, values = (var_479_cast_fp16, var_639_cast_fp16))[name = tensor("op_673_cast_fp16")]; tensor var_675_equation_0 = const()[name = tensor("op_675_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_675_cast_fp16 = einsum(equation = var_675_equation_0, values = (var_479_cast_fp16, var_640_cast_fp16))[name = tensor("op_675_cast_fp16")]; tensor var_677_equation_0 = const()[name = tensor("op_677_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_677_cast_fp16 = einsum(equation = var_677_equation_0, values = (var_479_cast_fp16, var_641_cast_fp16))[name = tensor("op_677_cast_fp16")]; tensor var_679_equation_0 = const()[name = tensor("op_679_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_679_cast_fp16 = einsum(equation = var_679_equation_0, values = (var_483_cast_fp16, var_642_cast_fp16))[name = tensor("op_679_cast_fp16")]; tensor var_681_equation_0 = const()[name = tensor("op_681_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_681_cast_fp16 = einsum(equation = var_681_equation_0, values = (var_483_cast_fp16, var_643_cast_fp16))[name = tensor("op_681_cast_fp16")]; tensor var_683_equation_0 = const()[name = tensor("op_683_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_683_cast_fp16 = einsum(equation = var_683_equation_0, values = (var_483_cast_fp16, var_644_cast_fp16))[name = tensor("op_683_cast_fp16")]; tensor var_685_equation_0 = const()[name = tensor("op_685_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_685_cast_fp16 = einsum(equation = var_685_equation_0, values = (var_483_cast_fp16, var_645_cast_fp16))[name = tensor("op_685_cast_fp16")]; tensor var_687_equation_0 = const()[name = tensor("op_687_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_687_cast_fp16 = einsum(equation = var_687_equation_0, values = (var_487_cast_fp16, var_646_cast_fp16))[name = tensor("op_687_cast_fp16")]; tensor var_689_equation_0 = const()[name = tensor("op_689_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_689_cast_fp16 = einsum(equation = var_689_equation_0, values = (var_487_cast_fp16, var_647_cast_fp16))[name = tensor("op_689_cast_fp16")]; tensor var_691_equation_0 = const()[name = tensor("op_691_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_691_cast_fp16 = einsum(equation = var_691_equation_0, values = (var_487_cast_fp16, var_648_cast_fp16))[name = tensor("op_691_cast_fp16")]; tensor var_693_equation_0 = const()[name = tensor("op_693_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_693_cast_fp16 = einsum(equation = var_693_equation_0, values = (var_487_cast_fp16, var_649_cast_fp16))[name = tensor("op_693_cast_fp16")]; tensor var_695_equation_0 = const()[name = tensor("op_695_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_695_cast_fp16 = einsum(equation = var_695_equation_0, values = (var_491_cast_fp16, var_650_cast_fp16))[name = tensor("op_695_cast_fp16")]; tensor var_697_equation_0 = const()[name = tensor("op_697_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_697_cast_fp16 = einsum(equation = var_697_equation_0, values = (var_491_cast_fp16, var_651_cast_fp16))[name = tensor("op_697_cast_fp16")]; tensor var_699_equation_0 = const()[name = tensor("op_699_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_699_cast_fp16 = einsum(equation = var_699_equation_0, values = (var_491_cast_fp16, var_652_cast_fp16))[name = tensor("op_699_cast_fp16")]; tensor var_701_equation_0 = const()[name = tensor("op_701_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_701_cast_fp16 = einsum(equation = var_701_equation_0, values = (var_491_cast_fp16, var_653_cast_fp16))[name = tensor("op_701_cast_fp16")]; tensor var_703_equation_0 = const()[name = tensor("op_703_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_703_cast_fp16 = einsum(equation = var_703_equation_0, values = (var_495_cast_fp16, var_654_cast_fp16))[name = tensor("op_703_cast_fp16")]; tensor var_705_equation_0 = const()[name = tensor("op_705_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_705_cast_fp16 = einsum(equation = var_705_equation_0, values = (var_495_cast_fp16, var_655_cast_fp16))[name = tensor("op_705_cast_fp16")]; tensor var_707_equation_0 = const()[name = tensor("op_707_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_707_cast_fp16 = einsum(equation = var_707_equation_0, values = (var_495_cast_fp16, var_656_cast_fp16))[name = tensor("op_707_cast_fp16")]; tensor var_709_equation_0 = const()[name = tensor("op_709_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_709_cast_fp16 = einsum(equation = var_709_equation_0, values = (var_495_cast_fp16, var_657_cast_fp16))[name = tensor("op_709_cast_fp16")]; tensor var_711_equation_0 = const()[name = tensor("op_711_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_711_cast_fp16 = einsum(equation = var_711_equation_0, values = (var_499_cast_fp16, var_658_cast_fp16))[name = tensor("op_711_cast_fp16")]; tensor var_713_equation_0 = const()[name = tensor("op_713_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_713_cast_fp16 = einsum(equation = var_713_equation_0, values = (var_499_cast_fp16, var_659_cast_fp16))[name = tensor("op_713_cast_fp16")]; tensor var_715_equation_0 = const()[name = tensor("op_715_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_715_cast_fp16 = einsum(equation = var_715_equation_0, values = (var_499_cast_fp16, var_660_cast_fp16))[name = tensor("op_715_cast_fp16")]; tensor var_717_equation_0 = const()[name = tensor("op_717_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_717_cast_fp16 = einsum(equation = var_717_equation_0, values = (var_499_cast_fp16, var_661_cast_fp16))[name = tensor("op_717_cast_fp16")]; tensor var_719_equation_0 = const()[name = tensor("op_719_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_719_cast_fp16 = einsum(equation = var_719_equation_0, values = (var_503_cast_fp16, var_662_cast_fp16))[name = tensor("op_719_cast_fp16")]; tensor var_721_equation_0 = const()[name = tensor("op_721_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_721_cast_fp16 = einsum(equation = var_721_equation_0, values = (var_503_cast_fp16, var_663_cast_fp16))[name = tensor("op_721_cast_fp16")]; tensor var_723_equation_0 = const()[name = tensor("op_723_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_723_cast_fp16 = einsum(equation = var_723_equation_0, values = (var_503_cast_fp16, var_664_cast_fp16))[name = tensor("op_723_cast_fp16")]; tensor var_725_equation_0 = const()[name = tensor("op_725_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_725_cast_fp16 = einsum(equation = var_725_equation_0, values = (var_503_cast_fp16, var_665_cast_fp16))[name = tensor("op_725_cast_fp16")]; tensor var_727_equation_0 = const()[name = tensor("op_727_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_727_cast_fp16 = einsum(equation = var_727_equation_0, values = (var_507_cast_fp16, var_666_cast_fp16))[name = tensor("op_727_cast_fp16")]; tensor var_729_equation_0 = const()[name = tensor("op_729_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_729_cast_fp16 = einsum(equation = var_729_equation_0, values = (var_507_cast_fp16, var_667_cast_fp16))[name = tensor("op_729_cast_fp16")]; tensor var_731_equation_0 = const()[name = tensor("op_731_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_731_cast_fp16 = einsum(equation = var_731_equation_0, values = (var_507_cast_fp16, var_668_cast_fp16))[name = tensor("op_731_cast_fp16")]; tensor var_733_equation_0 = const()[name = tensor("op_733_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_733_cast_fp16 = einsum(equation = var_733_equation_0, values = (var_507_cast_fp16, var_669_cast_fp16))[name = tensor("op_733_cast_fp16")]; tensor var_735_interleave_0 = const()[name = tensor("op_735_interleave_0"), val = tensor(false)]; tensor var_735_cast_fp16 = concat(axis = var_122, interleave = var_735_interleave_0, values = (var_671_cast_fp16, var_673_cast_fp16, var_675_cast_fp16, var_677_cast_fp16))[name = tensor("op_735_cast_fp16")]; tensor var_737_interleave_0 = const()[name = tensor("op_737_interleave_0"), val = tensor(false)]; tensor var_737_cast_fp16 = concat(axis = var_122, interleave = var_737_interleave_0, values = (var_679_cast_fp16, var_681_cast_fp16, var_683_cast_fp16, var_685_cast_fp16))[name = tensor("op_737_cast_fp16")]; tensor var_739_interleave_0 = const()[name = tensor("op_739_interleave_0"), val = tensor(false)]; tensor var_739_cast_fp16 = concat(axis = var_122, interleave = var_739_interleave_0, values = (var_687_cast_fp16, var_689_cast_fp16, var_691_cast_fp16, var_693_cast_fp16))[name = tensor("op_739_cast_fp16")]; tensor var_741_interleave_0 = const()[name = tensor("op_741_interleave_0"), val = tensor(false)]; tensor var_741_cast_fp16 = concat(axis = var_122, interleave = var_741_interleave_0, values = (var_695_cast_fp16, var_697_cast_fp16, var_699_cast_fp16, var_701_cast_fp16))[name = tensor("op_741_cast_fp16")]; tensor var_743_interleave_0 = const()[name = tensor("op_743_interleave_0"), val = tensor(false)]; tensor var_743_cast_fp16 = concat(axis = var_122, interleave = var_743_interleave_0, values = (var_703_cast_fp16, var_705_cast_fp16, var_707_cast_fp16, var_709_cast_fp16))[name = tensor("op_743_cast_fp16")]; tensor var_745_interleave_0 = const()[name = tensor("op_745_interleave_0"), val = tensor(false)]; tensor var_745_cast_fp16 = concat(axis = var_122, interleave = var_745_interleave_0, values = (var_711_cast_fp16, var_713_cast_fp16, var_715_cast_fp16, var_717_cast_fp16))[name = tensor("op_745_cast_fp16")]; tensor var_747_interleave_0 = const()[name = tensor("op_747_interleave_0"), val = tensor(false)]; tensor var_747_cast_fp16 = concat(axis = var_122, interleave = var_747_interleave_0, values = (var_719_cast_fp16, var_721_cast_fp16, var_723_cast_fp16, var_725_cast_fp16))[name = tensor("op_747_cast_fp16")]; tensor var_749_interleave_0 = const()[name = tensor("op_749_interleave_0"), val = tensor(false)]; tensor var_749_cast_fp16 = concat(axis = var_122, interleave = var_749_interleave_0, values = (var_727_cast_fp16, var_729_cast_fp16, var_731_cast_fp16, var_733_cast_fp16))[name = tensor("op_749_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_135, interleave = input_1_interleave_0, values = (var_735_cast_fp16, var_737_cast_fp16, var_739_cast_fp16, var_741_cast_fp16, var_743_cast_fp16, var_745_cast_fp16, var_747_cast_fp16, var_749_cast_fp16))[name = tensor("input_1_cast_fp16")]; tensor var_754 = const()[name = tensor("op_754"), val = tensor([1, 1])]; tensor var_756 = const()[name = tensor("op_756"), 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(4936640)))]; 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(5460992)))]; tensor obj_3_cast_fp16 = conv(bias = layers_0_self_attn_o_proj_bias_to_fp16, dilations = var_756, groups = var_135, pad = obj_3_pad_0, pad_type = obj_3_pad_type_0, strides = var_754, 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_762 = const()[name = tensor("op_762"), val = tensor([1])]; tensor channels_mean_3_cast_fp16 = reduce_mean(axes = var_762, keep_dims = var_136, 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_766 = const()[name = tensor("op_766"), val = tensor([1])]; tensor var_767_cast_fp16 = reduce_mean(axes = var_766, keep_dims = var_136, x = zero_mean_sq_3_cast_fp16)[name = tensor("op_767_cast_fp16")]; tensor var_768_to_fp16 = const()[name = tensor("op_768_to_fp16"), val = tensor(0x1.5p-17)]; tensor var_769_cast_fp16 = add(x = var_767_cast_fp16, y = var_768_to_fp16)[name = tensor("op_769_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_769_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(5462080)))]; 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(5463168)))]; 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_780 = const()[name = tensor("op_780"), val = tensor([1, 1])]; tensor var_782 = const()[name = tensor("op_782"), 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(5464256)))]; 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(7561472)))]; tensor input_5_cast_fp16 = conv(bias = layers_0_fc1_bias_to_fp16, dilations = var_782, groups = var_135, pad = input_5_pad_0, pad_type = input_5_pad_type_0, strides = var_780, 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_788 = const()[name = tensor("op_788"), val = tensor([1, 1])]; tensor var_790 = const()[name = tensor("op_790"), 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(7565632)))]; 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(9662848)))]; tensor hidden_states_5_cast_fp16 = conv(bias = layers_0_fc2_bias_to_fp16, dilations = var_790, groups = var_135, pad = hidden_states_5_pad_0, pad_type = hidden_states_5_pad_type_0, strides = var_788, 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_797 = const()[name = tensor("op_797"), val = tensor(3)]; tensor var_810 = const()[name = tensor("op_810"), val = tensor(1)]; tensor var_811 = const()[name = tensor("op_811"), val = tensor(true)]; tensor var_821 = const()[name = tensor("op_821"), val = tensor([1])]; tensor channels_mean_5_cast_fp16 = reduce_mean(axes = var_821, keep_dims = var_811, 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_825 = const()[name = tensor("op_825"), val = tensor([1])]; tensor var_826_cast_fp16 = reduce_mean(axes = var_825, keep_dims = var_811, x = zero_mean_sq_5_cast_fp16)[name = tensor("op_826_cast_fp16")]; tensor var_827_to_fp16 = const()[name = tensor("op_827_to_fp16"), val = tensor(0x1.5p-17)]; tensor var_828_cast_fp16 = add(x = var_826_cast_fp16, y = var_827_to_fp16)[name = tensor("op_828_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_828_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(9663936)))]; 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(9665024)))]; 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_843 = const()[name = tensor("op_843"), val = tensor([1, 1])]; tensor var_845 = const()[name = tensor("op_845"), 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(9666112)))]; 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(10190464)))]; tensor query_3_cast_fp16 = conv(bias = layers_1_self_attn_q_proj_bias_to_fp16, dilations = var_845, groups = var_810, pad = query_3_pad_0, pad_type = query_3_pad_type_0, strides = var_843, weight = layers_1_self_attn_q_proj_weight_to_fp16, x = obj_5_cast_fp16)[name = tensor("query_3_cast_fp16")]; tensor var_849 = const()[name = tensor("op_849"), val = tensor([1, 1])]; tensor var_851 = const()[name = tensor("op_851"), 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(10191552)))]; tensor key_3_cast_fp16 = conv(dilations = var_851, groups = var_810, pad = key_3_pad_0, pad_type = key_3_pad_type_0, strides = var_849, weight = layers_1_self_attn_k_proj_weight_to_fp16, x = obj_5_cast_fp16)[name = tensor("key_3_cast_fp16")]; tensor var_856 = const()[name = tensor("op_856"), val = tensor([1, 1])]; tensor var_858 = const()[name = tensor("op_858"), 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(10715904)))]; 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(11240256)))]; tensor value_3_cast_fp16 = conv(bias = layers_1_self_attn_v_proj_bias_to_fp16, dilations = var_858, groups = var_810, pad = value_3_pad_0, pad_type = value_3_pad_type_0, strides = var_856, weight = layers_1_self_attn_v_proj_weight_to_fp16, x = obj_5_cast_fp16)[name = tensor("value_3_cast_fp16")]; tensor var_865_begin_0 = const()[name = tensor("op_865_begin_0"), val = tensor([0, 0, 0, 0])]; tensor var_865_end_0 = const()[name = tensor("op_865_end_0"), val = tensor([1, 64, 1, 1500])]; tensor var_865_end_mask_0 = const()[name = tensor("op_865_end_mask_0"), val = tensor([true, false, true, true])]; 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 = query_3_cast_fp16)[name = tensor("op_865_cast_fp16")]; tensor var_869_begin_0 = const()[name = tensor("op_869_begin_0"), val = tensor([0, 64, 0, 0])]; tensor var_869_end_0 = const()[name = tensor("op_869_end_0"), val = tensor([1, 128, 1, 1500])]; tensor var_869_end_mask_0 = const()[name = tensor("op_869_end_mask_0"), val = tensor([true, false, true, true])]; tensor var_869_cast_fp16 = slice_by_index(begin = var_869_begin_0, end = var_869_end_0, end_mask = var_869_end_mask_0, x = query_3_cast_fp16)[name = tensor("op_869_cast_fp16")]; tensor var_873_begin_0 = const()[name = tensor("op_873_begin_0"), val = tensor([0, 128, 0, 0])]; tensor var_873_end_0 = const()[name = tensor("op_873_end_0"), val = tensor([1, 192, 1, 1500])]; tensor var_873_end_mask_0 = const()[name = tensor("op_873_end_mask_0"), val = tensor([true, false, true, true])]; tensor var_873_cast_fp16 = slice_by_index(begin = var_873_begin_0, end = var_873_end_0, end_mask = var_873_end_mask_0, x = query_3_cast_fp16)[name = tensor("op_873_cast_fp16")]; tensor var_877_begin_0 = const()[name = tensor("op_877_begin_0"), val = tensor([0, 192, 0, 0])]; tensor var_877_end_0 = const()[name = tensor("op_877_end_0"), val = tensor([1, 256, 1, 1500])]; tensor var_877_end_mask_0 = const()[name = tensor("op_877_end_mask_0"), val = tensor([true, false, true, true])]; tensor var_877_cast_fp16 = slice_by_index(begin = var_877_begin_0, end = var_877_end_0, end_mask = var_877_end_mask_0, x = query_3_cast_fp16)[name = tensor("op_877_cast_fp16")]; tensor var_881_begin_0 = const()[name = tensor("op_881_begin_0"), val = tensor([0, 256, 0, 0])]; tensor var_881_end_0 = const()[name = tensor("op_881_end_0"), val = tensor([1, 320, 1, 1500])]; tensor var_881_end_mask_0 = const()[name = tensor("op_881_end_mask_0"), val = tensor([true, false, true, true])]; tensor var_881_cast_fp16 = slice_by_index(begin = var_881_begin_0, end = var_881_end_0, end_mask = var_881_end_mask_0, x = query_3_cast_fp16)[name = tensor("op_881_cast_fp16")]; tensor var_885_begin_0 = const()[name = tensor("op_885_begin_0"), val = tensor([0, 320, 0, 0])]; tensor var_885_end_0 = const()[name = tensor("op_885_end_0"), val = tensor([1, 384, 1, 1500])]; tensor var_885_end_mask_0 = const()[name = tensor("op_885_end_mask_0"), val = tensor([true, false, true, true])]; tensor var_885_cast_fp16 = slice_by_index(begin = var_885_begin_0, end = var_885_end_0, end_mask = var_885_end_mask_0, x = query_3_cast_fp16)[name = tensor("op_885_cast_fp16")]; tensor var_889_begin_0 = const()[name = tensor("op_889_begin_0"), val = tensor([0, 384, 0, 0])]; tensor var_889_end_0 = const()[name = tensor("op_889_end_0"), val = tensor([1, 448, 1, 1500])]; tensor var_889_end_mask_0 = const()[name = tensor("op_889_end_mask_0"), val = tensor([true, false, true, true])]; tensor var_889_cast_fp16 = slice_by_index(begin = var_889_begin_0, end = var_889_end_0, end_mask = var_889_end_mask_0, x = query_3_cast_fp16)[name = tensor("op_889_cast_fp16")]; tensor var_893_begin_0 = const()[name = tensor("op_893_begin_0"), val = tensor([0, 448, 0, 0])]; tensor var_893_end_0 = const()[name = tensor("op_893_end_0"), val = tensor([1, 512, 1, 1500])]; tensor var_893_end_mask_0 = const()[name = tensor("op_893_end_mask_0"), val = tensor([true, false, true, true])]; 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 = query_3_cast_fp16)[name = tensor("op_893_cast_fp16")]; tensor var_902_begin_0 = const()[name = tensor("op_902_begin_0"), val = tensor([0, 0, 0, 0])]; tensor var_902_end_0 = const()[name = tensor("op_902_end_0"), val = tensor([1, 64, 1, 375])]; tensor var_902_end_mask_0 = const()[name = tensor("op_902_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_902_cast_fp16 = slice_by_index(begin = var_902_begin_0, end = var_902_end_0, end_mask = var_902_end_mask_0, x = var_865_cast_fp16)[name = tensor("op_902_cast_fp16")]; tensor var_909_begin_0 = const()[name = tensor("op_909_begin_0"), val = tensor([0, 0, 0, 375])]; tensor var_909_end_0 = const()[name = tensor("op_909_end_0"), val = tensor([1, 64, 1, 750])]; tensor var_909_end_mask_0 = const()[name = tensor("op_909_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_909_cast_fp16 = slice_by_index(begin = var_909_begin_0, end = var_909_end_0, end_mask = var_909_end_mask_0, x = var_865_cast_fp16)[name = tensor("op_909_cast_fp16")]; tensor var_916_begin_0 = const()[name = tensor("op_916_begin_0"), val = tensor([0, 0, 0, 750])]; tensor var_916_end_0 = const()[name = tensor("op_916_end_0"), val = tensor([1, 64, 1, 1125])]; 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 = var_865_cast_fp16)[name = tensor("op_916_cast_fp16")]; tensor var_923_begin_0 = const()[name = tensor("op_923_begin_0"), val = tensor([0, 0, 0, 1125])]; tensor var_923_end_0 = const()[name = tensor("op_923_end_0"), val = tensor([1, 64, 1, 1500])]; tensor var_923_end_mask_0 = const()[name = tensor("op_923_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_923_cast_fp16 = slice_by_index(begin = var_923_begin_0, end = var_923_end_0, end_mask = var_923_end_mask_0, x = var_865_cast_fp16)[name = tensor("op_923_cast_fp16")]; tensor var_930_begin_0 = const()[name = tensor("op_930_begin_0"), val = tensor([0, 0, 0, 0])]; tensor var_930_end_0 = const()[name = tensor("op_930_end_0"), val = tensor([1, 64, 1, 375])]; tensor var_930_end_mask_0 = const()[name = tensor("op_930_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_930_cast_fp16 = slice_by_index(begin = var_930_begin_0, end = var_930_end_0, end_mask = var_930_end_mask_0, x = var_869_cast_fp16)[name = tensor("op_930_cast_fp16")]; tensor var_937_begin_0 = const()[name = tensor("op_937_begin_0"), val = tensor([0, 0, 0, 375])]; tensor var_937_end_0 = const()[name = tensor("op_937_end_0"), val = tensor([1, 64, 1, 750])]; tensor var_937_end_mask_0 = const()[name = tensor("op_937_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_937_cast_fp16 = slice_by_index(begin = var_937_begin_0, end = var_937_end_0, end_mask = var_937_end_mask_0, x = var_869_cast_fp16)[name = tensor("op_937_cast_fp16")]; tensor var_944_begin_0 = const()[name = tensor("op_944_begin_0"), val = tensor([0, 0, 0, 750])]; tensor var_944_end_0 = const()[name = tensor("op_944_end_0"), val = tensor([1, 64, 1, 1125])]; tensor var_944_end_mask_0 = const()[name = tensor("op_944_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_944_cast_fp16 = slice_by_index(begin = var_944_begin_0, end = var_944_end_0, end_mask = var_944_end_mask_0, x = var_869_cast_fp16)[name = tensor("op_944_cast_fp16")]; tensor var_951_begin_0 = const()[name = tensor("op_951_begin_0"), val = tensor([0, 0, 0, 1125])]; tensor var_951_end_0 = const()[name = tensor("op_951_end_0"), val = tensor([1, 64, 1, 1500])]; tensor var_951_end_mask_0 = const()[name = tensor("op_951_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_951_cast_fp16 = slice_by_index(begin = var_951_begin_0, end = var_951_end_0, end_mask = var_951_end_mask_0, x = var_869_cast_fp16)[name = tensor("op_951_cast_fp16")]; tensor var_958_begin_0 = const()[name = tensor("op_958_begin_0"), val = tensor([0, 0, 0, 0])]; tensor var_958_end_0 = const()[name = tensor("op_958_end_0"), val = tensor([1, 64, 1, 375])]; tensor var_958_end_mask_0 = const()[name = tensor("op_958_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_958_cast_fp16 = slice_by_index(begin = var_958_begin_0, end = var_958_end_0, end_mask = var_958_end_mask_0, x = var_873_cast_fp16)[name = tensor("op_958_cast_fp16")]; tensor var_965_begin_0 = const()[name = tensor("op_965_begin_0"), val = tensor([0, 0, 0, 375])]; tensor var_965_end_0 = const()[name = tensor("op_965_end_0"), val = tensor([1, 64, 1, 750])]; tensor var_965_end_mask_0 = const()[name = tensor("op_965_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_965_cast_fp16 = slice_by_index(begin = var_965_begin_0, end = var_965_end_0, end_mask = var_965_end_mask_0, x = var_873_cast_fp16)[name = tensor("op_965_cast_fp16")]; tensor var_972_begin_0 = const()[name = tensor("op_972_begin_0"), val = tensor([0, 0, 0, 750])]; tensor var_972_end_0 = const()[name = tensor("op_972_end_0"), val = tensor([1, 64, 1, 1125])]; tensor var_972_end_mask_0 = const()[name = tensor("op_972_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_972_cast_fp16 = slice_by_index(begin = var_972_begin_0, end = var_972_end_0, end_mask = var_972_end_mask_0, x = var_873_cast_fp16)[name = tensor("op_972_cast_fp16")]; tensor var_979_begin_0 = const()[name = tensor("op_979_begin_0"), val = tensor([0, 0, 0, 1125])]; tensor var_979_end_0 = const()[name = tensor("op_979_end_0"), val = tensor([1, 64, 1, 1500])]; tensor var_979_end_mask_0 = const()[name = tensor("op_979_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_979_cast_fp16 = slice_by_index(begin = var_979_begin_0, end = var_979_end_0, end_mask = var_979_end_mask_0, x = var_873_cast_fp16)[name = tensor("op_979_cast_fp16")]; tensor var_986_begin_0 = const()[name = tensor("op_986_begin_0"), val = tensor([0, 0, 0, 0])]; tensor var_986_end_0 = const()[name = tensor("op_986_end_0"), val = tensor([1, 64, 1, 375])]; tensor var_986_end_mask_0 = const()[name = tensor("op_986_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_986_cast_fp16 = slice_by_index(begin = var_986_begin_0, end = var_986_end_0, end_mask = var_986_end_mask_0, x = var_877_cast_fp16)[name = tensor("op_986_cast_fp16")]; tensor var_993_begin_0 = const()[name = tensor("op_993_begin_0"), val = tensor([0, 0, 0, 375])]; tensor var_993_end_0 = const()[name = tensor("op_993_end_0"), val = tensor([1, 64, 1, 750])]; tensor var_993_end_mask_0 = const()[name = tensor("op_993_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_993_cast_fp16 = slice_by_index(begin = var_993_begin_0, end = var_993_end_0, end_mask = var_993_end_mask_0, x = var_877_cast_fp16)[name = tensor("op_993_cast_fp16")]; tensor var_1000_begin_0 = const()[name = tensor("op_1000_begin_0"), val = tensor([0, 0, 0, 750])]; tensor var_1000_end_0 = const()[name = tensor("op_1000_end_0"), val = tensor([1, 64, 1, 1125])]; tensor var_1000_end_mask_0 = const()[name = tensor("op_1000_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_1000_cast_fp16 = slice_by_index(begin = var_1000_begin_0, end = var_1000_end_0, end_mask = var_1000_end_mask_0, x = var_877_cast_fp16)[name = tensor("op_1000_cast_fp16")]; tensor var_1007_begin_0 = const()[name = tensor("op_1007_begin_0"), val = tensor([0, 0, 0, 1125])]; tensor var_1007_end_0 = const()[name = tensor("op_1007_end_0"), val = tensor([1, 64, 1, 1500])]; tensor var_1007_end_mask_0 = const()[name = tensor("op_1007_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_1007_cast_fp16 = slice_by_index(begin = var_1007_begin_0, end = var_1007_end_0, end_mask = var_1007_end_mask_0, x = var_877_cast_fp16)[name = tensor("op_1007_cast_fp16")]; tensor var_1014_begin_0 = const()[name = tensor("op_1014_begin_0"), val = tensor([0, 0, 0, 0])]; tensor var_1014_end_0 = const()[name = tensor("op_1014_end_0"), val = tensor([1, 64, 1, 375])]; tensor var_1014_end_mask_0 = const()[name = tensor("op_1014_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_1014_cast_fp16 = slice_by_index(begin = var_1014_begin_0, end = var_1014_end_0, end_mask = var_1014_end_mask_0, x = var_881_cast_fp16)[name = tensor("op_1014_cast_fp16")]; tensor var_1021_begin_0 = const()[name = tensor("op_1021_begin_0"), val = tensor([0, 0, 0, 375])]; tensor var_1021_end_0 = const()[name = tensor("op_1021_end_0"), val = tensor([1, 64, 1, 750])]; tensor var_1021_end_mask_0 = const()[name = tensor("op_1021_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_1021_cast_fp16 = slice_by_index(begin = var_1021_begin_0, end = var_1021_end_0, end_mask = var_1021_end_mask_0, x = var_881_cast_fp16)[name = tensor("op_1021_cast_fp16")]; tensor var_1028_begin_0 = const()[name = tensor("op_1028_begin_0"), val = tensor([0, 0, 0, 750])]; tensor var_1028_end_0 = const()[name = tensor("op_1028_end_0"), val = tensor([1, 64, 1, 1125])]; tensor var_1028_end_mask_0 = const()[name = tensor("op_1028_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_1028_cast_fp16 = slice_by_index(begin = var_1028_begin_0, end = var_1028_end_0, end_mask = var_1028_end_mask_0, x = var_881_cast_fp16)[name = tensor("op_1028_cast_fp16")]; tensor var_1035_begin_0 = const()[name = tensor("op_1035_begin_0"), val = tensor([0, 0, 0, 1125])]; tensor var_1035_end_0 = const()[name = tensor("op_1035_end_0"), val = tensor([1, 64, 1, 1500])]; tensor var_1035_end_mask_0 = const()[name = tensor("op_1035_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_1035_cast_fp16 = slice_by_index(begin = var_1035_begin_0, end = var_1035_end_0, end_mask = var_1035_end_mask_0, x = var_881_cast_fp16)[name = tensor("op_1035_cast_fp16")]; tensor var_1042_begin_0 = const()[name = tensor("op_1042_begin_0"), val = tensor([0, 0, 0, 0])]; tensor var_1042_end_0 = const()[name = tensor("op_1042_end_0"), val = tensor([1, 64, 1, 375])]; tensor var_1042_end_mask_0 = const()[name = tensor("op_1042_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_1042_cast_fp16 = slice_by_index(begin = var_1042_begin_0, end = var_1042_end_0, end_mask = var_1042_end_mask_0, x = var_885_cast_fp16)[name = tensor("op_1042_cast_fp16")]; tensor var_1049_begin_0 = const()[name = tensor("op_1049_begin_0"), val = tensor([0, 0, 0, 375])]; tensor var_1049_end_0 = const()[name = tensor("op_1049_end_0"), val = tensor([1, 64, 1, 750])]; tensor var_1049_end_mask_0 = const()[name = tensor("op_1049_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_1049_cast_fp16 = slice_by_index(begin = var_1049_begin_0, end = var_1049_end_0, end_mask = var_1049_end_mask_0, x = var_885_cast_fp16)[name = tensor("op_1049_cast_fp16")]; tensor var_1056_begin_0 = const()[name = tensor("op_1056_begin_0"), val = tensor([0, 0, 0, 750])]; tensor var_1056_end_0 = const()[name = tensor("op_1056_end_0"), val = tensor([1, 64, 1, 1125])]; tensor var_1056_end_mask_0 = const()[name = tensor("op_1056_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_1056_cast_fp16 = slice_by_index(begin = var_1056_begin_0, end = var_1056_end_0, end_mask = var_1056_end_mask_0, x = var_885_cast_fp16)[name = tensor("op_1056_cast_fp16")]; tensor var_1063_begin_0 = const()[name = tensor("op_1063_begin_0"), val = tensor([0, 0, 0, 1125])]; tensor var_1063_end_0 = const()[name = tensor("op_1063_end_0"), val = tensor([1, 64, 1, 1500])]; tensor var_1063_end_mask_0 = const()[name = tensor("op_1063_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_1063_cast_fp16 = slice_by_index(begin = var_1063_begin_0, end = var_1063_end_0, end_mask = var_1063_end_mask_0, x = var_885_cast_fp16)[name = tensor("op_1063_cast_fp16")]; tensor var_1070_begin_0 = const()[name = tensor("op_1070_begin_0"), val = tensor([0, 0, 0, 0])]; tensor var_1070_end_0 = const()[name = tensor("op_1070_end_0"), val = tensor([1, 64, 1, 375])]; tensor var_1070_end_mask_0 = const()[name = tensor("op_1070_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_1070_cast_fp16 = slice_by_index(begin = var_1070_begin_0, end = var_1070_end_0, end_mask = var_1070_end_mask_0, x = var_889_cast_fp16)[name = tensor("op_1070_cast_fp16")]; tensor var_1077_begin_0 = const()[name = tensor("op_1077_begin_0"), val = tensor([0, 0, 0, 375])]; tensor var_1077_end_0 = const()[name = tensor("op_1077_end_0"), val = tensor([1, 64, 1, 750])]; tensor var_1077_end_mask_0 = const()[name = tensor("op_1077_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_1077_cast_fp16 = slice_by_index(begin = var_1077_begin_0, end = var_1077_end_0, end_mask = var_1077_end_mask_0, x = var_889_cast_fp16)[name = tensor("op_1077_cast_fp16")]; tensor var_1084_begin_0 = const()[name = tensor("op_1084_begin_0"), val = tensor([0, 0, 0, 750])]; tensor var_1084_end_0 = const()[name = tensor("op_1084_end_0"), val = tensor([1, 64, 1, 1125])]; tensor var_1084_end_mask_0 = const()[name = tensor("op_1084_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_1084_cast_fp16 = slice_by_index(begin = var_1084_begin_0, end = var_1084_end_0, end_mask = var_1084_end_mask_0, x = var_889_cast_fp16)[name = tensor("op_1084_cast_fp16")]; tensor var_1091_begin_0 = const()[name = tensor("op_1091_begin_0"), val = tensor([0, 0, 0, 1125])]; tensor var_1091_end_0 = const()[name = tensor("op_1091_end_0"), val = tensor([1, 64, 1, 1500])]; tensor var_1091_end_mask_0 = const()[name = tensor("op_1091_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_1091_cast_fp16 = slice_by_index(begin = var_1091_begin_0, end = var_1091_end_0, end_mask = var_1091_end_mask_0, x = var_889_cast_fp16)[name = tensor("op_1091_cast_fp16")]; tensor var_1098_begin_0 = const()[name = tensor("op_1098_begin_0"), val = tensor([0, 0, 0, 0])]; tensor var_1098_end_0 = const()[name = tensor("op_1098_end_0"), val = tensor([1, 64, 1, 375])]; tensor var_1098_end_mask_0 = const()[name = tensor("op_1098_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_1098_cast_fp16 = slice_by_index(begin = var_1098_begin_0, end = var_1098_end_0, end_mask = var_1098_end_mask_0, x = var_893_cast_fp16)[name = tensor("op_1098_cast_fp16")]; tensor var_1105_begin_0 = const()[name = tensor("op_1105_begin_0"), val = tensor([0, 0, 0, 375])]; tensor var_1105_end_0 = const()[name = tensor("op_1105_end_0"), val = tensor([1, 64, 1, 750])]; tensor var_1105_end_mask_0 = const()[name = tensor("op_1105_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_1105_cast_fp16 = slice_by_index(begin = var_1105_begin_0, end = var_1105_end_0, end_mask = var_1105_end_mask_0, x = var_893_cast_fp16)[name = tensor("op_1105_cast_fp16")]; tensor var_1112_begin_0 = const()[name = tensor("op_1112_begin_0"), val = tensor([0, 0, 0, 750])]; tensor var_1112_end_0 = const()[name = tensor("op_1112_end_0"), val = tensor([1, 64, 1, 1125])]; tensor var_1112_end_mask_0 = const()[name = tensor("op_1112_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_1112_cast_fp16 = slice_by_index(begin = var_1112_begin_0, end = var_1112_end_0, end_mask = var_1112_end_mask_0, x = var_893_cast_fp16)[name = tensor("op_1112_cast_fp16")]; tensor var_1119_begin_0 = const()[name = tensor("op_1119_begin_0"), val = tensor([0, 0, 0, 1125])]; tensor var_1119_end_0 = const()[name = tensor("op_1119_end_0"), val = tensor([1, 64, 1, 1500])]; tensor var_1119_end_mask_0 = const()[name = tensor("op_1119_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_1119_cast_fp16 = slice_by_index(begin = var_1119_begin_0, end = var_1119_end_0, end_mask = var_1119_end_mask_0, x = var_893_cast_fp16)[name = tensor("op_1119_cast_fp16")]; tensor k_3_perm_0 = const()[name = tensor("k_3_perm_0"), val = tensor([0, 3, 2, 1])]; tensor var_1124_begin_0 = const()[name = tensor("op_1124_begin_0"), val = tensor([0, 0, 0, 0])]; tensor var_1124_end_0 = const()[name = tensor("op_1124_end_0"), val = tensor([1, 1500, 1, 64])]; tensor var_1124_end_mask_0 = const()[name = tensor("op_1124_end_mask_0"), val = tensor([true, true, true, false])]; tensor transpose_4 = transpose(perm = k_3_perm_0, x = key_3_cast_fp16)[name = tensor("transpose_4")]; tensor var_1124_cast_fp16 = slice_by_index(begin = var_1124_begin_0, end = var_1124_end_0, end_mask = var_1124_end_mask_0, x = transpose_4)[name = tensor("op_1124_cast_fp16")]; tensor var_1128_begin_0 = const()[name = tensor("op_1128_begin_0"), val = tensor([0, 0, 0, 64])]; tensor var_1128_end_0 = const()[name = tensor("op_1128_end_0"), val = tensor([1, 1500, 1, 128])]; tensor var_1128_end_mask_0 = const()[name = tensor("op_1128_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_1128_cast_fp16 = slice_by_index(begin = var_1128_begin_0, end = var_1128_end_0, end_mask = var_1128_end_mask_0, x = transpose_4)[name = tensor("op_1128_cast_fp16")]; tensor var_1132_begin_0 = const()[name = tensor("op_1132_begin_0"), val = tensor([0, 0, 0, 128])]; tensor var_1132_end_0 = const()[name = tensor("op_1132_end_0"), val = tensor([1, 1500, 1, 192])]; tensor var_1132_end_mask_0 = const()[name = tensor("op_1132_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_1132_cast_fp16 = slice_by_index(begin = var_1132_begin_0, end = var_1132_end_0, end_mask = var_1132_end_mask_0, x = transpose_4)[name = tensor("op_1132_cast_fp16")]; tensor var_1136_begin_0 = const()[name = tensor("op_1136_begin_0"), val = tensor([0, 0, 0, 192])]; tensor var_1136_end_0 = const()[name = tensor("op_1136_end_0"), val = tensor([1, 1500, 1, 256])]; tensor var_1136_end_mask_0 = const()[name = tensor("op_1136_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_1136_cast_fp16 = slice_by_index(begin = var_1136_begin_0, end = var_1136_end_0, end_mask = var_1136_end_mask_0, x = transpose_4)[name = tensor("op_1136_cast_fp16")]; tensor var_1140_begin_0 = const()[name = tensor("op_1140_begin_0"), val = tensor([0, 0, 0, 256])]; tensor var_1140_end_0 = const()[name = tensor("op_1140_end_0"), val = tensor([1, 1500, 1, 320])]; tensor var_1140_end_mask_0 = const()[name = tensor("op_1140_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_1140_cast_fp16 = slice_by_index(begin = var_1140_begin_0, end = var_1140_end_0, end_mask = var_1140_end_mask_0, x = transpose_4)[name = tensor("op_1140_cast_fp16")]; tensor var_1144_begin_0 = const()[name = tensor("op_1144_begin_0"), val = tensor([0, 0, 0, 320])]; tensor var_1144_end_0 = const()[name = tensor("op_1144_end_0"), val = tensor([1, 1500, 1, 384])]; tensor var_1144_end_mask_0 = const()[name = tensor("op_1144_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_1144_cast_fp16 = slice_by_index(begin = var_1144_begin_0, end = var_1144_end_0, end_mask = var_1144_end_mask_0, x = transpose_4)[name = tensor("op_1144_cast_fp16")]; tensor var_1148_begin_0 = const()[name = tensor("op_1148_begin_0"), val = tensor([0, 0, 0, 384])]; tensor var_1148_end_0 = const()[name = tensor("op_1148_end_0"), val = tensor([1, 1500, 1, 448])]; tensor var_1148_end_mask_0 = const()[name = tensor("op_1148_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_1148_cast_fp16 = slice_by_index(begin = var_1148_begin_0, end = var_1148_end_0, end_mask = var_1148_end_mask_0, x = transpose_4)[name = tensor("op_1148_cast_fp16")]; tensor var_1152_begin_0 = const()[name = tensor("op_1152_begin_0"), val = tensor([0, 0, 0, 448])]; tensor var_1152_end_0 = const()[name = tensor("op_1152_end_0"), val = tensor([1, 1500, 1, 512])]; tensor var_1152_end_mask_0 = const()[name = tensor("op_1152_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_1152_cast_fp16 = slice_by_index(begin = var_1152_begin_0, end = var_1152_end_0, end_mask = var_1152_end_mask_0, x = transpose_4)[name = tensor("op_1152_cast_fp16")]; tensor var_1154_begin_0 = const()[name = tensor("op_1154_begin_0"), val = tensor([0, 0, 0, 0])]; tensor var_1154_end_0 = const()[name = tensor("op_1154_end_0"), val = tensor([1, 64, 1, 1500])]; tensor var_1154_end_mask_0 = const()[name = tensor("op_1154_end_mask_0"), val = tensor([true, false, true, true])]; tensor var_1154_cast_fp16 = slice_by_index(begin = var_1154_begin_0, end = var_1154_end_0, end_mask = var_1154_end_mask_0, x = value_3_cast_fp16)[name = tensor("op_1154_cast_fp16")]; tensor var_1158_begin_0 = const()[name = tensor("op_1158_begin_0"), val = tensor([0, 64, 0, 0])]; tensor var_1158_end_0 = const()[name = tensor("op_1158_end_0"), val = tensor([1, 128, 1, 1500])]; tensor var_1158_end_mask_0 = const()[name = tensor("op_1158_end_mask_0"), val = tensor([true, false, true, true])]; tensor var_1158_cast_fp16 = slice_by_index(begin = var_1158_begin_0, end = var_1158_end_0, end_mask = var_1158_end_mask_0, x = value_3_cast_fp16)[name = tensor("op_1158_cast_fp16")]; tensor var_1162_begin_0 = const()[name = tensor("op_1162_begin_0"), val = tensor([0, 128, 0, 0])]; tensor var_1162_end_0 = const()[name = tensor("op_1162_end_0"), val = tensor([1, 192, 1, 1500])]; tensor var_1162_end_mask_0 = const()[name = tensor("op_1162_end_mask_0"), val = tensor([true, false, true, true])]; tensor var_1162_cast_fp16 = slice_by_index(begin = var_1162_begin_0, end = var_1162_end_0, end_mask = var_1162_end_mask_0, x = value_3_cast_fp16)[name = tensor("op_1162_cast_fp16")]; tensor var_1166_begin_0 = const()[name = tensor("op_1166_begin_0"), val = tensor([0, 192, 0, 0])]; tensor var_1166_end_0 = const()[name = tensor("op_1166_end_0"), val = tensor([1, 256, 1, 1500])]; tensor var_1166_end_mask_0 = const()[name = tensor("op_1166_end_mask_0"), val = tensor([true, false, true, true])]; tensor var_1166_cast_fp16 = slice_by_index(begin = var_1166_begin_0, end = var_1166_end_0, end_mask = var_1166_end_mask_0, x = value_3_cast_fp16)[name = tensor("op_1166_cast_fp16")]; tensor var_1170_begin_0 = const()[name = tensor("op_1170_begin_0"), val = tensor([0, 256, 0, 0])]; tensor var_1170_end_0 = const()[name = tensor("op_1170_end_0"), val = tensor([1, 320, 1, 1500])]; tensor var_1170_end_mask_0 = const()[name = tensor("op_1170_end_mask_0"), val = tensor([true, false, true, true])]; tensor var_1170_cast_fp16 = slice_by_index(begin = var_1170_begin_0, end = var_1170_end_0, end_mask = var_1170_end_mask_0, x = value_3_cast_fp16)[name = tensor("op_1170_cast_fp16")]; tensor var_1174_begin_0 = const()[name = tensor("op_1174_begin_0"), val = tensor([0, 320, 0, 0])]; tensor var_1174_end_0 = const()[name = tensor("op_1174_end_0"), val = tensor([1, 384, 1, 1500])]; tensor var_1174_end_mask_0 = const()[name = tensor("op_1174_end_mask_0"), val = tensor([true, false, true, true])]; tensor var_1174_cast_fp16 = slice_by_index(begin = var_1174_begin_0, end = var_1174_end_0, end_mask = var_1174_end_mask_0, x = value_3_cast_fp16)[name = tensor("op_1174_cast_fp16")]; tensor var_1178_begin_0 = const()[name = tensor("op_1178_begin_0"), val = tensor([0, 384, 0, 0])]; tensor var_1178_end_0 = const()[name = tensor("op_1178_end_0"), val = tensor([1, 448, 1, 1500])]; tensor var_1178_end_mask_0 = const()[name = tensor("op_1178_end_mask_0"), val = tensor([true, false, true, true])]; tensor var_1178_cast_fp16 = slice_by_index(begin = var_1178_begin_0, end = var_1178_end_0, end_mask = var_1178_end_mask_0, x = value_3_cast_fp16)[name = tensor("op_1178_cast_fp16")]; tensor var_1182_begin_0 = const()[name = tensor("op_1182_begin_0"), val = tensor([0, 448, 0, 0])]; tensor var_1182_end_0 = const()[name = tensor("op_1182_end_0"), val = tensor([1, 512, 1, 1500])]; tensor var_1182_end_mask_0 = const()[name = tensor("op_1182_end_mask_0"), val = tensor([true, false, true, true])]; tensor var_1182_cast_fp16 = slice_by_index(begin = var_1182_begin_0, end = var_1182_end_0, end_mask = var_1182_end_mask_0, x = value_3_cast_fp16)[name = tensor("op_1182_cast_fp16")]; tensor var_1186_equation_0 = const()[name = tensor("op_1186_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_1186_cast_fp16 = einsum(equation = var_1186_equation_0, values = (var_1124_cast_fp16, var_902_cast_fp16))[name = tensor("op_1186_cast_fp16")]; tensor var_1187_to_fp16 = const()[name = tensor("op_1187_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_65_cast_fp16 = mul(x = var_1186_cast_fp16, y = var_1187_to_fp16)[name = tensor("aw_chunk_65_cast_fp16")]; tensor var_1190_equation_0 = const()[name = tensor("op_1190_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_1190_cast_fp16 = einsum(equation = var_1190_equation_0, values = (var_1124_cast_fp16, var_909_cast_fp16))[name = tensor("op_1190_cast_fp16")]; tensor var_1191_to_fp16 = const()[name = tensor("op_1191_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_67_cast_fp16 = mul(x = var_1190_cast_fp16, y = var_1191_to_fp16)[name = tensor("aw_chunk_67_cast_fp16")]; tensor var_1194_equation_0 = const()[name = tensor("op_1194_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_1194_cast_fp16 = einsum(equation = var_1194_equation_0, values = (var_1124_cast_fp16, var_916_cast_fp16))[name = tensor("op_1194_cast_fp16")]; tensor var_1195_to_fp16 = const()[name = tensor("op_1195_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_69_cast_fp16 = mul(x = var_1194_cast_fp16, y = var_1195_to_fp16)[name = tensor("aw_chunk_69_cast_fp16")]; tensor var_1198_equation_0 = const()[name = tensor("op_1198_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_1198_cast_fp16 = einsum(equation = var_1198_equation_0, values = (var_1124_cast_fp16, var_923_cast_fp16))[name = tensor("op_1198_cast_fp16")]; tensor var_1199_to_fp16 = const()[name = tensor("op_1199_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_71_cast_fp16 = mul(x = var_1198_cast_fp16, y = var_1199_to_fp16)[name = tensor("aw_chunk_71_cast_fp16")]; tensor var_1202_equation_0 = const()[name = tensor("op_1202_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_1202_cast_fp16 = einsum(equation = var_1202_equation_0, values = (var_1128_cast_fp16, var_930_cast_fp16))[name = tensor("op_1202_cast_fp16")]; tensor var_1203_to_fp16 = const()[name = tensor("op_1203_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_73_cast_fp16 = mul(x = var_1202_cast_fp16, y = var_1203_to_fp16)[name = tensor("aw_chunk_73_cast_fp16")]; tensor var_1206_equation_0 = const()[name = tensor("op_1206_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_1206_cast_fp16 = einsum(equation = var_1206_equation_0, values = (var_1128_cast_fp16, var_937_cast_fp16))[name = tensor("op_1206_cast_fp16")]; tensor var_1207_to_fp16 = const()[name = tensor("op_1207_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_75_cast_fp16 = mul(x = var_1206_cast_fp16, y = var_1207_to_fp16)[name = tensor("aw_chunk_75_cast_fp16")]; tensor var_1210_equation_0 = const()[name = tensor("op_1210_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_1210_cast_fp16 = einsum(equation = var_1210_equation_0, values = (var_1128_cast_fp16, var_944_cast_fp16))[name = tensor("op_1210_cast_fp16")]; tensor var_1211_to_fp16 = const()[name = tensor("op_1211_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_77_cast_fp16 = mul(x = var_1210_cast_fp16, y = var_1211_to_fp16)[name = tensor("aw_chunk_77_cast_fp16")]; tensor var_1214_equation_0 = const()[name = tensor("op_1214_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_1214_cast_fp16 = einsum(equation = var_1214_equation_0, values = (var_1128_cast_fp16, var_951_cast_fp16))[name = tensor("op_1214_cast_fp16")]; tensor var_1215_to_fp16 = const()[name = tensor("op_1215_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_79_cast_fp16 = mul(x = var_1214_cast_fp16, y = var_1215_to_fp16)[name = tensor("aw_chunk_79_cast_fp16")]; tensor var_1218_equation_0 = const()[name = tensor("op_1218_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_1218_cast_fp16 = einsum(equation = var_1218_equation_0, values = (var_1132_cast_fp16, var_958_cast_fp16))[name = tensor("op_1218_cast_fp16")]; tensor var_1219_to_fp16 = const()[name = tensor("op_1219_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_81_cast_fp16 = mul(x = var_1218_cast_fp16, y = var_1219_to_fp16)[name = tensor("aw_chunk_81_cast_fp16")]; tensor var_1222_equation_0 = const()[name = tensor("op_1222_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_1222_cast_fp16 = einsum(equation = var_1222_equation_0, values = (var_1132_cast_fp16, var_965_cast_fp16))[name = tensor("op_1222_cast_fp16")]; tensor var_1223_to_fp16 = const()[name = tensor("op_1223_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_83_cast_fp16 = mul(x = var_1222_cast_fp16, y = var_1223_to_fp16)[name = tensor("aw_chunk_83_cast_fp16")]; tensor var_1226_equation_0 = const()[name = tensor("op_1226_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_1226_cast_fp16 = einsum(equation = var_1226_equation_0, values = (var_1132_cast_fp16, var_972_cast_fp16))[name = tensor("op_1226_cast_fp16")]; tensor var_1227_to_fp16 = const()[name = tensor("op_1227_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_85_cast_fp16 = mul(x = var_1226_cast_fp16, y = var_1227_to_fp16)[name = tensor("aw_chunk_85_cast_fp16")]; tensor var_1230_equation_0 = const()[name = tensor("op_1230_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_1230_cast_fp16 = einsum(equation = var_1230_equation_0, values = (var_1132_cast_fp16, var_979_cast_fp16))[name = tensor("op_1230_cast_fp16")]; tensor var_1231_to_fp16 = const()[name = tensor("op_1231_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_87_cast_fp16 = mul(x = var_1230_cast_fp16, y = var_1231_to_fp16)[name = tensor("aw_chunk_87_cast_fp16")]; tensor var_1234_equation_0 = const()[name = tensor("op_1234_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_1234_cast_fp16 = einsum(equation = var_1234_equation_0, values = (var_1136_cast_fp16, var_986_cast_fp16))[name = tensor("op_1234_cast_fp16")]; tensor var_1235_to_fp16 = const()[name = tensor("op_1235_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_89_cast_fp16 = mul(x = var_1234_cast_fp16, y = var_1235_to_fp16)[name = tensor("aw_chunk_89_cast_fp16")]; tensor var_1238_equation_0 = const()[name = tensor("op_1238_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_1238_cast_fp16 = einsum(equation = var_1238_equation_0, values = (var_1136_cast_fp16, var_993_cast_fp16))[name = tensor("op_1238_cast_fp16")]; tensor var_1239_to_fp16 = const()[name = tensor("op_1239_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_91_cast_fp16 = mul(x = var_1238_cast_fp16, y = var_1239_to_fp16)[name = tensor("aw_chunk_91_cast_fp16")]; tensor var_1242_equation_0 = const()[name = tensor("op_1242_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_1242_cast_fp16 = einsum(equation = var_1242_equation_0, values = (var_1136_cast_fp16, var_1000_cast_fp16))[name = tensor("op_1242_cast_fp16")]; tensor var_1243_to_fp16 = const()[name = tensor("op_1243_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_93_cast_fp16 = mul(x = var_1242_cast_fp16, y = var_1243_to_fp16)[name = tensor("aw_chunk_93_cast_fp16")]; tensor var_1246_equation_0 = const()[name = tensor("op_1246_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_1246_cast_fp16 = einsum(equation = var_1246_equation_0, values = (var_1136_cast_fp16, var_1007_cast_fp16))[name = tensor("op_1246_cast_fp16")]; tensor var_1247_to_fp16 = const()[name = tensor("op_1247_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_95_cast_fp16 = mul(x = var_1246_cast_fp16, y = var_1247_to_fp16)[name = tensor("aw_chunk_95_cast_fp16")]; tensor var_1250_equation_0 = const()[name = tensor("op_1250_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_1250_cast_fp16 = einsum(equation = var_1250_equation_0, values = (var_1140_cast_fp16, var_1014_cast_fp16))[name = tensor("op_1250_cast_fp16")]; tensor var_1251_to_fp16 = const()[name = tensor("op_1251_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_97_cast_fp16 = mul(x = var_1250_cast_fp16, y = var_1251_to_fp16)[name = tensor("aw_chunk_97_cast_fp16")]; tensor var_1254_equation_0 = const()[name = tensor("op_1254_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_1254_cast_fp16 = einsum(equation = var_1254_equation_0, values = (var_1140_cast_fp16, var_1021_cast_fp16))[name = tensor("op_1254_cast_fp16")]; tensor var_1255_to_fp16 = const()[name = tensor("op_1255_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_99_cast_fp16 = mul(x = var_1254_cast_fp16, y = var_1255_to_fp16)[name = tensor("aw_chunk_99_cast_fp16")]; tensor var_1258_equation_0 = const()[name = tensor("op_1258_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_1258_cast_fp16 = einsum(equation = var_1258_equation_0, values = (var_1140_cast_fp16, var_1028_cast_fp16))[name = tensor("op_1258_cast_fp16")]; tensor var_1259_to_fp16 = const()[name = tensor("op_1259_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_101_cast_fp16 = mul(x = var_1258_cast_fp16, y = var_1259_to_fp16)[name = tensor("aw_chunk_101_cast_fp16")]; tensor var_1262_equation_0 = const()[name = tensor("op_1262_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_1262_cast_fp16 = einsum(equation = var_1262_equation_0, values = (var_1140_cast_fp16, var_1035_cast_fp16))[name = tensor("op_1262_cast_fp16")]; tensor var_1263_to_fp16 = const()[name = tensor("op_1263_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_103_cast_fp16 = mul(x = var_1262_cast_fp16, y = var_1263_to_fp16)[name = tensor("aw_chunk_103_cast_fp16")]; tensor var_1266_equation_0 = const()[name = tensor("op_1266_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_1266_cast_fp16 = einsum(equation = var_1266_equation_0, values = (var_1144_cast_fp16, var_1042_cast_fp16))[name = tensor("op_1266_cast_fp16")]; tensor var_1267_to_fp16 = const()[name = tensor("op_1267_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_105_cast_fp16 = mul(x = var_1266_cast_fp16, y = var_1267_to_fp16)[name = tensor("aw_chunk_105_cast_fp16")]; tensor var_1270_equation_0 = const()[name = tensor("op_1270_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_1270_cast_fp16 = einsum(equation = var_1270_equation_0, values = (var_1144_cast_fp16, var_1049_cast_fp16))[name = tensor("op_1270_cast_fp16")]; tensor var_1271_to_fp16 = const()[name = tensor("op_1271_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_107_cast_fp16 = mul(x = var_1270_cast_fp16, y = var_1271_to_fp16)[name = tensor("aw_chunk_107_cast_fp16")]; tensor var_1274_equation_0 = const()[name = tensor("op_1274_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_1274_cast_fp16 = einsum(equation = var_1274_equation_0, values = (var_1144_cast_fp16, var_1056_cast_fp16))[name = tensor("op_1274_cast_fp16")]; tensor var_1275_to_fp16 = const()[name = tensor("op_1275_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_109_cast_fp16 = mul(x = var_1274_cast_fp16, y = var_1275_to_fp16)[name = tensor("aw_chunk_109_cast_fp16")]; tensor var_1278_equation_0 = const()[name = tensor("op_1278_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_1278_cast_fp16 = einsum(equation = var_1278_equation_0, values = (var_1144_cast_fp16, var_1063_cast_fp16))[name = tensor("op_1278_cast_fp16")]; tensor var_1279_to_fp16 = const()[name = tensor("op_1279_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_111_cast_fp16 = mul(x = var_1278_cast_fp16, y = var_1279_to_fp16)[name = tensor("aw_chunk_111_cast_fp16")]; tensor var_1282_equation_0 = const()[name = tensor("op_1282_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_1282_cast_fp16 = einsum(equation = var_1282_equation_0, values = (var_1148_cast_fp16, var_1070_cast_fp16))[name = tensor("op_1282_cast_fp16")]; tensor var_1283_to_fp16 = const()[name = tensor("op_1283_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_113_cast_fp16 = mul(x = var_1282_cast_fp16, y = var_1283_to_fp16)[name = tensor("aw_chunk_113_cast_fp16")]; tensor var_1286_equation_0 = const()[name = tensor("op_1286_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_1286_cast_fp16 = einsum(equation = var_1286_equation_0, values = (var_1148_cast_fp16, var_1077_cast_fp16))[name = tensor("op_1286_cast_fp16")]; tensor var_1287_to_fp16 = const()[name = tensor("op_1287_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_115_cast_fp16 = mul(x = var_1286_cast_fp16, y = var_1287_to_fp16)[name = tensor("aw_chunk_115_cast_fp16")]; tensor var_1290_equation_0 = const()[name = tensor("op_1290_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_1290_cast_fp16 = einsum(equation = var_1290_equation_0, values = (var_1148_cast_fp16, var_1084_cast_fp16))[name = tensor("op_1290_cast_fp16")]; tensor var_1291_to_fp16 = const()[name = tensor("op_1291_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_117_cast_fp16 = mul(x = var_1290_cast_fp16, y = var_1291_to_fp16)[name = tensor("aw_chunk_117_cast_fp16")]; tensor var_1294_equation_0 = const()[name = tensor("op_1294_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_1294_cast_fp16 = einsum(equation = var_1294_equation_0, values = (var_1148_cast_fp16, var_1091_cast_fp16))[name = tensor("op_1294_cast_fp16")]; tensor var_1295_to_fp16 = const()[name = tensor("op_1295_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_119_cast_fp16 = mul(x = var_1294_cast_fp16, y = var_1295_to_fp16)[name = tensor("aw_chunk_119_cast_fp16")]; tensor var_1298_equation_0 = const()[name = tensor("op_1298_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_1298_cast_fp16 = einsum(equation = var_1298_equation_0, values = (var_1152_cast_fp16, var_1098_cast_fp16))[name = tensor("op_1298_cast_fp16")]; tensor var_1299_to_fp16 = const()[name = tensor("op_1299_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_121_cast_fp16 = mul(x = var_1298_cast_fp16, y = var_1299_to_fp16)[name = tensor("aw_chunk_121_cast_fp16")]; tensor var_1302_equation_0 = const()[name = tensor("op_1302_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_1302_cast_fp16 = einsum(equation = var_1302_equation_0, values = (var_1152_cast_fp16, var_1105_cast_fp16))[name = tensor("op_1302_cast_fp16")]; tensor var_1303_to_fp16 = const()[name = tensor("op_1303_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_123_cast_fp16 = mul(x = var_1302_cast_fp16, y = var_1303_to_fp16)[name = tensor("aw_chunk_123_cast_fp16")]; tensor var_1306_equation_0 = const()[name = tensor("op_1306_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_1306_cast_fp16 = einsum(equation = var_1306_equation_0, values = (var_1152_cast_fp16, var_1112_cast_fp16))[name = tensor("op_1306_cast_fp16")]; tensor var_1307_to_fp16 = const()[name = tensor("op_1307_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_125_cast_fp16 = mul(x = var_1306_cast_fp16, y = var_1307_to_fp16)[name = tensor("aw_chunk_125_cast_fp16")]; tensor var_1310_equation_0 = const()[name = tensor("op_1310_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_1310_cast_fp16 = einsum(equation = var_1310_equation_0, values = (var_1152_cast_fp16, var_1119_cast_fp16))[name = tensor("op_1310_cast_fp16")]; tensor var_1311_to_fp16 = const()[name = tensor("op_1311_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_127_cast_fp16 = mul(x = var_1310_cast_fp16, y = var_1311_to_fp16)[name = tensor("aw_chunk_127_cast_fp16")]; tensor var_1313_cast_fp16 = softmax(axis = var_810, x = aw_chunk_65_cast_fp16)[name = tensor("op_1313_cast_fp16")]; tensor var_1314_cast_fp16 = softmax(axis = var_810, x = aw_chunk_67_cast_fp16)[name = tensor("op_1314_cast_fp16")]; tensor var_1315_cast_fp16 = softmax(axis = var_810, x = aw_chunk_69_cast_fp16)[name = tensor("op_1315_cast_fp16")]; tensor var_1316_cast_fp16 = softmax(axis = var_810, x = aw_chunk_71_cast_fp16)[name = tensor("op_1316_cast_fp16")]; tensor var_1317_cast_fp16 = softmax(axis = var_810, x = aw_chunk_73_cast_fp16)[name = tensor("op_1317_cast_fp16")]; tensor var_1318_cast_fp16 = softmax(axis = var_810, x = aw_chunk_75_cast_fp16)[name = tensor("op_1318_cast_fp16")]; tensor var_1319_cast_fp16 = softmax(axis = var_810, x = aw_chunk_77_cast_fp16)[name = tensor("op_1319_cast_fp16")]; tensor var_1320_cast_fp16 = softmax(axis = var_810, x = aw_chunk_79_cast_fp16)[name = tensor("op_1320_cast_fp16")]; tensor var_1321_cast_fp16 = softmax(axis = var_810, x = aw_chunk_81_cast_fp16)[name = tensor("op_1321_cast_fp16")]; tensor var_1322_cast_fp16 = softmax(axis = var_810, x = aw_chunk_83_cast_fp16)[name = tensor("op_1322_cast_fp16")]; tensor var_1323_cast_fp16 = softmax(axis = var_810, x = aw_chunk_85_cast_fp16)[name = tensor("op_1323_cast_fp16")]; tensor var_1324_cast_fp16 = softmax(axis = var_810, x = aw_chunk_87_cast_fp16)[name = tensor("op_1324_cast_fp16")]; tensor var_1325_cast_fp16 = softmax(axis = var_810, x = aw_chunk_89_cast_fp16)[name = tensor("op_1325_cast_fp16")]; tensor var_1326_cast_fp16 = softmax(axis = var_810, x = aw_chunk_91_cast_fp16)[name = tensor("op_1326_cast_fp16")]; tensor var_1327_cast_fp16 = softmax(axis = var_810, x = aw_chunk_93_cast_fp16)[name = tensor("op_1327_cast_fp16")]; tensor var_1328_cast_fp16 = softmax(axis = var_810, x = aw_chunk_95_cast_fp16)[name = tensor("op_1328_cast_fp16")]; tensor var_1329_cast_fp16 = softmax(axis = var_810, x = aw_chunk_97_cast_fp16)[name = tensor("op_1329_cast_fp16")]; tensor var_1330_cast_fp16 = softmax(axis = var_810, x = aw_chunk_99_cast_fp16)[name = tensor("op_1330_cast_fp16")]; tensor var_1331_cast_fp16 = softmax(axis = var_810, x = aw_chunk_101_cast_fp16)[name = tensor("op_1331_cast_fp16")]; tensor var_1332_cast_fp16 = softmax(axis = var_810, x = aw_chunk_103_cast_fp16)[name = tensor("op_1332_cast_fp16")]; tensor var_1333_cast_fp16 = softmax(axis = var_810, x = aw_chunk_105_cast_fp16)[name = tensor("op_1333_cast_fp16")]; tensor var_1334_cast_fp16 = softmax(axis = var_810, x = aw_chunk_107_cast_fp16)[name = tensor("op_1334_cast_fp16")]; tensor var_1335_cast_fp16 = softmax(axis = var_810, x = aw_chunk_109_cast_fp16)[name = tensor("op_1335_cast_fp16")]; tensor var_1336_cast_fp16 = softmax(axis = var_810, x = aw_chunk_111_cast_fp16)[name = tensor("op_1336_cast_fp16")]; tensor var_1337_cast_fp16 = softmax(axis = var_810, x = aw_chunk_113_cast_fp16)[name = tensor("op_1337_cast_fp16")]; tensor var_1338_cast_fp16 = softmax(axis = var_810, x = aw_chunk_115_cast_fp16)[name = tensor("op_1338_cast_fp16")]; tensor var_1339_cast_fp16 = softmax(axis = var_810, x = aw_chunk_117_cast_fp16)[name = tensor("op_1339_cast_fp16")]; tensor var_1340_cast_fp16 = softmax(axis = var_810, x = aw_chunk_119_cast_fp16)[name = tensor("op_1340_cast_fp16")]; tensor var_1341_cast_fp16 = softmax(axis = var_810, x = aw_chunk_121_cast_fp16)[name = tensor("op_1341_cast_fp16")]; tensor var_1342_cast_fp16 = softmax(axis = var_810, x = aw_chunk_123_cast_fp16)[name = tensor("op_1342_cast_fp16")]; tensor var_1343_cast_fp16 = softmax(axis = var_810, x = aw_chunk_125_cast_fp16)[name = tensor("op_1343_cast_fp16")]; tensor var_1344_cast_fp16 = softmax(axis = var_810, x = aw_chunk_127_cast_fp16)[name = tensor("op_1344_cast_fp16")]; tensor var_1346_equation_0 = const()[name = tensor("op_1346_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_1346_cast_fp16 = einsum(equation = var_1346_equation_0, values = (var_1154_cast_fp16, var_1313_cast_fp16))[name = tensor("op_1346_cast_fp16")]; tensor var_1348_equation_0 = const()[name = tensor("op_1348_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_1348_cast_fp16 = einsum(equation = var_1348_equation_0, values = (var_1154_cast_fp16, var_1314_cast_fp16))[name = tensor("op_1348_cast_fp16")]; tensor var_1350_equation_0 = const()[name = tensor("op_1350_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_1350_cast_fp16 = einsum(equation = var_1350_equation_0, values = (var_1154_cast_fp16, var_1315_cast_fp16))[name = tensor("op_1350_cast_fp16")]; tensor var_1352_equation_0 = const()[name = tensor("op_1352_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_1352_cast_fp16 = einsum(equation = var_1352_equation_0, values = (var_1154_cast_fp16, var_1316_cast_fp16))[name = tensor("op_1352_cast_fp16")]; tensor var_1354_equation_0 = const()[name = tensor("op_1354_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_1354_cast_fp16 = einsum(equation = var_1354_equation_0, values = (var_1158_cast_fp16, var_1317_cast_fp16))[name = tensor("op_1354_cast_fp16")]; tensor var_1356_equation_0 = const()[name = tensor("op_1356_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_1356_cast_fp16 = einsum(equation = var_1356_equation_0, values = (var_1158_cast_fp16, var_1318_cast_fp16))[name = tensor("op_1356_cast_fp16")]; tensor var_1358_equation_0 = const()[name = tensor("op_1358_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_1358_cast_fp16 = einsum(equation = var_1358_equation_0, values = (var_1158_cast_fp16, var_1319_cast_fp16))[name = tensor("op_1358_cast_fp16")]; tensor var_1360_equation_0 = const()[name = tensor("op_1360_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_1360_cast_fp16 = einsum(equation = var_1360_equation_0, values = (var_1158_cast_fp16, var_1320_cast_fp16))[name = tensor("op_1360_cast_fp16")]; tensor var_1362_equation_0 = const()[name = tensor("op_1362_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_1362_cast_fp16 = einsum(equation = var_1362_equation_0, values = (var_1162_cast_fp16, var_1321_cast_fp16))[name = tensor("op_1362_cast_fp16")]; tensor var_1364_equation_0 = const()[name = tensor("op_1364_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_1364_cast_fp16 = einsum(equation = var_1364_equation_0, values = (var_1162_cast_fp16, var_1322_cast_fp16))[name = tensor("op_1364_cast_fp16")]; tensor var_1366_equation_0 = const()[name = tensor("op_1366_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_1366_cast_fp16 = einsum(equation = var_1366_equation_0, values = (var_1162_cast_fp16, var_1323_cast_fp16))[name = tensor("op_1366_cast_fp16")]; tensor var_1368_equation_0 = const()[name = tensor("op_1368_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_1368_cast_fp16 = einsum(equation = var_1368_equation_0, values = (var_1162_cast_fp16, var_1324_cast_fp16))[name = tensor("op_1368_cast_fp16")]; tensor var_1370_equation_0 = const()[name = tensor("op_1370_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_1370_cast_fp16 = einsum(equation = var_1370_equation_0, values = (var_1166_cast_fp16, var_1325_cast_fp16))[name = tensor("op_1370_cast_fp16")]; tensor var_1372_equation_0 = const()[name = tensor("op_1372_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_1372_cast_fp16 = einsum(equation = var_1372_equation_0, values = (var_1166_cast_fp16, var_1326_cast_fp16))[name = tensor("op_1372_cast_fp16")]; tensor var_1374_equation_0 = const()[name = tensor("op_1374_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_1374_cast_fp16 = einsum(equation = var_1374_equation_0, values = (var_1166_cast_fp16, var_1327_cast_fp16))[name = tensor("op_1374_cast_fp16")]; tensor var_1376_equation_0 = const()[name = tensor("op_1376_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_1376_cast_fp16 = einsum(equation = var_1376_equation_0, values = (var_1166_cast_fp16, var_1328_cast_fp16))[name = tensor("op_1376_cast_fp16")]; tensor var_1378_equation_0 = const()[name = tensor("op_1378_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_1378_cast_fp16 = einsum(equation = var_1378_equation_0, values = (var_1170_cast_fp16, var_1329_cast_fp16))[name = tensor("op_1378_cast_fp16")]; tensor var_1380_equation_0 = const()[name = tensor("op_1380_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_1380_cast_fp16 = einsum(equation = var_1380_equation_0, values = (var_1170_cast_fp16, var_1330_cast_fp16))[name = tensor("op_1380_cast_fp16")]; tensor var_1382_equation_0 = const()[name = tensor("op_1382_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_1382_cast_fp16 = einsum(equation = var_1382_equation_0, values = (var_1170_cast_fp16, var_1331_cast_fp16))[name = tensor("op_1382_cast_fp16")]; tensor var_1384_equation_0 = const()[name = tensor("op_1384_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_1384_cast_fp16 = einsum(equation = var_1384_equation_0, values = (var_1170_cast_fp16, var_1332_cast_fp16))[name = tensor("op_1384_cast_fp16")]; tensor var_1386_equation_0 = const()[name = tensor("op_1386_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_1386_cast_fp16 = einsum(equation = var_1386_equation_0, values = (var_1174_cast_fp16, var_1333_cast_fp16))[name = tensor("op_1386_cast_fp16")]; tensor var_1388_equation_0 = const()[name = tensor("op_1388_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_1388_cast_fp16 = einsum(equation = var_1388_equation_0, values = (var_1174_cast_fp16, var_1334_cast_fp16))[name = tensor("op_1388_cast_fp16")]; tensor var_1390_equation_0 = const()[name = tensor("op_1390_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_1390_cast_fp16 = einsum(equation = var_1390_equation_0, values = (var_1174_cast_fp16, var_1335_cast_fp16))[name = tensor("op_1390_cast_fp16")]; tensor var_1392_equation_0 = const()[name = tensor("op_1392_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_1392_cast_fp16 = einsum(equation = var_1392_equation_0, values = (var_1174_cast_fp16, var_1336_cast_fp16))[name = tensor("op_1392_cast_fp16")]; tensor var_1394_equation_0 = const()[name = tensor("op_1394_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_1394_cast_fp16 = einsum(equation = var_1394_equation_0, values = (var_1178_cast_fp16, var_1337_cast_fp16))[name = tensor("op_1394_cast_fp16")]; tensor var_1396_equation_0 = const()[name = tensor("op_1396_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_1396_cast_fp16 = einsum(equation = var_1396_equation_0, values = (var_1178_cast_fp16, var_1338_cast_fp16))[name = tensor("op_1396_cast_fp16")]; tensor var_1398_equation_0 = const()[name = tensor("op_1398_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_1398_cast_fp16 = einsum(equation = var_1398_equation_0, values = (var_1178_cast_fp16, var_1339_cast_fp16))[name = tensor("op_1398_cast_fp16")]; tensor var_1400_equation_0 = const()[name = tensor("op_1400_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_1400_cast_fp16 = einsum(equation = var_1400_equation_0, values = (var_1178_cast_fp16, var_1340_cast_fp16))[name = tensor("op_1400_cast_fp16")]; tensor var_1402_equation_0 = const()[name = tensor("op_1402_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_1402_cast_fp16 = einsum(equation = var_1402_equation_0, values = (var_1182_cast_fp16, var_1341_cast_fp16))[name = tensor("op_1402_cast_fp16")]; tensor var_1404_equation_0 = const()[name = tensor("op_1404_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_1404_cast_fp16 = einsum(equation = var_1404_equation_0, values = (var_1182_cast_fp16, var_1342_cast_fp16))[name = tensor("op_1404_cast_fp16")]; tensor var_1406_equation_0 = const()[name = tensor("op_1406_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_1406_cast_fp16 = einsum(equation = var_1406_equation_0, values = (var_1182_cast_fp16, var_1343_cast_fp16))[name = tensor("op_1406_cast_fp16")]; tensor var_1408_equation_0 = const()[name = tensor("op_1408_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_1408_cast_fp16 = einsum(equation = var_1408_equation_0, values = (var_1182_cast_fp16, var_1344_cast_fp16))[name = tensor("op_1408_cast_fp16")]; tensor var_1410_interleave_0 = const()[name = tensor("op_1410_interleave_0"), val = tensor(false)]; tensor var_1410_cast_fp16 = concat(axis = var_797, interleave = var_1410_interleave_0, values = (var_1346_cast_fp16, var_1348_cast_fp16, var_1350_cast_fp16, var_1352_cast_fp16))[name = tensor("op_1410_cast_fp16")]; tensor var_1412_interleave_0 = const()[name = tensor("op_1412_interleave_0"), val = tensor(false)]; tensor var_1412_cast_fp16 = concat(axis = var_797, interleave = var_1412_interleave_0, values = (var_1354_cast_fp16, var_1356_cast_fp16, var_1358_cast_fp16, var_1360_cast_fp16))[name = tensor("op_1412_cast_fp16")]; tensor var_1414_interleave_0 = const()[name = tensor("op_1414_interleave_0"), val = tensor(false)]; tensor var_1414_cast_fp16 = concat(axis = var_797, interleave = var_1414_interleave_0, values = (var_1362_cast_fp16, var_1364_cast_fp16, var_1366_cast_fp16, var_1368_cast_fp16))[name = tensor("op_1414_cast_fp16")]; tensor var_1416_interleave_0 = const()[name = tensor("op_1416_interleave_0"), val = tensor(false)]; tensor var_1416_cast_fp16 = concat(axis = var_797, interleave = var_1416_interleave_0, values = (var_1370_cast_fp16, var_1372_cast_fp16, var_1374_cast_fp16, var_1376_cast_fp16))[name = tensor("op_1416_cast_fp16")]; tensor var_1418_interleave_0 = const()[name = tensor("op_1418_interleave_0"), val = tensor(false)]; tensor var_1418_cast_fp16 = concat(axis = var_797, interleave = var_1418_interleave_0, values = (var_1378_cast_fp16, var_1380_cast_fp16, var_1382_cast_fp16, var_1384_cast_fp16))[name = tensor("op_1418_cast_fp16")]; tensor var_1420_interleave_0 = const()[name = tensor("op_1420_interleave_0"), val = tensor(false)]; tensor var_1420_cast_fp16 = concat(axis = var_797, interleave = var_1420_interleave_0, values = (var_1386_cast_fp16, var_1388_cast_fp16, var_1390_cast_fp16, var_1392_cast_fp16))[name = tensor("op_1420_cast_fp16")]; tensor var_1422_interleave_0 = const()[name = tensor("op_1422_interleave_0"), val = tensor(false)]; tensor var_1422_cast_fp16 = concat(axis = var_797, interleave = var_1422_interleave_0, values = (var_1394_cast_fp16, var_1396_cast_fp16, var_1398_cast_fp16, var_1400_cast_fp16))[name = tensor("op_1422_cast_fp16")]; tensor var_1424_interleave_0 = const()[name = tensor("op_1424_interleave_0"), val = tensor(false)]; tensor var_1424_cast_fp16 = concat(axis = var_797, interleave = var_1424_interleave_0, values = (var_1402_cast_fp16, var_1404_cast_fp16, var_1406_cast_fp16, var_1408_cast_fp16))[name = tensor("op_1424_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_810, interleave = input_9_interleave_0, values = (var_1410_cast_fp16, var_1412_cast_fp16, var_1414_cast_fp16, var_1416_cast_fp16, var_1418_cast_fp16, var_1420_cast_fp16, var_1422_cast_fp16, var_1424_cast_fp16))[name = tensor("input_9_cast_fp16")]; tensor var_1429 = const()[name = tensor("op_1429"), val = tensor([1, 1])]; tensor var_1431 = const()[name = tensor("op_1431"), 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(11241344)))]; 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(11765696)))]; tensor obj_7_cast_fp16 = conv(bias = layers_1_self_attn_o_proj_bias_to_fp16, dilations = var_1431, groups = var_810, pad = obj_7_pad_0, pad_type = obj_7_pad_type_0, strides = var_1429, 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_1437 = const()[name = tensor("op_1437"), val = tensor([1])]; tensor channels_mean_7_cast_fp16 = reduce_mean(axes = var_1437, keep_dims = var_811, 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_1441 = const()[name = tensor("op_1441"), val = tensor([1])]; tensor var_1442_cast_fp16 = reduce_mean(axes = var_1441, keep_dims = var_811, x = zero_mean_sq_7_cast_fp16)[name = tensor("op_1442_cast_fp16")]; tensor var_1443_to_fp16 = const()[name = tensor("op_1443_to_fp16"), val = tensor(0x1.5p-17)]; tensor var_1444_cast_fp16 = add(x = var_1442_cast_fp16, y = var_1443_to_fp16)[name = tensor("op_1444_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_1444_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(11766784)))]; 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(11767872)))]; 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_1455 = const()[name = tensor("op_1455"), val = tensor([1, 1])]; tensor var_1457 = const()[name = tensor("op_1457"), 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(11768960)))]; 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(13866176)))]; tensor input_13_cast_fp16 = conv(bias = layers_1_fc1_bias_to_fp16, dilations = var_1457, groups = var_810, pad = input_13_pad_0, pad_type = input_13_pad_type_0, strides = var_1455, 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_1463 = const()[name = tensor("op_1463"), val = tensor([1, 1])]; tensor var_1465 = const()[name = tensor("op_1465"), 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(13870336)))]; 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(15967552)))]; tensor hidden_states_7_cast_fp16 = conv(bias = layers_1_fc2_bias_to_fp16, dilations = var_1465, groups = var_810, pad = hidden_states_7_pad_0, pad_type = hidden_states_7_pad_type_0, strides = var_1463, 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_1472 = const()[name = tensor("op_1472"), val = tensor(3)]; tensor var_1485 = const()[name = tensor("op_1485"), val = tensor(1)]; tensor var_1486 = const()[name = tensor("op_1486"), val = tensor(true)]; tensor var_1496 = const()[name = tensor("op_1496"), val = tensor([1])]; tensor channels_mean_9_cast_fp16 = reduce_mean(axes = var_1496, keep_dims = var_1486, 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_1500 = const()[name = tensor("op_1500"), val = tensor([1])]; tensor var_1501_cast_fp16 = reduce_mean(axes = var_1500, keep_dims = var_1486, x = zero_mean_sq_9_cast_fp16)[name = tensor("op_1501_cast_fp16")]; tensor var_1502_to_fp16 = const()[name = tensor("op_1502_to_fp16"), val = tensor(0x1.5p-17)]; tensor var_1503_cast_fp16 = add(x = var_1501_cast_fp16, y = var_1502_to_fp16)[name = tensor("op_1503_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_1503_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(15968640)))]; 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(15969728)))]; 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_1518 = const()[name = tensor("op_1518"), val = tensor([1, 1])]; tensor var_1520 = const()[name = tensor("op_1520"), 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(15970816)))]; 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(16495168)))]; tensor query_5_cast_fp16 = conv(bias = layers_2_self_attn_q_proj_bias_to_fp16, dilations = var_1520, groups = var_1485, pad = query_5_pad_0, pad_type = query_5_pad_type_0, strides = var_1518, weight = layers_2_self_attn_q_proj_weight_to_fp16, x = obj_9_cast_fp16)[name = tensor("query_5_cast_fp16")]; tensor var_1524 = const()[name = tensor("op_1524"), val = tensor([1, 1])]; tensor var_1526 = const()[name = tensor("op_1526"), 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(16496256)))]; tensor key_5_cast_fp16 = conv(dilations = var_1526, groups = var_1485, pad = key_5_pad_0, pad_type = key_5_pad_type_0, strides = var_1524, weight = layers_2_self_attn_k_proj_weight_to_fp16, x = obj_9_cast_fp16)[name = tensor("key_5_cast_fp16")]; tensor var_1531 = const()[name = tensor("op_1531"), val = tensor([1, 1])]; tensor var_1533 = const()[name = tensor("op_1533"), 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(17020608)))]; 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(17544960)))]; tensor value_5_cast_fp16 = conv(bias = layers_2_self_attn_v_proj_bias_to_fp16, dilations = var_1533, groups = var_1485, pad = value_5_pad_0, pad_type = value_5_pad_type_0, strides = var_1531, weight = layers_2_self_attn_v_proj_weight_to_fp16, x = obj_9_cast_fp16)[name = tensor("value_5_cast_fp16")]; tensor var_1540_begin_0 = const()[name = tensor("op_1540_begin_0"), val = tensor([0, 0, 0, 0])]; tensor var_1540_end_0 = const()[name = tensor("op_1540_end_0"), val = tensor([1, 64, 1, 1500])]; tensor var_1540_end_mask_0 = const()[name = tensor("op_1540_end_mask_0"), val = tensor([true, false, true, true])]; tensor var_1540_cast_fp16 = slice_by_index(begin = var_1540_begin_0, end = var_1540_end_0, end_mask = var_1540_end_mask_0, x = query_5_cast_fp16)[name = tensor("op_1540_cast_fp16")]; tensor var_1544_begin_0 = const()[name = tensor("op_1544_begin_0"), val = tensor([0, 64, 0, 0])]; tensor var_1544_end_0 = const()[name = tensor("op_1544_end_0"), val = tensor([1, 128, 1, 1500])]; tensor var_1544_end_mask_0 = const()[name = tensor("op_1544_end_mask_0"), val = tensor([true, false, true, true])]; tensor var_1544_cast_fp16 = slice_by_index(begin = var_1544_begin_0, end = var_1544_end_0, end_mask = var_1544_end_mask_0, x = query_5_cast_fp16)[name = tensor("op_1544_cast_fp16")]; tensor var_1548_begin_0 = const()[name = tensor("op_1548_begin_0"), val = tensor([0, 128, 0, 0])]; tensor var_1548_end_0 = const()[name = tensor("op_1548_end_0"), val = tensor([1, 192, 1, 1500])]; tensor var_1548_end_mask_0 = const()[name = tensor("op_1548_end_mask_0"), val = tensor([true, false, true, true])]; tensor var_1548_cast_fp16 = slice_by_index(begin = var_1548_begin_0, end = var_1548_end_0, end_mask = var_1548_end_mask_0, x = query_5_cast_fp16)[name = tensor("op_1548_cast_fp16")]; tensor var_1552_begin_0 = const()[name = tensor("op_1552_begin_0"), val = tensor([0, 192, 0, 0])]; tensor var_1552_end_0 = const()[name = tensor("op_1552_end_0"), val = tensor([1, 256, 1, 1500])]; tensor var_1552_end_mask_0 = const()[name = tensor("op_1552_end_mask_0"), val = tensor([true, false, true, true])]; tensor var_1552_cast_fp16 = slice_by_index(begin = var_1552_begin_0, end = var_1552_end_0, end_mask = var_1552_end_mask_0, x = query_5_cast_fp16)[name = tensor("op_1552_cast_fp16")]; tensor var_1556_begin_0 = const()[name = tensor("op_1556_begin_0"), val = tensor([0, 256, 0, 0])]; tensor var_1556_end_0 = const()[name = tensor("op_1556_end_0"), val = tensor([1, 320, 1, 1500])]; tensor var_1556_end_mask_0 = const()[name = tensor("op_1556_end_mask_0"), val = tensor([true, false, true, true])]; tensor var_1556_cast_fp16 = slice_by_index(begin = var_1556_begin_0, end = var_1556_end_0, end_mask = var_1556_end_mask_0, x = query_5_cast_fp16)[name = tensor("op_1556_cast_fp16")]; tensor var_1560_begin_0 = const()[name = tensor("op_1560_begin_0"), val = tensor([0, 320, 0, 0])]; tensor var_1560_end_0 = const()[name = tensor("op_1560_end_0"), val = tensor([1, 384, 1, 1500])]; tensor var_1560_end_mask_0 = const()[name = tensor("op_1560_end_mask_0"), val = tensor([true, false, true, true])]; tensor var_1560_cast_fp16 = slice_by_index(begin = var_1560_begin_0, end = var_1560_end_0, end_mask = var_1560_end_mask_0, x = query_5_cast_fp16)[name = tensor("op_1560_cast_fp16")]; tensor var_1564_begin_0 = const()[name = tensor("op_1564_begin_0"), val = tensor([0, 384, 0, 0])]; tensor var_1564_end_0 = const()[name = tensor("op_1564_end_0"), val = tensor([1, 448, 1, 1500])]; tensor var_1564_end_mask_0 = const()[name = tensor("op_1564_end_mask_0"), val = tensor([true, false, true, true])]; tensor var_1564_cast_fp16 = slice_by_index(begin = var_1564_begin_0, end = var_1564_end_0, end_mask = var_1564_end_mask_0, x = query_5_cast_fp16)[name = tensor("op_1564_cast_fp16")]; tensor var_1568_begin_0 = const()[name = tensor("op_1568_begin_0"), val = tensor([0, 448, 0, 0])]; tensor var_1568_end_0 = const()[name = tensor("op_1568_end_0"), val = tensor([1, 512, 1, 1500])]; tensor var_1568_end_mask_0 = const()[name = tensor("op_1568_end_mask_0"), val = tensor([true, false, true, true])]; tensor var_1568_cast_fp16 = slice_by_index(begin = var_1568_begin_0, end = var_1568_end_0, end_mask = var_1568_end_mask_0, x = query_5_cast_fp16)[name = tensor("op_1568_cast_fp16")]; tensor var_1577_begin_0 = const()[name = tensor("op_1577_begin_0"), val = tensor([0, 0, 0, 0])]; tensor var_1577_end_0 = const()[name = tensor("op_1577_end_0"), val = tensor([1, 64, 1, 375])]; tensor var_1577_end_mask_0 = const()[name = tensor("op_1577_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_1577_cast_fp16 = slice_by_index(begin = var_1577_begin_0, end = var_1577_end_0, end_mask = var_1577_end_mask_0, x = var_1540_cast_fp16)[name = tensor("op_1577_cast_fp16")]; tensor var_1584_begin_0 = const()[name = tensor("op_1584_begin_0"), val = tensor([0, 0, 0, 375])]; tensor var_1584_end_0 = const()[name = tensor("op_1584_end_0"), val = tensor([1, 64, 1, 750])]; tensor var_1584_end_mask_0 = const()[name = tensor("op_1584_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_1584_cast_fp16 = slice_by_index(begin = var_1584_begin_0, end = var_1584_end_0, end_mask = var_1584_end_mask_0, x = var_1540_cast_fp16)[name = tensor("op_1584_cast_fp16")]; tensor var_1591_begin_0 = const()[name = tensor("op_1591_begin_0"), val = tensor([0, 0, 0, 750])]; tensor var_1591_end_0 = const()[name = tensor("op_1591_end_0"), val = tensor([1, 64, 1, 1125])]; tensor var_1591_end_mask_0 = const()[name = tensor("op_1591_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_1591_cast_fp16 = slice_by_index(begin = var_1591_begin_0, end = var_1591_end_0, end_mask = var_1591_end_mask_0, x = var_1540_cast_fp16)[name = tensor("op_1591_cast_fp16")]; tensor var_1598_begin_0 = const()[name = tensor("op_1598_begin_0"), val = tensor([0, 0, 0, 1125])]; tensor var_1598_end_0 = const()[name = tensor("op_1598_end_0"), val = tensor([1, 64, 1, 1500])]; tensor var_1598_end_mask_0 = const()[name = tensor("op_1598_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_1598_cast_fp16 = slice_by_index(begin = var_1598_begin_0, end = var_1598_end_0, end_mask = var_1598_end_mask_0, x = var_1540_cast_fp16)[name = tensor("op_1598_cast_fp16")]; tensor var_1605_begin_0 = const()[name = tensor("op_1605_begin_0"), val = tensor([0, 0, 0, 0])]; tensor var_1605_end_0 = const()[name = tensor("op_1605_end_0"), val = tensor([1, 64, 1, 375])]; tensor var_1605_end_mask_0 = const()[name = tensor("op_1605_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_1605_cast_fp16 = slice_by_index(begin = var_1605_begin_0, end = var_1605_end_0, end_mask = var_1605_end_mask_0, x = var_1544_cast_fp16)[name = tensor("op_1605_cast_fp16")]; tensor var_1612_begin_0 = const()[name = tensor("op_1612_begin_0"), val = tensor([0, 0, 0, 375])]; tensor var_1612_end_0 = const()[name = tensor("op_1612_end_0"), val = tensor([1, 64, 1, 750])]; tensor var_1612_end_mask_0 = const()[name = tensor("op_1612_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_1612_cast_fp16 = slice_by_index(begin = var_1612_begin_0, end = var_1612_end_0, end_mask = var_1612_end_mask_0, x = var_1544_cast_fp16)[name = tensor("op_1612_cast_fp16")]; tensor var_1619_begin_0 = const()[name = tensor("op_1619_begin_0"), val = tensor([0, 0, 0, 750])]; tensor var_1619_end_0 = const()[name = tensor("op_1619_end_0"), val = tensor([1, 64, 1, 1125])]; tensor var_1619_end_mask_0 = const()[name = tensor("op_1619_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_1619_cast_fp16 = slice_by_index(begin = var_1619_begin_0, end = var_1619_end_0, end_mask = var_1619_end_mask_0, x = var_1544_cast_fp16)[name = tensor("op_1619_cast_fp16")]; tensor var_1626_begin_0 = const()[name = tensor("op_1626_begin_0"), val = tensor([0, 0, 0, 1125])]; tensor var_1626_end_0 = const()[name = tensor("op_1626_end_0"), val = tensor([1, 64, 1, 1500])]; tensor var_1626_end_mask_0 = const()[name = tensor("op_1626_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_1626_cast_fp16 = slice_by_index(begin = var_1626_begin_0, end = var_1626_end_0, end_mask = var_1626_end_mask_0, x = var_1544_cast_fp16)[name = tensor("op_1626_cast_fp16")]; tensor var_1633_begin_0 = const()[name = tensor("op_1633_begin_0"), val = tensor([0, 0, 0, 0])]; tensor var_1633_end_0 = const()[name = tensor("op_1633_end_0"), val = tensor([1, 64, 1, 375])]; tensor var_1633_end_mask_0 = const()[name = tensor("op_1633_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_1633_cast_fp16 = slice_by_index(begin = var_1633_begin_0, end = var_1633_end_0, end_mask = var_1633_end_mask_0, x = var_1548_cast_fp16)[name = tensor("op_1633_cast_fp16")]; tensor var_1640_begin_0 = const()[name = tensor("op_1640_begin_0"), val = tensor([0, 0, 0, 375])]; tensor var_1640_end_0 = const()[name = tensor("op_1640_end_0"), val = tensor([1, 64, 1, 750])]; tensor var_1640_end_mask_0 = const()[name = tensor("op_1640_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_1640_cast_fp16 = slice_by_index(begin = var_1640_begin_0, end = var_1640_end_0, end_mask = var_1640_end_mask_0, x = var_1548_cast_fp16)[name = tensor("op_1640_cast_fp16")]; tensor var_1647_begin_0 = const()[name = tensor("op_1647_begin_0"), val = tensor([0, 0, 0, 750])]; tensor var_1647_end_0 = const()[name = tensor("op_1647_end_0"), val = tensor([1, 64, 1, 1125])]; tensor var_1647_end_mask_0 = const()[name = tensor("op_1647_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_1647_cast_fp16 = slice_by_index(begin = var_1647_begin_0, end = var_1647_end_0, end_mask = var_1647_end_mask_0, x = var_1548_cast_fp16)[name = tensor("op_1647_cast_fp16")]; tensor var_1654_begin_0 = const()[name = tensor("op_1654_begin_0"), val = tensor([0, 0, 0, 1125])]; tensor var_1654_end_0 = const()[name = tensor("op_1654_end_0"), val = tensor([1, 64, 1, 1500])]; tensor var_1654_end_mask_0 = const()[name = tensor("op_1654_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_1654_cast_fp16 = slice_by_index(begin = var_1654_begin_0, end = var_1654_end_0, end_mask = var_1654_end_mask_0, x = var_1548_cast_fp16)[name = tensor("op_1654_cast_fp16")]; tensor var_1661_begin_0 = const()[name = tensor("op_1661_begin_0"), val = tensor([0, 0, 0, 0])]; tensor var_1661_end_0 = const()[name = tensor("op_1661_end_0"), val = tensor([1, 64, 1, 375])]; tensor var_1661_end_mask_0 = const()[name = tensor("op_1661_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_1661_cast_fp16 = slice_by_index(begin = var_1661_begin_0, end = var_1661_end_0, end_mask = var_1661_end_mask_0, x = var_1552_cast_fp16)[name = tensor("op_1661_cast_fp16")]; tensor var_1668_begin_0 = const()[name = tensor("op_1668_begin_0"), val = tensor([0, 0, 0, 375])]; tensor var_1668_end_0 = const()[name = tensor("op_1668_end_0"), val = tensor([1, 64, 1, 750])]; tensor var_1668_end_mask_0 = const()[name = tensor("op_1668_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_1668_cast_fp16 = slice_by_index(begin = var_1668_begin_0, end = var_1668_end_0, end_mask = var_1668_end_mask_0, x = var_1552_cast_fp16)[name = tensor("op_1668_cast_fp16")]; tensor var_1675_begin_0 = const()[name = tensor("op_1675_begin_0"), val = tensor([0, 0, 0, 750])]; tensor var_1675_end_0 = const()[name = tensor("op_1675_end_0"), val = tensor([1, 64, 1, 1125])]; tensor var_1675_end_mask_0 = const()[name = tensor("op_1675_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_1675_cast_fp16 = slice_by_index(begin = var_1675_begin_0, end = var_1675_end_0, end_mask = var_1675_end_mask_0, x = var_1552_cast_fp16)[name = tensor("op_1675_cast_fp16")]; tensor var_1682_begin_0 = const()[name = tensor("op_1682_begin_0"), val = tensor([0, 0, 0, 1125])]; tensor var_1682_end_0 = const()[name = tensor("op_1682_end_0"), val = tensor([1, 64, 1, 1500])]; tensor var_1682_end_mask_0 = const()[name = tensor("op_1682_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_1682_cast_fp16 = slice_by_index(begin = var_1682_begin_0, end = var_1682_end_0, end_mask = var_1682_end_mask_0, x = var_1552_cast_fp16)[name = tensor("op_1682_cast_fp16")]; tensor var_1689_begin_0 = const()[name = tensor("op_1689_begin_0"), val = tensor([0, 0, 0, 0])]; tensor var_1689_end_0 = const()[name = tensor("op_1689_end_0"), val = tensor([1, 64, 1, 375])]; tensor var_1689_end_mask_0 = const()[name = tensor("op_1689_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_1689_cast_fp16 = slice_by_index(begin = var_1689_begin_0, end = var_1689_end_0, end_mask = var_1689_end_mask_0, x = var_1556_cast_fp16)[name = tensor("op_1689_cast_fp16")]; tensor var_1696_begin_0 = const()[name = tensor("op_1696_begin_0"), val = tensor([0, 0, 0, 375])]; tensor var_1696_end_0 = const()[name = tensor("op_1696_end_0"), val = tensor([1, 64, 1, 750])]; tensor var_1696_end_mask_0 = const()[name = tensor("op_1696_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_1696_cast_fp16 = slice_by_index(begin = var_1696_begin_0, end = var_1696_end_0, end_mask = var_1696_end_mask_0, x = var_1556_cast_fp16)[name = tensor("op_1696_cast_fp16")]; tensor var_1703_begin_0 = const()[name = tensor("op_1703_begin_0"), val = tensor([0, 0, 0, 750])]; tensor var_1703_end_0 = const()[name = tensor("op_1703_end_0"), val = tensor([1, 64, 1, 1125])]; tensor var_1703_end_mask_0 = const()[name = tensor("op_1703_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_1703_cast_fp16 = slice_by_index(begin = var_1703_begin_0, end = var_1703_end_0, end_mask = var_1703_end_mask_0, x = var_1556_cast_fp16)[name = tensor("op_1703_cast_fp16")]; tensor var_1710_begin_0 = const()[name = tensor("op_1710_begin_0"), val = tensor([0, 0, 0, 1125])]; tensor var_1710_end_0 = const()[name = tensor("op_1710_end_0"), val = tensor([1, 64, 1, 1500])]; tensor var_1710_end_mask_0 = const()[name = tensor("op_1710_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_1710_cast_fp16 = slice_by_index(begin = var_1710_begin_0, end = var_1710_end_0, end_mask = var_1710_end_mask_0, x = var_1556_cast_fp16)[name = tensor("op_1710_cast_fp16")]; tensor var_1717_begin_0 = const()[name = tensor("op_1717_begin_0"), val = tensor([0, 0, 0, 0])]; tensor var_1717_end_0 = const()[name = tensor("op_1717_end_0"), val = tensor([1, 64, 1, 375])]; tensor var_1717_end_mask_0 = const()[name = tensor("op_1717_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_1717_cast_fp16 = slice_by_index(begin = var_1717_begin_0, end = var_1717_end_0, end_mask = var_1717_end_mask_0, x = var_1560_cast_fp16)[name = tensor("op_1717_cast_fp16")]; tensor var_1724_begin_0 = const()[name = tensor("op_1724_begin_0"), val = tensor([0, 0, 0, 375])]; tensor var_1724_end_0 = const()[name = tensor("op_1724_end_0"), val = tensor([1, 64, 1, 750])]; tensor var_1724_end_mask_0 = const()[name = tensor("op_1724_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_1724_cast_fp16 = slice_by_index(begin = var_1724_begin_0, end = var_1724_end_0, end_mask = var_1724_end_mask_0, x = var_1560_cast_fp16)[name = tensor("op_1724_cast_fp16")]; tensor var_1731_begin_0 = const()[name = tensor("op_1731_begin_0"), val = tensor([0, 0, 0, 750])]; tensor var_1731_end_0 = const()[name = tensor("op_1731_end_0"), val = tensor([1, 64, 1, 1125])]; tensor var_1731_end_mask_0 = const()[name = tensor("op_1731_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_1731_cast_fp16 = slice_by_index(begin = var_1731_begin_0, end = var_1731_end_0, end_mask = var_1731_end_mask_0, x = var_1560_cast_fp16)[name = tensor("op_1731_cast_fp16")]; tensor var_1738_begin_0 = const()[name = tensor("op_1738_begin_0"), val = tensor([0, 0, 0, 1125])]; tensor var_1738_end_0 = const()[name = tensor("op_1738_end_0"), val = tensor([1, 64, 1, 1500])]; tensor var_1738_end_mask_0 = const()[name = tensor("op_1738_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_1738_cast_fp16 = slice_by_index(begin = var_1738_begin_0, end = var_1738_end_0, end_mask = var_1738_end_mask_0, x = var_1560_cast_fp16)[name = tensor("op_1738_cast_fp16")]; tensor var_1745_begin_0 = const()[name = tensor("op_1745_begin_0"), val = tensor([0, 0, 0, 0])]; tensor var_1745_end_0 = const()[name = tensor("op_1745_end_0"), val = tensor([1, 64, 1, 375])]; tensor var_1745_end_mask_0 = const()[name = tensor("op_1745_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_1745_cast_fp16 = slice_by_index(begin = var_1745_begin_0, end = var_1745_end_0, end_mask = var_1745_end_mask_0, x = var_1564_cast_fp16)[name = tensor("op_1745_cast_fp16")]; tensor var_1752_begin_0 = const()[name = tensor("op_1752_begin_0"), val = tensor([0, 0, 0, 375])]; tensor var_1752_end_0 = const()[name = tensor("op_1752_end_0"), val = tensor([1, 64, 1, 750])]; tensor var_1752_end_mask_0 = const()[name = tensor("op_1752_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_1752_cast_fp16 = slice_by_index(begin = var_1752_begin_0, end = var_1752_end_0, end_mask = var_1752_end_mask_0, x = var_1564_cast_fp16)[name = tensor("op_1752_cast_fp16")]; tensor var_1759_begin_0 = const()[name = tensor("op_1759_begin_0"), val = tensor([0, 0, 0, 750])]; tensor var_1759_end_0 = const()[name = tensor("op_1759_end_0"), val = tensor([1, 64, 1, 1125])]; tensor var_1759_end_mask_0 = const()[name = tensor("op_1759_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_1759_cast_fp16 = slice_by_index(begin = var_1759_begin_0, end = var_1759_end_0, end_mask = var_1759_end_mask_0, x = var_1564_cast_fp16)[name = tensor("op_1759_cast_fp16")]; tensor var_1766_begin_0 = const()[name = tensor("op_1766_begin_0"), val = tensor([0, 0, 0, 1125])]; tensor var_1766_end_0 = const()[name = tensor("op_1766_end_0"), val = tensor([1, 64, 1, 1500])]; tensor var_1766_end_mask_0 = const()[name = tensor("op_1766_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_1766_cast_fp16 = slice_by_index(begin = var_1766_begin_0, end = var_1766_end_0, end_mask = var_1766_end_mask_0, x = var_1564_cast_fp16)[name = tensor("op_1766_cast_fp16")]; tensor var_1773_begin_0 = const()[name = tensor("op_1773_begin_0"), val = tensor([0, 0, 0, 0])]; tensor var_1773_end_0 = const()[name = tensor("op_1773_end_0"), val = tensor([1, 64, 1, 375])]; tensor var_1773_end_mask_0 = const()[name = tensor("op_1773_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_1773_cast_fp16 = slice_by_index(begin = var_1773_begin_0, end = var_1773_end_0, end_mask = var_1773_end_mask_0, x = var_1568_cast_fp16)[name = tensor("op_1773_cast_fp16")]; tensor var_1780_begin_0 = const()[name = tensor("op_1780_begin_0"), val = tensor([0, 0, 0, 375])]; tensor var_1780_end_0 = const()[name = tensor("op_1780_end_0"), val = tensor([1, 64, 1, 750])]; tensor var_1780_end_mask_0 = const()[name = tensor("op_1780_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_1780_cast_fp16 = slice_by_index(begin = var_1780_begin_0, end = var_1780_end_0, end_mask = var_1780_end_mask_0, x = var_1568_cast_fp16)[name = tensor("op_1780_cast_fp16")]; tensor var_1787_begin_0 = const()[name = tensor("op_1787_begin_0"), val = tensor([0, 0, 0, 750])]; tensor var_1787_end_0 = const()[name = tensor("op_1787_end_0"), val = tensor([1, 64, 1, 1125])]; tensor var_1787_end_mask_0 = const()[name = tensor("op_1787_end_mask_0"), val = tensor([true, true, true, false])]; 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 = var_1568_cast_fp16)[name = tensor("op_1787_cast_fp16")]; tensor var_1794_begin_0 = const()[name = tensor("op_1794_begin_0"), val = tensor([0, 0, 0, 1125])]; tensor var_1794_end_0 = const()[name = tensor("op_1794_end_0"), val = tensor([1, 64, 1, 1500])]; tensor var_1794_end_mask_0 = const()[name = tensor("op_1794_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_1794_cast_fp16 = slice_by_index(begin = var_1794_begin_0, end = var_1794_end_0, end_mask = var_1794_end_mask_0, x = var_1568_cast_fp16)[name = tensor("op_1794_cast_fp16")]; tensor k_5_perm_0 = const()[name = tensor("k_5_perm_0"), val = tensor([0, 3, 2, 1])]; tensor var_1799_begin_0 = const()[name = tensor("op_1799_begin_0"), val = tensor([0, 0, 0, 0])]; tensor var_1799_end_0 = const()[name = tensor("op_1799_end_0"), val = tensor([1, 1500, 1, 64])]; tensor var_1799_end_mask_0 = const()[name = tensor("op_1799_end_mask_0"), val = tensor([true, true, true, false])]; tensor transpose_3 = transpose(perm = k_5_perm_0, x = key_5_cast_fp16)[name = tensor("transpose_3")]; 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 = transpose_3)[name = tensor("op_1799_cast_fp16")]; tensor var_1803_begin_0 = const()[name = tensor("op_1803_begin_0"), val = tensor([0, 0, 0, 64])]; tensor var_1803_end_0 = const()[name = tensor("op_1803_end_0"), val = tensor([1, 1500, 1, 128])]; tensor var_1803_end_mask_0 = const()[name = tensor("op_1803_end_mask_0"), val = tensor([true, true, true, false])]; 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 = transpose_3)[name = tensor("op_1803_cast_fp16")]; tensor var_1807_begin_0 = const()[name = tensor("op_1807_begin_0"), val = tensor([0, 0, 0, 128])]; tensor var_1807_end_0 = const()[name = tensor("op_1807_end_0"), val = tensor([1, 1500, 1, 192])]; tensor var_1807_end_mask_0 = const()[name = tensor("op_1807_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_1807_cast_fp16 = slice_by_index(begin = var_1807_begin_0, end = var_1807_end_0, end_mask = var_1807_end_mask_0, x = transpose_3)[name = tensor("op_1807_cast_fp16")]; tensor var_1811_begin_0 = const()[name = tensor("op_1811_begin_0"), val = tensor([0, 0, 0, 192])]; tensor var_1811_end_0 = const()[name = tensor("op_1811_end_0"), val = tensor([1, 1500, 1, 256])]; tensor var_1811_end_mask_0 = const()[name = tensor("op_1811_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_1811_cast_fp16 = slice_by_index(begin = var_1811_begin_0, end = var_1811_end_0, end_mask = var_1811_end_mask_0, x = transpose_3)[name = tensor("op_1811_cast_fp16")]; tensor var_1815_begin_0 = const()[name = tensor("op_1815_begin_0"), val = tensor([0, 0, 0, 256])]; tensor var_1815_end_0 = const()[name = tensor("op_1815_end_0"), val = tensor([1, 1500, 1, 320])]; tensor var_1815_end_mask_0 = const()[name = tensor("op_1815_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_1815_cast_fp16 = slice_by_index(begin = var_1815_begin_0, end = var_1815_end_0, end_mask = var_1815_end_mask_0, x = transpose_3)[name = tensor("op_1815_cast_fp16")]; tensor var_1819_begin_0 = const()[name = tensor("op_1819_begin_0"), val = tensor([0, 0, 0, 320])]; tensor var_1819_end_0 = const()[name = tensor("op_1819_end_0"), val = tensor([1, 1500, 1, 384])]; 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 = transpose_3)[name = tensor("op_1819_cast_fp16")]; tensor var_1823_begin_0 = const()[name = tensor("op_1823_begin_0"), val = tensor([0, 0, 0, 384])]; tensor var_1823_end_0 = const()[name = tensor("op_1823_end_0"), val = tensor([1, 1500, 1, 448])]; tensor var_1823_end_mask_0 = const()[name = tensor("op_1823_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_1823_cast_fp16 = slice_by_index(begin = var_1823_begin_0, end = var_1823_end_0, end_mask = var_1823_end_mask_0, x = transpose_3)[name = tensor("op_1823_cast_fp16")]; tensor var_1827_begin_0 = const()[name = tensor("op_1827_begin_0"), val = tensor([0, 0, 0, 448])]; tensor var_1827_end_0 = const()[name = tensor("op_1827_end_0"), val = tensor([1, 1500, 1, 512])]; tensor var_1827_end_mask_0 = const()[name = tensor("op_1827_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_1827_cast_fp16 = slice_by_index(begin = var_1827_begin_0, end = var_1827_end_0, end_mask = var_1827_end_mask_0, x = transpose_3)[name = tensor("op_1827_cast_fp16")]; tensor var_1829_begin_0 = const()[name = tensor("op_1829_begin_0"), val = tensor([0, 0, 0, 0])]; tensor var_1829_end_0 = const()[name = tensor("op_1829_end_0"), val = tensor([1, 64, 1, 1500])]; tensor var_1829_end_mask_0 = const()[name = tensor("op_1829_end_mask_0"), val = tensor([true, false, true, true])]; tensor var_1829_cast_fp16 = slice_by_index(begin = var_1829_begin_0, end = var_1829_end_0, end_mask = var_1829_end_mask_0, x = value_5_cast_fp16)[name = tensor("op_1829_cast_fp16")]; tensor var_1833_begin_0 = const()[name = tensor("op_1833_begin_0"), val = tensor([0, 64, 0, 0])]; tensor var_1833_end_0 = const()[name = tensor("op_1833_end_0"), val = tensor([1, 128, 1, 1500])]; tensor var_1833_end_mask_0 = const()[name = tensor("op_1833_end_mask_0"), val = tensor([true, false, true, true])]; 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 = value_5_cast_fp16)[name = tensor("op_1833_cast_fp16")]; tensor var_1837_begin_0 = const()[name = tensor("op_1837_begin_0"), val = tensor([0, 128, 0, 0])]; tensor var_1837_end_0 = const()[name = tensor("op_1837_end_0"), val = tensor([1, 192, 1, 1500])]; tensor var_1837_end_mask_0 = const()[name = tensor("op_1837_end_mask_0"), val = tensor([true, false, true, true])]; tensor var_1837_cast_fp16 = slice_by_index(begin = var_1837_begin_0, end = var_1837_end_0, end_mask = var_1837_end_mask_0, x = value_5_cast_fp16)[name = tensor("op_1837_cast_fp16")]; tensor var_1841_begin_0 = const()[name = tensor("op_1841_begin_0"), val = tensor([0, 192, 0, 0])]; tensor var_1841_end_0 = const()[name = tensor("op_1841_end_0"), val = tensor([1, 256, 1, 1500])]; tensor var_1841_end_mask_0 = const()[name = tensor("op_1841_end_mask_0"), val = tensor([true, false, true, true])]; tensor var_1841_cast_fp16 = slice_by_index(begin = var_1841_begin_0, end = var_1841_end_0, end_mask = var_1841_end_mask_0, x = value_5_cast_fp16)[name = tensor("op_1841_cast_fp16")]; tensor var_1845_begin_0 = const()[name = tensor("op_1845_begin_0"), val = tensor([0, 256, 0, 0])]; tensor var_1845_end_0 = const()[name = tensor("op_1845_end_0"), val = tensor([1, 320, 1, 1500])]; tensor var_1845_end_mask_0 = const()[name = tensor("op_1845_end_mask_0"), val = tensor([true, false, true, true])]; tensor var_1845_cast_fp16 = slice_by_index(begin = var_1845_begin_0, end = var_1845_end_0, end_mask = var_1845_end_mask_0, x = value_5_cast_fp16)[name = tensor("op_1845_cast_fp16")]; tensor var_1849_begin_0 = const()[name = tensor("op_1849_begin_0"), val = tensor([0, 320, 0, 0])]; tensor var_1849_end_0 = const()[name = tensor("op_1849_end_0"), val = tensor([1, 384, 1, 1500])]; tensor var_1849_end_mask_0 = const()[name = tensor("op_1849_end_mask_0"), val = tensor([true, false, true, true])]; tensor var_1849_cast_fp16 = slice_by_index(begin = var_1849_begin_0, end = var_1849_end_0, end_mask = var_1849_end_mask_0, x = value_5_cast_fp16)[name = tensor("op_1849_cast_fp16")]; tensor var_1853_begin_0 = const()[name = tensor("op_1853_begin_0"), val = tensor([0, 384, 0, 0])]; tensor var_1853_end_0 = const()[name = tensor("op_1853_end_0"), val = tensor([1, 448, 1, 1500])]; tensor var_1853_end_mask_0 = const()[name = tensor("op_1853_end_mask_0"), val = tensor([true, false, true, true])]; tensor var_1853_cast_fp16 = slice_by_index(begin = var_1853_begin_0, end = var_1853_end_0, end_mask = var_1853_end_mask_0, x = value_5_cast_fp16)[name = tensor("op_1853_cast_fp16")]; tensor var_1857_begin_0 = const()[name = tensor("op_1857_begin_0"), val = tensor([0, 448, 0, 0])]; tensor var_1857_end_0 = const()[name = tensor("op_1857_end_0"), val = tensor([1, 512, 1, 1500])]; tensor var_1857_end_mask_0 = const()[name = tensor("op_1857_end_mask_0"), val = tensor([true, false, true, true])]; tensor var_1857_cast_fp16 = slice_by_index(begin = var_1857_begin_0, end = var_1857_end_0, end_mask = var_1857_end_mask_0, x = value_5_cast_fp16)[name = tensor("op_1857_cast_fp16")]; tensor var_1861_equation_0 = const()[name = tensor("op_1861_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_1861_cast_fp16 = einsum(equation = var_1861_equation_0, values = (var_1799_cast_fp16, var_1577_cast_fp16))[name = tensor("op_1861_cast_fp16")]; tensor var_1862_to_fp16 = const()[name = tensor("op_1862_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_129_cast_fp16 = mul(x = var_1861_cast_fp16, y = var_1862_to_fp16)[name = tensor("aw_chunk_129_cast_fp16")]; tensor var_1865_equation_0 = const()[name = tensor("op_1865_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_1865_cast_fp16 = einsum(equation = var_1865_equation_0, values = (var_1799_cast_fp16, var_1584_cast_fp16))[name = tensor("op_1865_cast_fp16")]; tensor var_1866_to_fp16 = const()[name = tensor("op_1866_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_131_cast_fp16 = mul(x = var_1865_cast_fp16, y = var_1866_to_fp16)[name = tensor("aw_chunk_131_cast_fp16")]; tensor var_1869_equation_0 = const()[name = tensor("op_1869_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_1869_cast_fp16 = einsum(equation = var_1869_equation_0, values = (var_1799_cast_fp16, var_1591_cast_fp16))[name = tensor("op_1869_cast_fp16")]; tensor var_1870_to_fp16 = const()[name = tensor("op_1870_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_133_cast_fp16 = mul(x = var_1869_cast_fp16, y = var_1870_to_fp16)[name = tensor("aw_chunk_133_cast_fp16")]; tensor var_1873_equation_0 = const()[name = tensor("op_1873_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_1873_cast_fp16 = einsum(equation = var_1873_equation_0, values = (var_1799_cast_fp16, var_1598_cast_fp16))[name = tensor("op_1873_cast_fp16")]; tensor var_1874_to_fp16 = const()[name = tensor("op_1874_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_135_cast_fp16 = mul(x = var_1873_cast_fp16, y = var_1874_to_fp16)[name = tensor("aw_chunk_135_cast_fp16")]; tensor var_1877_equation_0 = const()[name = tensor("op_1877_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_1877_cast_fp16 = einsum(equation = var_1877_equation_0, values = (var_1803_cast_fp16, var_1605_cast_fp16))[name = tensor("op_1877_cast_fp16")]; tensor var_1878_to_fp16 = const()[name = tensor("op_1878_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_137_cast_fp16 = mul(x = var_1877_cast_fp16, y = var_1878_to_fp16)[name = tensor("aw_chunk_137_cast_fp16")]; tensor var_1881_equation_0 = const()[name = tensor("op_1881_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_1881_cast_fp16 = einsum(equation = var_1881_equation_0, values = (var_1803_cast_fp16, var_1612_cast_fp16))[name = tensor("op_1881_cast_fp16")]; tensor var_1882_to_fp16 = const()[name = tensor("op_1882_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_139_cast_fp16 = mul(x = var_1881_cast_fp16, y = var_1882_to_fp16)[name = tensor("aw_chunk_139_cast_fp16")]; tensor var_1885_equation_0 = const()[name = tensor("op_1885_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_1885_cast_fp16 = einsum(equation = var_1885_equation_0, values = (var_1803_cast_fp16, var_1619_cast_fp16))[name = tensor("op_1885_cast_fp16")]; tensor var_1886_to_fp16 = const()[name = tensor("op_1886_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_141_cast_fp16 = mul(x = var_1885_cast_fp16, y = var_1886_to_fp16)[name = tensor("aw_chunk_141_cast_fp16")]; tensor var_1889_equation_0 = const()[name = tensor("op_1889_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_1889_cast_fp16 = einsum(equation = var_1889_equation_0, values = (var_1803_cast_fp16, var_1626_cast_fp16))[name = tensor("op_1889_cast_fp16")]; tensor var_1890_to_fp16 = const()[name = tensor("op_1890_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_143_cast_fp16 = mul(x = var_1889_cast_fp16, y = var_1890_to_fp16)[name = tensor("aw_chunk_143_cast_fp16")]; tensor var_1893_equation_0 = const()[name = tensor("op_1893_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_1893_cast_fp16 = einsum(equation = var_1893_equation_0, values = (var_1807_cast_fp16, var_1633_cast_fp16))[name = tensor("op_1893_cast_fp16")]; tensor var_1894_to_fp16 = const()[name = tensor("op_1894_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_145_cast_fp16 = mul(x = var_1893_cast_fp16, y = var_1894_to_fp16)[name = tensor("aw_chunk_145_cast_fp16")]; tensor var_1897_equation_0 = const()[name = tensor("op_1897_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_1897_cast_fp16 = einsum(equation = var_1897_equation_0, values = (var_1807_cast_fp16, var_1640_cast_fp16))[name = tensor("op_1897_cast_fp16")]; tensor var_1898_to_fp16 = const()[name = tensor("op_1898_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_147_cast_fp16 = mul(x = var_1897_cast_fp16, y = var_1898_to_fp16)[name = tensor("aw_chunk_147_cast_fp16")]; tensor var_1901_equation_0 = const()[name = tensor("op_1901_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_1901_cast_fp16 = einsum(equation = var_1901_equation_0, values = (var_1807_cast_fp16, var_1647_cast_fp16))[name = tensor("op_1901_cast_fp16")]; tensor var_1902_to_fp16 = const()[name = tensor("op_1902_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_149_cast_fp16 = mul(x = var_1901_cast_fp16, y = var_1902_to_fp16)[name = tensor("aw_chunk_149_cast_fp16")]; tensor var_1905_equation_0 = const()[name = tensor("op_1905_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_1905_cast_fp16 = einsum(equation = var_1905_equation_0, values = (var_1807_cast_fp16, var_1654_cast_fp16))[name = tensor("op_1905_cast_fp16")]; tensor var_1906_to_fp16 = const()[name = tensor("op_1906_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_151_cast_fp16 = mul(x = var_1905_cast_fp16, y = var_1906_to_fp16)[name = tensor("aw_chunk_151_cast_fp16")]; tensor var_1909_equation_0 = const()[name = tensor("op_1909_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_1909_cast_fp16 = einsum(equation = var_1909_equation_0, values = (var_1811_cast_fp16, var_1661_cast_fp16))[name = tensor("op_1909_cast_fp16")]; tensor var_1910_to_fp16 = const()[name = tensor("op_1910_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_153_cast_fp16 = mul(x = var_1909_cast_fp16, y = var_1910_to_fp16)[name = tensor("aw_chunk_153_cast_fp16")]; tensor var_1913_equation_0 = const()[name = tensor("op_1913_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_1913_cast_fp16 = einsum(equation = var_1913_equation_0, values = (var_1811_cast_fp16, var_1668_cast_fp16))[name = tensor("op_1913_cast_fp16")]; tensor var_1914_to_fp16 = const()[name = tensor("op_1914_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_155_cast_fp16 = mul(x = var_1913_cast_fp16, y = var_1914_to_fp16)[name = tensor("aw_chunk_155_cast_fp16")]; tensor var_1917_equation_0 = const()[name = tensor("op_1917_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_1917_cast_fp16 = einsum(equation = var_1917_equation_0, values = (var_1811_cast_fp16, var_1675_cast_fp16))[name = tensor("op_1917_cast_fp16")]; tensor var_1918_to_fp16 = const()[name = tensor("op_1918_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_157_cast_fp16 = mul(x = var_1917_cast_fp16, y = var_1918_to_fp16)[name = tensor("aw_chunk_157_cast_fp16")]; tensor var_1921_equation_0 = const()[name = tensor("op_1921_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_1921_cast_fp16 = einsum(equation = var_1921_equation_0, values = (var_1811_cast_fp16, var_1682_cast_fp16))[name = tensor("op_1921_cast_fp16")]; tensor var_1922_to_fp16 = const()[name = tensor("op_1922_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_159_cast_fp16 = mul(x = var_1921_cast_fp16, y = var_1922_to_fp16)[name = tensor("aw_chunk_159_cast_fp16")]; tensor var_1925_equation_0 = const()[name = tensor("op_1925_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_1925_cast_fp16 = einsum(equation = var_1925_equation_0, values = (var_1815_cast_fp16, var_1689_cast_fp16))[name = tensor("op_1925_cast_fp16")]; tensor var_1926_to_fp16 = const()[name = tensor("op_1926_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_161_cast_fp16 = mul(x = var_1925_cast_fp16, y = var_1926_to_fp16)[name = tensor("aw_chunk_161_cast_fp16")]; tensor var_1929_equation_0 = const()[name = tensor("op_1929_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_1929_cast_fp16 = einsum(equation = var_1929_equation_0, values = (var_1815_cast_fp16, var_1696_cast_fp16))[name = tensor("op_1929_cast_fp16")]; tensor var_1930_to_fp16 = const()[name = tensor("op_1930_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_163_cast_fp16 = mul(x = var_1929_cast_fp16, y = var_1930_to_fp16)[name = tensor("aw_chunk_163_cast_fp16")]; tensor var_1933_equation_0 = const()[name = tensor("op_1933_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_1933_cast_fp16 = einsum(equation = var_1933_equation_0, values = (var_1815_cast_fp16, var_1703_cast_fp16))[name = tensor("op_1933_cast_fp16")]; tensor var_1934_to_fp16 = const()[name = tensor("op_1934_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_165_cast_fp16 = mul(x = var_1933_cast_fp16, y = var_1934_to_fp16)[name = tensor("aw_chunk_165_cast_fp16")]; tensor var_1937_equation_0 = const()[name = tensor("op_1937_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_1937_cast_fp16 = einsum(equation = var_1937_equation_0, values = (var_1815_cast_fp16, var_1710_cast_fp16))[name = tensor("op_1937_cast_fp16")]; tensor var_1938_to_fp16 = const()[name = tensor("op_1938_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_167_cast_fp16 = mul(x = var_1937_cast_fp16, y = var_1938_to_fp16)[name = tensor("aw_chunk_167_cast_fp16")]; tensor var_1941_equation_0 = const()[name = tensor("op_1941_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_1941_cast_fp16 = einsum(equation = var_1941_equation_0, values = (var_1819_cast_fp16, var_1717_cast_fp16))[name = tensor("op_1941_cast_fp16")]; tensor var_1942_to_fp16 = const()[name = tensor("op_1942_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_169_cast_fp16 = mul(x = var_1941_cast_fp16, y = var_1942_to_fp16)[name = tensor("aw_chunk_169_cast_fp16")]; tensor var_1945_equation_0 = const()[name = tensor("op_1945_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_1945_cast_fp16 = einsum(equation = var_1945_equation_0, values = (var_1819_cast_fp16, var_1724_cast_fp16))[name = tensor("op_1945_cast_fp16")]; tensor var_1946_to_fp16 = const()[name = tensor("op_1946_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_171_cast_fp16 = mul(x = var_1945_cast_fp16, y = var_1946_to_fp16)[name = tensor("aw_chunk_171_cast_fp16")]; tensor var_1949_equation_0 = const()[name = tensor("op_1949_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_1949_cast_fp16 = einsum(equation = var_1949_equation_0, values = (var_1819_cast_fp16, var_1731_cast_fp16))[name = tensor("op_1949_cast_fp16")]; tensor var_1950_to_fp16 = const()[name = tensor("op_1950_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_173_cast_fp16 = mul(x = var_1949_cast_fp16, y = var_1950_to_fp16)[name = tensor("aw_chunk_173_cast_fp16")]; tensor var_1953_equation_0 = const()[name = tensor("op_1953_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_1953_cast_fp16 = einsum(equation = var_1953_equation_0, values = (var_1819_cast_fp16, var_1738_cast_fp16))[name = tensor("op_1953_cast_fp16")]; tensor var_1954_to_fp16 = const()[name = tensor("op_1954_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_175_cast_fp16 = mul(x = var_1953_cast_fp16, y = var_1954_to_fp16)[name = tensor("aw_chunk_175_cast_fp16")]; tensor var_1957_equation_0 = const()[name = tensor("op_1957_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_1957_cast_fp16 = einsum(equation = var_1957_equation_0, values = (var_1823_cast_fp16, var_1745_cast_fp16))[name = tensor("op_1957_cast_fp16")]; tensor var_1958_to_fp16 = const()[name = tensor("op_1958_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_177_cast_fp16 = mul(x = var_1957_cast_fp16, y = var_1958_to_fp16)[name = tensor("aw_chunk_177_cast_fp16")]; tensor var_1961_equation_0 = const()[name = tensor("op_1961_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_1961_cast_fp16 = einsum(equation = var_1961_equation_0, values = (var_1823_cast_fp16, var_1752_cast_fp16))[name = tensor("op_1961_cast_fp16")]; tensor var_1962_to_fp16 = const()[name = tensor("op_1962_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_179_cast_fp16 = mul(x = var_1961_cast_fp16, y = var_1962_to_fp16)[name = tensor("aw_chunk_179_cast_fp16")]; tensor var_1965_equation_0 = const()[name = tensor("op_1965_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_1965_cast_fp16 = einsum(equation = var_1965_equation_0, values = (var_1823_cast_fp16, var_1759_cast_fp16))[name = tensor("op_1965_cast_fp16")]; tensor var_1966_to_fp16 = const()[name = tensor("op_1966_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_181_cast_fp16 = mul(x = var_1965_cast_fp16, y = var_1966_to_fp16)[name = tensor("aw_chunk_181_cast_fp16")]; tensor var_1969_equation_0 = const()[name = tensor("op_1969_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_1969_cast_fp16 = einsum(equation = var_1969_equation_0, values = (var_1823_cast_fp16, var_1766_cast_fp16))[name = tensor("op_1969_cast_fp16")]; tensor var_1970_to_fp16 = const()[name = tensor("op_1970_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_183_cast_fp16 = mul(x = var_1969_cast_fp16, y = var_1970_to_fp16)[name = tensor("aw_chunk_183_cast_fp16")]; tensor var_1973_equation_0 = const()[name = tensor("op_1973_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_1973_cast_fp16 = einsum(equation = var_1973_equation_0, values = (var_1827_cast_fp16, var_1773_cast_fp16))[name = tensor("op_1973_cast_fp16")]; tensor var_1974_to_fp16 = const()[name = tensor("op_1974_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_185_cast_fp16 = mul(x = var_1973_cast_fp16, y = var_1974_to_fp16)[name = tensor("aw_chunk_185_cast_fp16")]; tensor var_1977_equation_0 = const()[name = tensor("op_1977_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_1977_cast_fp16 = einsum(equation = var_1977_equation_0, values = (var_1827_cast_fp16, var_1780_cast_fp16))[name = tensor("op_1977_cast_fp16")]; tensor var_1978_to_fp16 = const()[name = tensor("op_1978_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_187_cast_fp16 = mul(x = var_1977_cast_fp16, y = var_1978_to_fp16)[name = tensor("aw_chunk_187_cast_fp16")]; tensor var_1981_equation_0 = const()[name = tensor("op_1981_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_1981_cast_fp16 = einsum(equation = var_1981_equation_0, values = (var_1827_cast_fp16, var_1787_cast_fp16))[name = tensor("op_1981_cast_fp16")]; tensor var_1982_to_fp16 = const()[name = tensor("op_1982_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_189_cast_fp16 = mul(x = var_1981_cast_fp16, y = var_1982_to_fp16)[name = tensor("aw_chunk_189_cast_fp16")]; tensor var_1985_equation_0 = const()[name = tensor("op_1985_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_1985_cast_fp16 = einsum(equation = var_1985_equation_0, values = (var_1827_cast_fp16, var_1794_cast_fp16))[name = tensor("op_1985_cast_fp16")]; tensor var_1986_to_fp16 = const()[name = tensor("op_1986_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_191_cast_fp16 = mul(x = var_1985_cast_fp16, y = var_1986_to_fp16)[name = tensor("aw_chunk_191_cast_fp16")]; tensor var_1988_cast_fp16 = softmax(axis = var_1485, x = aw_chunk_129_cast_fp16)[name = tensor("op_1988_cast_fp16")]; tensor var_1989_cast_fp16 = softmax(axis = var_1485, x = aw_chunk_131_cast_fp16)[name = tensor("op_1989_cast_fp16")]; tensor var_1990_cast_fp16 = softmax(axis = var_1485, x = aw_chunk_133_cast_fp16)[name = tensor("op_1990_cast_fp16")]; tensor var_1991_cast_fp16 = softmax(axis = var_1485, x = aw_chunk_135_cast_fp16)[name = tensor("op_1991_cast_fp16")]; tensor var_1992_cast_fp16 = softmax(axis = var_1485, x = aw_chunk_137_cast_fp16)[name = tensor("op_1992_cast_fp16")]; tensor var_1993_cast_fp16 = softmax(axis = var_1485, x = aw_chunk_139_cast_fp16)[name = tensor("op_1993_cast_fp16")]; tensor var_1994_cast_fp16 = softmax(axis = var_1485, x = aw_chunk_141_cast_fp16)[name = tensor("op_1994_cast_fp16")]; tensor var_1995_cast_fp16 = softmax(axis = var_1485, x = aw_chunk_143_cast_fp16)[name = tensor("op_1995_cast_fp16")]; tensor var_1996_cast_fp16 = softmax(axis = var_1485, x = aw_chunk_145_cast_fp16)[name = tensor("op_1996_cast_fp16")]; tensor var_1997_cast_fp16 = softmax(axis = var_1485, x = aw_chunk_147_cast_fp16)[name = tensor("op_1997_cast_fp16")]; tensor var_1998_cast_fp16 = softmax(axis = var_1485, x = aw_chunk_149_cast_fp16)[name = tensor("op_1998_cast_fp16")]; tensor var_1999_cast_fp16 = softmax(axis = var_1485, x = aw_chunk_151_cast_fp16)[name = tensor("op_1999_cast_fp16")]; tensor var_2000_cast_fp16 = softmax(axis = var_1485, x = aw_chunk_153_cast_fp16)[name = tensor("op_2000_cast_fp16")]; tensor var_2001_cast_fp16 = softmax(axis = var_1485, x = aw_chunk_155_cast_fp16)[name = tensor("op_2001_cast_fp16")]; tensor var_2002_cast_fp16 = softmax(axis = var_1485, x = aw_chunk_157_cast_fp16)[name = tensor("op_2002_cast_fp16")]; tensor var_2003_cast_fp16 = softmax(axis = var_1485, x = aw_chunk_159_cast_fp16)[name = tensor("op_2003_cast_fp16")]; tensor var_2004_cast_fp16 = softmax(axis = var_1485, x = aw_chunk_161_cast_fp16)[name = tensor("op_2004_cast_fp16")]; tensor var_2005_cast_fp16 = softmax(axis = var_1485, x = aw_chunk_163_cast_fp16)[name = tensor("op_2005_cast_fp16")]; tensor var_2006_cast_fp16 = softmax(axis = var_1485, x = aw_chunk_165_cast_fp16)[name = tensor("op_2006_cast_fp16")]; tensor var_2007_cast_fp16 = softmax(axis = var_1485, x = aw_chunk_167_cast_fp16)[name = tensor("op_2007_cast_fp16")]; tensor var_2008_cast_fp16 = softmax(axis = var_1485, x = aw_chunk_169_cast_fp16)[name = tensor("op_2008_cast_fp16")]; tensor var_2009_cast_fp16 = softmax(axis = var_1485, x = aw_chunk_171_cast_fp16)[name = tensor("op_2009_cast_fp16")]; tensor var_2010_cast_fp16 = softmax(axis = var_1485, x = aw_chunk_173_cast_fp16)[name = tensor("op_2010_cast_fp16")]; tensor var_2011_cast_fp16 = softmax(axis = var_1485, x = aw_chunk_175_cast_fp16)[name = tensor("op_2011_cast_fp16")]; tensor var_2012_cast_fp16 = softmax(axis = var_1485, x = aw_chunk_177_cast_fp16)[name = tensor("op_2012_cast_fp16")]; tensor var_2013_cast_fp16 = softmax(axis = var_1485, x = aw_chunk_179_cast_fp16)[name = tensor("op_2013_cast_fp16")]; tensor var_2014_cast_fp16 = softmax(axis = var_1485, x = aw_chunk_181_cast_fp16)[name = tensor("op_2014_cast_fp16")]; tensor var_2015_cast_fp16 = softmax(axis = var_1485, x = aw_chunk_183_cast_fp16)[name = tensor("op_2015_cast_fp16")]; tensor var_2016_cast_fp16 = softmax(axis = var_1485, x = aw_chunk_185_cast_fp16)[name = tensor("op_2016_cast_fp16")]; tensor var_2017_cast_fp16 = softmax(axis = var_1485, x = aw_chunk_187_cast_fp16)[name = tensor("op_2017_cast_fp16")]; tensor var_2018_cast_fp16 = softmax(axis = var_1485, x = aw_chunk_189_cast_fp16)[name = tensor("op_2018_cast_fp16")]; tensor var_2019_cast_fp16 = softmax(axis = var_1485, x = aw_chunk_191_cast_fp16)[name = tensor("op_2019_cast_fp16")]; tensor var_2021_equation_0 = const()[name = tensor("op_2021_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_2021_cast_fp16 = einsum(equation = var_2021_equation_0, values = (var_1829_cast_fp16, var_1988_cast_fp16))[name = tensor("op_2021_cast_fp16")]; tensor var_2023_equation_0 = const()[name = tensor("op_2023_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_2023_cast_fp16 = einsum(equation = var_2023_equation_0, values = (var_1829_cast_fp16, var_1989_cast_fp16))[name = tensor("op_2023_cast_fp16")]; tensor var_2025_equation_0 = const()[name = tensor("op_2025_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_2025_cast_fp16 = einsum(equation = var_2025_equation_0, values = (var_1829_cast_fp16, var_1990_cast_fp16))[name = tensor("op_2025_cast_fp16")]; tensor var_2027_equation_0 = const()[name = tensor("op_2027_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_2027_cast_fp16 = einsum(equation = var_2027_equation_0, values = (var_1829_cast_fp16, var_1991_cast_fp16))[name = tensor("op_2027_cast_fp16")]; tensor var_2029_equation_0 = const()[name = tensor("op_2029_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_2029_cast_fp16 = einsum(equation = var_2029_equation_0, values = (var_1833_cast_fp16, var_1992_cast_fp16))[name = tensor("op_2029_cast_fp16")]; tensor var_2031_equation_0 = const()[name = tensor("op_2031_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_2031_cast_fp16 = einsum(equation = var_2031_equation_0, values = (var_1833_cast_fp16, var_1993_cast_fp16))[name = tensor("op_2031_cast_fp16")]; tensor var_2033_equation_0 = const()[name = tensor("op_2033_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_2033_cast_fp16 = einsum(equation = var_2033_equation_0, values = (var_1833_cast_fp16, var_1994_cast_fp16))[name = tensor("op_2033_cast_fp16")]; tensor var_2035_equation_0 = const()[name = tensor("op_2035_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_2035_cast_fp16 = einsum(equation = var_2035_equation_0, values = (var_1833_cast_fp16, var_1995_cast_fp16))[name = tensor("op_2035_cast_fp16")]; tensor var_2037_equation_0 = const()[name = tensor("op_2037_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_2037_cast_fp16 = einsum(equation = var_2037_equation_0, values = (var_1837_cast_fp16, var_1996_cast_fp16))[name = tensor("op_2037_cast_fp16")]; tensor var_2039_equation_0 = const()[name = tensor("op_2039_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_2039_cast_fp16 = einsum(equation = var_2039_equation_0, values = (var_1837_cast_fp16, var_1997_cast_fp16))[name = tensor("op_2039_cast_fp16")]; tensor var_2041_equation_0 = const()[name = tensor("op_2041_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_2041_cast_fp16 = einsum(equation = var_2041_equation_0, values = (var_1837_cast_fp16, var_1998_cast_fp16))[name = tensor("op_2041_cast_fp16")]; tensor var_2043_equation_0 = const()[name = tensor("op_2043_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_2043_cast_fp16 = einsum(equation = var_2043_equation_0, values = (var_1837_cast_fp16, var_1999_cast_fp16))[name = tensor("op_2043_cast_fp16")]; tensor var_2045_equation_0 = const()[name = tensor("op_2045_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_2045_cast_fp16 = einsum(equation = var_2045_equation_0, values = (var_1841_cast_fp16, var_2000_cast_fp16))[name = tensor("op_2045_cast_fp16")]; tensor var_2047_equation_0 = const()[name = tensor("op_2047_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_2047_cast_fp16 = einsum(equation = var_2047_equation_0, values = (var_1841_cast_fp16, var_2001_cast_fp16))[name = tensor("op_2047_cast_fp16")]; tensor var_2049_equation_0 = const()[name = tensor("op_2049_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_2049_cast_fp16 = einsum(equation = var_2049_equation_0, values = (var_1841_cast_fp16, var_2002_cast_fp16))[name = tensor("op_2049_cast_fp16")]; tensor var_2051_equation_0 = const()[name = tensor("op_2051_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_2051_cast_fp16 = einsum(equation = var_2051_equation_0, values = (var_1841_cast_fp16, var_2003_cast_fp16))[name = tensor("op_2051_cast_fp16")]; tensor var_2053_equation_0 = const()[name = tensor("op_2053_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_2053_cast_fp16 = einsum(equation = var_2053_equation_0, values = (var_1845_cast_fp16, var_2004_cast_fp16))[name = tensor("op_2053_cast_fp16")]; tensor var_2055_equation_0 = const()[name = tensor("op_2055_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_2055_cast_fp16 = einsum(equation = var_2055_equation_0, values = (var_1845_cast_fp16, var_2005_cast_fp16))[name = tensor("op_2055_cast_fp16")]; tensor var_2057_equation_0 = const()[name = tensor("op_2057_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_2057_cast_fp16 = einsum(equation = var_2057_equation_0, values = (var_1845_cast_fp16, var_2006_cast_fp16))[name = tensor("op_2057_cast_fp16")]; tensor var_2059_equation_0 = const()[name = tensor("op_2059_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_2059_cast_fp16 = einsum(equation = var_2059_equation_0, values = (var_1845_cast_fp16, var_2007_cast_fp16))[name = tensor("op_2059_cast_fp16")]; tensor var_2061_equation_0 = const()[name = tensor("op_2061_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_2061_cast_fp16 = einsum(equation = var_2061_equation_0, values = (var_1849_cast_fp16, var_2008_cast_fp16))[name = tensor("op_2061_cast_fp16")]; tensor var_2063_equation_0 = const()[name = tensor("op_2063_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_2063_cast_fp16 = einsum(equation = var_2063_equation_0, values = (var_1849_cast_fp16, var_2009_cast_fp16))[name = tensor("op_2063_cast_fp16")]; tensor var_2065_equation_0 = const()[name = tensor("op_2065_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_2065_cast_fp16 = einsum(equation = var_2065_equation_0, values = (var_1849_cast_fp16, var_2010_cast_fp16))[name = tensor("op_2065_cast_fp16")]; tensor var_2067_equation_0 = const()[name = tensor("op_2067_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_2067_cast_fp16 = einsum(equation = var_2067_equation_0, values = (var_1849_cast_fp16, var_2011_cast_fp16))[name = tensor("op_2067_cast_fp16")]; tensor var_2069_equation_0 = const()[name = tensor("op_2069_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_2069_cast_fp16 = einsum(equation = var_2069_equation_0, values = (var_1853_cast_fp16, var_2012_cast_fp16))[name = tensor("op_2069_cast_fp16")]; tensor var_2071_equation_0 = const()[name = tensor("op_2071_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_2071_cast_fp16 = einsum(equation = var_2071_equation_0, values = (var_1853_cast_fp16, var_2013_cast_fp16))[name = tensor("op_2071_cast_fp16")]; tensor var_2073_equation_0 = const()[name = tensor("op_2073_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_2073_cast_fp16 = einsum(equation = var_2073_equation_0, values = (var_1853_cast_fp16, var_2014_cast_fp16))[name = tensor("op_2073_cast_fp16")]; tensor var_2075_equation_0 = const()[name = tensor("op_2075_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_2075_cast_fp16 = einsum(equation = var_2075_equation_0, values = (var_1853_cast_fp16, var_2015_cast_fp16))[name = tensor("op_2075_cast_fp16")]; tensor var_2077_equation_0 = const()[name = tensor("op_2077_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_2077_cast_fp16 = einsum(equation = var_2077_equation_0, values = (var_1857_cast_fp16, var_2016_cast_fp16))[name = tensor("op_2077_cast_fp16")]; tensor var_2079_equation_0 = const()[name = tensor("op_2079_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_2079_cast_fp16 = einsum(equation = var_2079_equation_0, values = (var_1857_cast_fp16, var_2017_cast_fp16))[name = tensor("op_2079_cast_fp16")]; tensor var_2081_equation_0 = const()[name = tensor("op_2081_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_2081_cast_fp16 = einsum(equation = var_2081_equation_0, values = (var_1857_cast_fp16, var_2018_cast_fp16))[name = tensor("op_2081_cast_fp16")]; tensor var_2083_equation_0 = const()[name = tensor("op_2083_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_2083_cast_fp16 = einsum(equation = var_2083_equation_0, values = (var_1857_cast_fp16, var_2019_cast_fp16))[name = tensor("op_2083_cast_fp16")]; tensor var_2085_interleave_0 = const()[name = tensor("op_2085_interleave_0"), val = tensor(false)]; tensor var_2085_cast_fp16 = concat(axis = var_1472, interleave = var_2085_interleave_0, values = (var_2021_cast_fp16, var_2023_cast_fp16, var_2025_cast_fp16, var_2027_cast_fp16))[name = tensor("op_2085_cast_fp16")]; tensor var_2087_interleave_0 = const()[name = tensor("op_2087_interleave_0"), val = tensor(false)]; tensor var_2087_cast_fp16 = concat(axis = var_1472, interleave = var_2087_interleave_0, values = (var_2029_cast_fp16, var_2031_cast_fp16, var_2033_cast_fp16, var_2035_cast_fp16))[name = tensor("op_2087_cast_fp16")]; tensor var_2089_interleave_0 = const()[name = tensor("op_2089_interleave_0"), val = tensor(false)]; tensor var_2089_cast_fp16 = concat(axis = var_1472, interleave = var_2089_interleave_0, values = (var_2037_cast_fp16, var_2039_cast_fp16, var_2041_cast_fp16, var_2043_cast_fp16))[name = tensor("op_2089_cast_fp16")]; tensor var_2091_interleave_0 = const()[name = tensor("op_2091_interleave_0"), val = tensor(false)]; tensor var_2091_cast_fp16 = concat(axis = var_1472, interleave = var_2091_interleave_0, values = (var_2045_cast_fp16, var_2047_cast_fp16, var_2049_cast_fp16, var_2051_cast_fp16))[name = tensor("op_2091_cast_fp16")]; tensor var_2093_interleave_0 = const()[name = tensor("op_2093_interleave_0"), val = tensor(false)]; tensor var_2093_cast_fp16 = concat(axis = var_1472, interleave = var_2093_interleave_0, values = (var_2053_cast_fp16, var_2055_cast_fp16, var_2057_cast_fp16, var_2059_cast_fp16))[name = tensor("op_2093_cast_fp16")]; tensor var_2095_interleave_0 = const()[name = tensor("op_2095_interleave_0"), val = tensor(false)]; tensor var_2095_cast_fp16 = concat(axis = var_1472, interleave = var_2095_interleave_0, values = (var_2061_cast_fp16, var_2063_cast_fp16, var_2065_cast_fp16, var_2067_cast_fp16))[name = tensor("op_2095_cast_fp16")]; tensor var_2097_interleave_0 = const()[name = tensor("op_2097_interleave_0"), val = tensor(false)]; tensor var_2097_cast_fp16 = concat(axis = var_1472, interleave = var_2097_interleave_0, values = (var_2069_cast_fp16, var_2071_cast_fp16, var_2073_cast_fp16, var_2075_cast_fp16))[name = tensor("op_2097_cast_fp16")]; tensor var_2099_interleave_0 = const()[name = tensor("op_2099_interleave_0"), val = tensor(false)]; tensor var_2099_cast_fp16 = concat(axis = var_1472, interleave = var_2099_interleave_0, values = (var_2077_cast_fp16, var_2079_cast_fp16, var_2081_cast_fp16, var_2083_cast_fp16))[name = tensor("op_2099_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_1485, interleave = input_17_interleave_0, values = (var_2085_cast_fp16, var_2087_cast_fp16, var_2089_cast_fp16, var_2091_cast_fp16, var_2093_cast_fp16, var_2095_cast_fp16, var_2097_cast_fp16, var_2099_cast_fp16))[name = tensor("input_17_cast_fp16")]; tensor var_2104 = const()[name = tensor("op_2104"), val = tensor([1, 1])]; tensor var_2106 = const()[name = tensor("op_2106"), 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(17546048)))]; 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(18070400)))]; tensor obj_11_cast_fp16 = conv(bias = layers_2_self_attn_o_proj_bias_to_fp16, dilations = var_2106, groups = var_1485, pad = obj_11_pad_0, pad_type = obj_11_pad_type_0, strides = var_2104, 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_2112 = const()[name = tensor("op_2112"), val = tensor([1])]; tensor channels_mean_11_cast_fp16 = reduce_mean(axes = var_2112, keep_dims = var_1486, 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_2116 = const()[name = tensor("op_2116"), val = tensor([1])]; tensor var_2117_cast_fp16 = reduce_mean(axes = var_2116, keep_dims = var_1486, x = zero_mean_sq_11_cast_fp16)[name = tensor("op_2117_cast_fp16")]; tensor var_2118_to_fp16 = const()[name = tensor("op_2118_to_fp16"), val = tensor(0x1.5p-17)]; tensor var_2119_cast_fp16 = add(x = var_2117_cast_fp16, y = var_2118_to_fp16)[name = tensor("op_2119_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_2119_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(18071488)))]; 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(18072576)))]; 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_2130 = const()[name = tensor("op_2130"), val = tensor([1, 1])]; tensor var_2132 = const()[name = tensor("op_2132"), 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(18073664)))]; 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(20170880)))]; tensor input_21_cast_fp16 = conv(bias = layers_2_fc1_bias_to_fp16, dilations = var_2132, groups = var_1485, pad = input_21_pad_0, pad_type = input_21_pad_type_0, strides = var_2130, 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_2138 = const()[name = tensor("op_2138"), val = tensor([1, 1])]; tensor var_2140 = const()[name = tensor("op_2140"), 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(20175040)))]; 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(22272256)))]; tensor hidden_states_9_cast_fp16 = conv(bias = layers_2_fc2_bias_to_fp16, dilations = var_2140, groups = var_1485, pad = hidden_states_9_pad_0, pad_type = hidden_states_9_pad_type_0, strides = var_2138, 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_2147 = const()[name = tensor("op_2147"), val = tensor(3)]; tensor var_2160 = const()[name = tensor("op_2160"), val = tensor(1)]; tensor var_2161 = const()[name = tensor("op_2161"), val = tensor(true)]; tensor var_2171 = const()[name = tensor("op_2171"), val = tensor([1])]; tensor channels_mean_13_cast_fp16 = reduce_mean(axes = var_2171, keep_dims = var_2161, 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_2175 = const()[name = tensor("op_2175"), val = tensor([1])]; tensor var_2176_cast_fp16 = reduce_mean(axes = var_2175, keep_dims = var_2161, x = zero_mean_sq_13_cast_fp16)[name = tensor("op_2176_cast_fp16")]; tensor var_2177_to_fp16 = const()[name = tensor("op_2177_to_fp16"), val = tensor(0x1.5p-17)]; tensor var_2178_cast_fp16 = add(x = var_2176_cast_fp16, y = var_2177_to_fp16)[name = tensor("op_2178_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_2178_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(22273344)))]; 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(22274432)))]; 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_2193 = const()[name = tensor("op_2193"), val = tensor([1, 1])]; tensor var_2195 = const()[name = tensor("op_2195"), val = tensor([1, 1])]; tensor query_7_pad_type_0 = const()[name = tensor("query_7_pad_type_0"), val = tensor("custom")]; tensor query_7_pad_0 = const()[name = tensor("query_7_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(22275520)))]; 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(22799872)))]; tensor query_7_cast_fp16 = conv(bias = layers_3_self_attn_q_proj_bias_to_fp16, dilations = var_2195, groups = var_2160, pad = query_7_pad_0, pad_type = query_7_pad_type_0, strides = var_2193, weight = layers_3_self_attn_q_proj_weight_to_fp16, x = obj_13_cast_fp16)[name = tensor("query_7_cast_fp16")]; tensor var_2199 = const()[name = tensor("op_2199"), val = tensor([1, 1])]; tensor var_2201 = const()[name = tensor("op_2201"), val = tensor([1, 1])]; tensor key_7_pad_type_0 = const()[name = tensor("key_7_pad_type_0"), val = tensor("custom")]; tensor key_7_pad_0 = const()[name = tensor("key_7_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(22800960)))]; tensor key_7_cast_fp16 = conv(dilations = var_2201, groups = var_2160, pad = key_7_pad_0, pad_type = key_7_pad_type_0, strides = var_2199, weight = layers_3_self_attn_k_proj_weight_to_fp16, x = obj_13_cast_fp16)[name = tensor("key_7_cast_fp16")]; tensor var_2206 = const()[name = tensor("op_2206"), val = tensor([1, 1])]; tensor var_2208 = const()[name = tensor("op_2208"), val = tensor([1, 1])]; tensor value_7_pad_type_0 = const()[name = tensor("value_7_pad_type_0"), val = tensor("custom")]; tensor value_7_pad_0 = const()[name = tensor("value_7_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(23325312)))]; 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(23849664)))]; tensor value_7_cast_fp16 = conv(bias = layers_3_self_attn_v_proj_bias_to_fp16, dilations = var_2208, groups = var_2160, pad = value_7_pad_0, pad_type = value_7_pad_type_0, strides = var_2206, weight = layers_3_self_attn_v_proj_weight_to_fp16, x = obj_13_cast_fp16)[name = tensor("value_7_cast_fp16")]; tensor var_2215_begin_0 = const()[name = tensor("op_2215_begin_0"), val = tensor([0, 0, 0, 0])]; tensor var_2215_end_0 = const()[name = tensor("op_2215_end_0"), val = tensor([1, 64, 1, 1500])]; tensor var_2215_end_mask_0 = const()[name = tensor("op_2215_end_mask_0"), val = tensor([true, false, true, true])]; tensor var_2215_cast_fp16 = slice_by_index(begin = var_2215_begin_0, end = var_2215_end_0, end_mask = var_2215_end_mask_0, x = query_7_cast_fp16)[name = tensor("op_2215_cast_fp16")]; tensor var_2219_begin_0 = const()[name = tensor("op_2219_begin_0"), val = tensor([0, 64, 0, 0])]; tensor var_2219_end_0 = const()[name = tensor("op_2219_end_0"), val = tensor([1, 128, 1, 1500])]; tensor var_2219_end_mask_0 = const()[name = tensor("op_2219_end_mask_0"), val = tensor([true, false, true, true])]; tensor var_2219_cast_fp16 = slice_by_index(begin = var_2219_begin_0, end = var_2219_end_0, end_mask = var_2219_end_mask_0, x = query_7_cast_fp16)[name = tensor("op_2219_cast_fp16")]; tensor var_2223_begin_0 = const()[name = tensor("op_2223_begin_0"), val = tensor([0, 128, 0, 0])]; tensor var_2223_end_0 = const()[name = tensor("op_2223_end_0"), val = tensor([1, 192, 1, 1500])]; tensor var_2223_end_mask_0 = const()[name = tensor("op_2223_end_mask_0"), val = tensor([true, false, true, true])]; tensor var_2223_cast_fp16 = slice_by_index(begin = var_2223_begin_0, end = var_2223_end_0, end_mask = var_2223_end_mask_0, x = query_7_cast_fp16)[name = tensor("op_2223_cast_fp16")]; tensor var_2227_begin_0 = const()[name = tensor("op_2227_begin_0"), val = tensor([0, 192, 0, 0])]; tensor var_2227_end_0 = const()[name = tensor("op_2227_end_0"), val = tensor([1, 256, 1, 1500])]; tensor var_2227_end_mask_0 = const()[name = tensor("op_2227_end_mask_0"), val = tensor([true, false, true, true])]; tensor var_2227_cast_fp16 = slice_by_index(begin = var_2227_begin_0, end = var_2227_end_0, end_mask = var_2227_end_mask_0, x = query_7_cast_fp16)[name = tensor("op_2227_cast_fp16")]; tensor var_2231_begin_0 = const()[name = tensor("op_2231_begin_0"), val = tensor([0, 256, 0, 0])]; tensor var_2231_end_0 = const()[name = tensor("op_2231_end_0"), val = tensor([1, 320, 1, 1500])]; tensor var_2231_end_mask_0 = const()[name = tensor("op_2231_end_mask_0"), val = tensor([true, false, true, true])]; tensor var_2231_cast_fp16 = slice_by_index(begin = var_2231_begin_0, end = var_2231_end_0, end_mask = var_2231_end_mask_0, x = query_7_cast_fp16)[name = tensor("op_2231_cast_fp16")]; tensor var_2235_begin_0 = const()[name = tensor("op_2235_begin_0"), val = tensor([0, 320, 0, 0])]; tensor var_2235_end_0 = const()[name = tensor("op_2235_end_0"), val = tensor([1, 384, 1, 1500])]; tensor var_2235_end_mask_0 = const()[name = tensor("op_2235_end_mask_0"), val = tensor([true, false, true, true])]; tensor var_2235_cast_fp16 = slice_by_index(begin = var_2235_begin_0, end = var_2235_end_0, end_mask = var_2235_end_mask_0, x = query_7_cast_fp16)[name = tensor("op_2235_cast_fp16")]; tensor var_2239_begin_0 = const()[name = tensor("op_2239_begin_0"), val = tensor([0, 384, 0, 0])]; tensor var_2239_end_0 = const()[name = tensor("op_2239_end_0"), val = tensor([1, 448, 1, 1500])]; tensor var_2239_end_mask_0 = const()[name = tensor("op_2239_end_mask_0"), val = tensor([true, false, true, true])]; tensor var_2239_cast_fp16 = slice_by_index(begin = var_2239_begin_0, end = var_2239_end_0, end_mask = var_2239_end_mask_0, x = query_7_cast_fp16)[name = tensor("op_2239_cast_fp16")]; tensor var_2243_begin_0 = const()[name = tensor("op_2243_begin_0"), val = tensor([0, 448, 0, 0])]; tensor var_2243_end_0 = const()[name = tensor("op_2243_end_0"), val = tensor([1, 512, 1, 1500])]; tensor var_2243_end_mask_0 = const()[name = tensor("op_2243_end_mask_0"), val = tensor([true, false, true, true])]; tensor var_2243_cast_fp16 = slice_by_index(begin = var_2243_begin_0, end = var_2243_end_0, end_mask = var_2243_end_mask_0, x = query_7_cast_fp16)[name = tensor("op_2243_cast_fp16")]; tensor var_2252_begin_0 = const()[name = tensor("op_2252_begin_0"), val = tensor([0, 0, 0, 0])]; tensor var_2252_end_0 = const()[name = tensor("op_2252_end_0"), val = tensor([1, 64, 1, 375])]; tensor var_2252_end_mask_0 = const()[name = tensor("op_2252_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_2252_cast_fp16 = slice_by_index(begin = var_2252_begin_0, end = var_2252_end_0, end_mask = var_2252_end_mask_0, x = var_2215_cast_fp16)[name = tensor("op_2252_cast_fp16")]; tensor var_2259_begin_0 = const()[name = tensor("op_2259_begin_0"), val = tensor([0, 0, 0, 375])]; tensor var_2259_end_0 = const()[name = tensor("op_2259_end_0"), val = tensor([1, 64, 1, 750])]; tensor var_2259_end_mask_0 = const()[name = tensor("op_2259_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_2259_cast_fp16 = slice_by_index(begin = var_2259_begin_0, end = var_2259_end_0, end_mask = var_2259_end_mask_0, x = var_2215_cast_fp16)[name = tensor("op_2259_cast_fp16")]; tensor var_2266_begin_0 = const()[name = tensor("op_2266_begin_0"), val = tensor([0, 0, 0, 750])]; tensor var_2266_end_0 = const()[name = tensor("op_2266_end_0"), val = tensor([1, 64, 1, 1125])]; tensor var_2266_end_mask_0 = const()[name = tensor("op_2266_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_2266_cast_fp16 = slice_by_index(begin = var_2266_begin_0, end = var_2266_end_0, end_mask = var_2266_end_mask_0, x = var_2215_cast_fp16)[name = tensor("op_2266_cast_fp16")]; tensor var_2273_begin_0 = const()[name = tensor("op_2273_begin_0"), val = tensor([0, 0, 0, 1125])]; tensor var_2273_end_0 = const()[name = tensor("op_2273_end_0"), val = tensor([1, 64, 1, 1500])]; tensor var_2273_end_mask_0 = const()[name = tensor("op_2273_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_2273_cast_fp16 = slice_by_index(begin = var_2273_begin_0, end = var_2273_end_0, end_mask = var_2273_end_mask_0, x = var_2215_cast_fp16)[name = tensor("op_2273_cast_fp16")]; tensor var_2280_begin_0 = const()[name = tensor("op_2280_begin_0"), val = tensor([0, 0, 0, 0])]; tensor var_2280_end_0 = const()[name = tensor("op_2280_end_0"), val = tensor([1, 64, 1, 375])]; tensor var_2280_end_mask_0 = const()[name = tensor("op_2280_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_2280_cast_fp16 = slice_by_index(begin = var_2280_begin_0, end = var_2280_end_0, end_mask = var_2280_end_mask_0, x = var_2219_cast_fp16)[name = tensor("op_2280_cast_fp16")]; tensor var_2287_begin_0 = const()[name = tensor("op_2287_begin_0"), val = tensor([0, 0, 0, 375])]; tensor var_2287_end_0 = const()[name = tensor("op_2287_end_0"), val = tensor([1, 64, 1, 750])]; tensor var_2287_end_mask_0 = const()[name = tensor("op_2287_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_2287_cast_fp16 = slice_by_index(begin = var_2287_begin_0, end = var_2287_end_0, end_mask = var_2287_end_mask_0, x = var_2219_cast_fp16)[name = tensor("op_2287_cast_fp16")]; tensor var_2294_begin_0 = const()[name = tensor("op_2294_begin_0"), val = tensor([0, 0, 0, 750])]; tensor var_2294_end_0 = const()[name = tensor("op_2294_end_0"), val = tensor([1, 64, 1, 1125])]; tensor var_2294_end_mask_0 = const()[name = tensor("op_2294_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_2294_cast_fp16 = slice_by_index(begin = var_2294_begin_0, end = var_2294_end_0, end_mask = var_2294_end_mask_0, x = var_2219_cast_fp16)[name = tensor("op_2294_cast_fp16")]; tensor var_2301_begin_0 = const()[name = tensor("op_2301_begin_0"), val = tensor([0, 0, 0, 1125])]; tensor var_2301_end_0 = const()[name = tensor("op_2301_end_0"), val = tensor([1, 64, 1, 1500])]; tensor var_2301_end_mask_0 = const()[name = tensor("op_2301_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_2301_cast_fp16 = slice_by_index(begin = var_2301_begin_0, end = var_2301_end_0, end_mask = var_2301_end_mask_0, x = var_2219_cast_fp16)[name = tensor("op_2301_cast_fp16")]; tensor var_2308_begin_0 = const()[name = tensor("op_2308_begin_0"), val = tensor([0, 0, 0, 0])]; tensor var_2308_end_0 = const()[name = tensor("op_2308_end_0"), val = tensor([1, 64, 1, 375])]; tensor var_2308_end_mask_0 = const()[name = tensor("op_2308_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_2308_cast_fp16 = slice_by_index(begin = var_2308_begin_0, end = var_2308_end_0, end_mask = var_2308_end_mask_0, x = var_2223_cast_fp16)[name = tensor("op_2308_cast_fp16")]; tensor var_2315_begin_0 = const()[name = tensor("op_2315_begin_0"), val = tensor([0, 0, 0, 375])]; tensor var_2315_end_0 = const()[name = tensor("op_2315_end_0"), val = tensor([1, 64, 1, 750])]; tensor var_2315_end_mask_0 = const()[name = tensor("op_2315_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_2315_cast_fp16 = slice_by_index(begin = var_2315_begin_0, end = var_2315_end_0, end_mask = var_2315_end_mask_0, x = var_2223_cast_fp16)[name = tensor("op_2315_cast_fp16")]; tensor var_2322_begin_0 = const()[name = tensor("op_2322_begin_0"), val = tensor([0, 0, 0, 750])]; tensor var_2322_end_0 = const()[name = tensor("op_2322_end_0"), val = tensor([1, 64, 1, 1125])]; tensor var_2322_end_mask_0 = const()[name = tensor("op_2322_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_2322_cast_fp16 = slice_by_index(begin = var_2322_begin_0, end = var_2322_end_0, end_mask = var_2322_end_mask_0, x = var_2223_cast_fp16)[name = tensor("op_2322_cast_fp16")]; tensor var_2329_begin_0 = const()[name = tensor("op_2329_begin_0"), val = tensor([0, 0, 0, 1125])]; tensor var_2329_end_0 = const()[name = tensor("op_2329_end_0"), val = tensor([1, 64, 1, 1500])]; tensor var_2329_end_mask_0 = const()[name = tensor("op_2329_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_2329_cast_fp16 = slice_by_index(begin = var_2329_begin_0, end = var_2329_end_0, end_mask = var_2329_end_mask_0, x = var_2223_cast_fp16)[name = tensor("op_2329_cast_fp16")]; tensor var_2336_begin_0 = const()[name = tensor("op_2336_begin_0"), val = tensor([0, 0, 0, 0])]; tensor var_2336_end_0 = const()[name = tensor("op_2336_end_0"), val = tensor([1, 64, 1, 375])]; tensor var_2336_end_mask_0 = const()[name = tensor("op_2336_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_2336_cast_fp16 = slice_by_index(begin = var_2336_begin_0, end = var_2336_end_0, end_mask = var_2336_end_mask_0, x = var_2227_cast_fp16)[name = tensor("op_2336_cast_fp16")]; tensor var_2343_begin_0 = const()[name = tensor("op_2343_begin_0"), val = tensor([0, 0, 0, 375])]; tensor var_2343_end_0 = const()[name = tensor("op_2343_end_0"), val = tensor([1, 64, 1, 750])]; tensor var_2343_end_mask_0 = const()[name = tensor("op_2343_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_2343_cast_fp16 = slice_by_index(begin = var_2343_begin_0, end = var_2343_end_0, end_mask = var_2343_end_mask_0, x = var_2227_cast_fp16)[name = tensor("op_2343_cast_fp16")]; tensor var_2350_begin_0 = const()[name = tensor("op_2350_begin_0"), val = tensor([0, 0, 0, 750])]; tensor var_2350_end_0 = const()[name = tensor("op_2350_end_0"), val = tensor([1, 64, 1, 1125])]; tensor var_2350_end_mask_0 = const()[name = tensor("op_2350_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_2350_cast_fp16 = slice_by_index(begin = var_2350_begin_0, end = var_2350_end_0, end_mask = var_2350_end_mask_0, x = var_2227_cast_fp16)[name = tensor("op_2350_cast_fp16")]; tensor var_2357_begin_0 = const()[name = tensor("op_2357_begin_0"), val = tensor([0, 0, 0, 1125])]; tensor var_2357_end_0 = const()[name = tensor("op_2357_end_0"), val = tensor([1, 64, 1, 1500])]; tensor var_2357_end_mask_0 = const()[name = tensor("op_2357_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_2357_cast_fp16 = slice_by_index(begin = var_2357_begin_0, end = var_2357_end_0, end_mask = var_2357_end_mask_0, x = var_2227_cast_fp16)[name = tensor("op_2357_cast_fp16")]; tensor var_2364_begin_0 = const()[name = tensor("op_2364_begin_0"), val = tensor([0, 0, 0, 0])]; tensor var_2364_end_0 = const()[name = tensor("op_2364_end_0"), val = tensor([1, 64, 1, 375])]; tensor var_2364_end_mask_0 = const()[name = tensor("op_2364_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_2364_cast_fp16 = slice_by_index(begin = var_2364_begin_0, end = var_2364_end_0, end_mask = var_2364_end_mask_0, x = var_2231_cast_fp16)[name = tensor("op_2364_cast_fp16")]; tensor var_2371_begin_0 = const()[name = tensor("op_2371_begin_0"), val = tensor([0, 0, 0, 375])]; tensor var_2371_end_0 = const()[name = tensor("op_2371_end_0"), val = tensor([1, 64, 1, 750])]; tensor var_2371_end_mask_0 = const()[name = tensor("op_2371_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_2371_cast_fp16 = slice_by_index(begin = var_2371_begin_0, end = var_2371_end_0, end_mask = var_2371_end_mask_0, x = var_2231_cast_fp16)[name = tensor("op_2371_cast_fp16")]; tensor var_2378_begin_0 = const()[name = tensor("op_2378_begin_0"), val = tensor([0, 0, 0, 750])]; tensor var_2378_end_0 = const()[name = tensor("op_2378_end_0"), val = tensor([1, 64, 1, 1125])]; tensor var_2378_end_mask_0 = const()[name = tensor("op_2378_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_2378_cast_fp16 = slice_by_index(begin = var_2378_begin_0, end = var_2378_end_0, end_mask = var_2378_end_mask_0, x = var_2231_cast_fp16)[name = tensor("op_2378_cast_fp16")]; tensor var_2385_begin_0 = const()[name = tensor("op_2385_begin_0"), val = tensor([0, 0, 0, 1125])]; tensor var_2385_end_0 = const()[name = tensor("op_2385_end_0"), val = tensor([1, 64, 1, 1500])]; tensor var_2385_end_mask_0 = const()[name = tensor("op_2385_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_2385_cast_fp16 = slice_by_index(begin = var_2385_begin_0, end = var_2385_end_0, end_mask = var_2385_end_mask_0, x = var_2231_cast_fp16)[name = tensor("op_2385_cast_fp16")]; tensor var_2392_begin_0 = const()[name = tensor("op_2392_begin_0"), val = tensor([0, 0, 0, 0])]; tensor var_2392_end_0 = const()[name = tensor("op_2392_end_0"), val = tensor([1, 64, 1, 375])]; tensor var_2392_end_mask_0 = const()[name = tensor("op_2392_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_2392_cast_fp16 = slice_by_index(begin = var_2392_begin_0, end = var_2392_end_0, end_mask = var_2392_end_mask_0, x = var_2235_cast_fp16)[name = tensor("op_2392_cast_fp16")]; tensor var_2399_begin_0 = const()[name = tensor("op_2399_begin_0"), val = tensor([0, 0, 0, 375])]; tensor var_2399_end_0 = const()[name = tensor("op_2399_end_0"), val = tensor([1, 64, 1, 750])]; tensor var_2399_end_mask_0 = const()[name = tensor("op_2399_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_2399_cast_fp16 = slice_by_index(begin = var_2399_begin_0, end = var_2399_end_0, end_mask = var_2399_end_mask_0, x = var_2235_cast_fp16)[name = tensor("op_2399_cast_fp16")]; tensor var_2406_begin_0 = const()[name = tensor("op_2406_begin_0"), val = tensor([0, 0, 0, 750])]; tensor var_2406_end_0 = const()[name = tensor("op_2406_end_0"), val = tensor([1, 64, 1, 1125])]; tensor var_2406_end_mask_0 = const()[name = tensor("op_2406_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_2406_cast_fp16 = slice_by_index(begin = var_2406_begin_0, end = var_2406_end_0, end_mask = var_2406_end_mask_0, x = var_2235_cast_fp16)[name = tensor("op_2406_cast_fp16")]; tensor var_2413_begin_0 = const()[name = tensor("op_2413_begin_0"), val = tensor([0, 0, 0, 1125])]; tensor var_2413_end_0 = const()[name = tensor("op_2413_end_0"), val = tensor([1, 64, 1, 1500])]; tensor var_2413_end_mask_0 = const()[name = tensor("op_2413_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_2413_cast_fp16 = slice_by_index(begin = var_2413_begin_0, end = var_2413_end_0, end_mask = var_2413_end_mask_0, x = var_2235_cast_fp16)[name = tensor("op_2413_cast_fp16")]; tensor var_2420_begin_0 = const()[name = tensor("op_2420_begin_0"), val = tensor([0, 0, 0, 0])]; tensor var_2420_end_0 = const()[name = tensor("op_2420_end_0"), val = tensor([1, 64, 1, 375])]; tensor var_2420_end_mask_0 = const()[name = tensor("op_2420_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_2420_cast_fp16 = slice_by_index(begin = var_2420_begin_0, end = var_2420_end_0, end_mask = var_2420_end_mask_0, x = var_2239_cast_fp16)[name = tensor("op_2420_cast_fp16")]; tensor var_2427_begin_0 = const()[name = tensor("op_2427_begin_0"), val = tensor([0, 0, 0, 375])]; tensor var_2427_end_0 = const()[name = tensor("op_2427_end_0"), val = tensor([1, 64, 1, 750])]; tensor var_2427_end_mask_0 = const()[name = tensor("op_2427_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_2427_cast_fp16 = slice_by_index(begin = var_2427_begin_0, end = var_2427_end_0, end_mask = var_2427_end_mask_0, x = var_2239_cast_fp16)[name = tensor("op_2427_cast_fp16")]; tensor var_2434_begin_0 = const()[name = tensor("op_2434_begin_0"), val = tensor([0, 0, 0, 750])]; tensor var_2434_end_0 = const()[name = tensor("op_2434_end_0"), val = tensor([1, 64, 1, 1125])]; tensor var_2434_end_mask_0 = const()[name = tensor("op_2434_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_2434_cast_fp16 = slice_by_index(begin = var_2434_begin_0, end = var_2434_end_0, end_mask = var_2434_end_mask_0, x = var_2239_cast_fp16)[name = tensor("op_2434_cast_fp16")]; tensor var_2441_begin_0 = const()[name = tensor("op_2441_begin_0"), val = tensor([0, 0, 0, 1125])]; tensor var_2441_end_0 = const()[name = tensor("op_2441_end_0"), val = tensor([1, 64, 1, 1500])]; tensor var_2441_end_mask_0 = const()[name = tensor("op_2441_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_2441_cast_fp16 = slice_by_index(begin = var_2441_begin_0, end = var_2441_end_0, end_mask = var_2441_end_mask_0, x = var_2239_cast_fp16)[name = tensor("op_2441_cast_fp16")]; tensor var_2448_begin_0 = const()[name = tensor("op_2448_begin_0"), val = tensor([0, 0, 0, 0])]; tensor var_2448_end_0 = const()[name = tensor("op_2448_end_0"), val = tensor([1, 64, 1, 375])]; tensor var_2448_end_mask_0 = const()[name = tensor("op_2448_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_2448_cast_fp16 = slice_by_index(begin = var_2448_begin_0, end = var_2448_end_0, end_mask = var_2448_end_mask_0, x = var_2243_cast_fp16)[name = tensor("op_2448_cast_fp16")]; tensor var_2455_begin_0 = const()[name = tensor("op_2455_begin_0"), val = tensor([0, 0, 0, 375])]; tensor var_2455_end_0 = const()[name = tensor("op_2455_end_0"), val = tensor([1, 64, 1, 750])]; tensor var_2455_end_mask_0 = const()[name = tensor("op_2455_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_2455_cast_fp16 = slice_by_index(begin = var_2455_begin_0, end = var_2455_end_0, end_mask = var_2455_end_mask_0, x = var_2243_cast_fp16)[name = tensor("op_2455_cast_fp16")]; tensor var_2462_begin_0 = const()[name = tensor("op_2462_begin_0"), val = tensor([0, 0, 0, 750])]; tensor var_2462_end_0 = const()[name = tensor("op_2462_end_0"), val = tensor([1, 64, 1, 1125])]; tensor var_2462_end_mask_0 = const()[name = tensor("op_2462_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_2462_cast_fp16 = slice_by_index(begin = var_2462_begin_0, end = var_2462_end_0, end_mask = var_2462_end_mask_0, x = var_2243_cast_fp16)[name = tensor("op_2462_cast_fp16")]; tensor var_2469_begin_0 = const()[name = tensor("op_2469_begin_0"), val = tensor([0, 0, 0, 1125])]; tensor var_2469_end_0 = const()[name = tensor("op_2469_end_0"), val = tensor([1, 64, 1, 1500])]; tensor var_2469_end_mask_0 = const()[name = tensor("op_2469_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_2469_cast_fp16 = slice_by_index(begin = var_2469_begin_0, end = var_2469_end_0, end_mask = var_2469_end_mask_0, x = var_2243_cast_fp16)[name = tensor("op_2469_cast_fp16")]; tensor k_7_perm_0 = const()[name = tensor("k_7_perm_0"), val = tensor([0, 3, 2, 1])]; tensor var_2474_begin_0 = const()[name = tensor("op_2474_begin_0"), val = tensor([0, 0, 0, 0])]; tensor var_2474_end_0 = const()[name = tensor("op_2474_end_0"), val = tensor([1, 1500, 1, 64])]; tensor var_2474_end_mask_0 = const()[name = tensor("op_2474_end_mask_0"), val = tensor([true, true, true, false])]; tensor transpose_2 = transpose(perm = k_7_perm_0, x = key_7_cast_fp16)[name = tensor("transpose_2")]; tensor var_2474_cast_fp16 = slice_by_index(begin = var_2474_begin_0, end = var_2474_end_0, end_mask = var_2474_end_mask_0, x = transpose_2)[name = tensor("op_2474_cast_fp16")]; tensor var_2478_begin_0 = const()[name = tensor("op_2478_begin_0"), val = tensor([0, 0, 0, 64])]; tensor var_2478_end_0 = const()[name = tensor("op_2478_end_0"), val = tensor([1, 1500, 1, 128])]; tensor var_2478_end_mask_0 = const()[name = tensor("op_2478_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_2478_cast_fp16 = slice_by_index(begin = var_2478_begin_0, end = var_2478_end_0, end_mask = var_2478_end_mask_0, x = transpose_2)[name = tensor("op_2478_cast_fp16")]; tensor var_2482_begin_0 = const()[name = tensor("op_2482_begin_0"), val = tensor([0, 0, 0, 128])]; tensor var_2482_end_0 = const()[name = tensor("op_2482_end_0"), val = tensor([1, 1500, 1, 192])]; tensor var_2482_end_mask_0 = const()[name = tensor("op_2482_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_2482_cast_fp16 = slice_by_index(begin = var_2482_begin_0, end = var_2482_end_0, end_mask = var_2482_end_mask_0, x = transpose_2)[name = tensor("op_2482_cast_fp16")]; tensor var_2486_begin_0 = const()[name = tensor("op_2486_begin_0"), val = tensor([0, 0, 0, 192])]; tensor var_2486_end_0 = const()[name = tensor("op_2486_end_0"), val = tensor([1, 1500, 1, 256])]; tensor var_2486_end_mask_0 = const()[name = tensor("op_2486_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_2486_cast_fp16 = slice_by_index(begin = var_2486_begin_0, end = var_2486_end_0, end_mask = var_2486_end_mask_0, x = transpose_2)[name = tensor("op_2486_cast_fp16")]; tensor var_2490_begin_0 = const()[name = tensor("op_2490_begin_0"), val = tensor([0, 0, 0, 256])]; tensor var_2490_end_0 = const()[name = tensor("op_2490_end_0"), val = tensor([1, 1500, 1, 320])]; tensor var_2490_end_mask_0 = const()[name = tensor("op_2490_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_2490_cast_fp16 = slice_by_index(begin = var_2490_begin_0, end = var_2490_end_0, end_mask = var_2490_end_mask_0, x = transpose_2)[name = tensor("op_2490_cast_fp16")]; tensor var_2494_begin_0 = const()[name = tensor("op_2494_begin_0"), val = tensor([0, 0, 0, 320])]; tensor var_2494_end_0 = const()[name = tensor("op_2494_end_0"), val = tensor([1, 1500, 1, 384])]; tensor var_2494_end_mask_0 = const()[name = tensor("op_2494_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_2494_cast_fp16 = slice_by_index(begin = var_2494_begin_0, end = var_2494_end_0, end_mask = var_2494_end_mask_0, x = transpose_2)[name = tensor("op_2494_cast_fp16")]; tensor var_2498_begin_0 = const()[name = tensor("op_2498_begin_0"), val = tensor([0, 0, 0, 384])]; tensor var_2498_end_0 = const()[name = tensor("op_2498_end_0"), val = tensor([1, 1500, 1, 448])]; tensor var_2498_end_mask_0 = const()[name = tensor("op_2498_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_2498_cast_fp16 = slice_by_index(begin = var_2498_begin_0, end = var_2498_end_0, end_mask = var_2498_end_mask_0, x = transpose_2)[name = tensor("op_2498_cast_fp16")]; tensor var_2502_begin_0 = const()[name = tensor("op_2502_begin_0"), val = tensor([0, 0, 0, 448])]; tensor var_2502_end_0 = const()[name = tensor("op_2502_end_0"), val = tensor([1, 1500, 1, 512])]; tensor var_2502_end_mask_0 = const()[name = tensor("op_2502_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_2502_cast_fp16 = slice_by_index(begin = var_2502_begin_0, end = var_2502_end_0, end_mask = var_2502_end_mask_0, x = transpose_2)[name = tensor("op_2502_cast_fp16")]; tensor var_2504_begin_0 = const()[name = tensor("op_2504_begin_0"), val = tensor([0, 0, 0, 0])]; tensor var_2504_end_0 = const()[name = tensor("op_2504_end_0"), val = tensor([1, 64, 1, 1500])]; tensor var_2504_end_mask_0 = const()[name = tensor("op_2504_end_mask_0"), val = tensor([true, false, true, true])]; tensor var_2504_cast_fp16 = slice_by_index(begin = var_2504_begin_0, end = var_2504_end_0, end_mask = var_2504_end_mask_0, x = value_7_cast_fp16)[name = tensor("op_2504_cast_fp16")]; tensor var_2508_begin_0 = const()[name = tensor("op_2508_begin_0"), val = tensor([0, 64, 0, 0])]; tensor var_2508_end_0 = const()[name = tensor("op_2508_end_0"), val = tensor([1, 128, 1, 1500])]; tensor var_2508_end_mask_0 = const()[name = tensor("op_2508_end_mask_0"), val = tensor([true, false, true, true])]; tensor var_2508_cast_fp16 = slice_by_index(begin = var_2508_begin_0, end = var_2508_end_0, end_mask = var_2508_end_mask_0, x = value_7_cast_fp16)[name = tensor("op_2508_cast_fp16")]; tensor var_2512_begin_0 = const()[name = tensor("op_2512_begin_0"), val = tensor([0, 128, 0, 0])]; tensor var_2512_end_0 = const()[name = tensor("op_2512_end_0"), val = tensor([1, 192, 1, 1500])]; tensor var_2512_end_mask_0 = const()[name = tensor("op_2512_end_mask_0"), val = tensor([true, false, true, true])]; tensor var_2512_cast_fp16 = slice_by_index(begin = var_2512_begin_0, end = var_2512_end_0, end_mask = var_2512_end_mask_0, x = value_7_cast_fp16)[name = tensor("op_2512_cast_fp16")]; tensor var_2516_begin_0 = const()[name = tensor("op_2516_begin_0"), val = tensor([0, 192, 0, 0])]; tensor var_2516_end_0 = const()[name = tensor("op_2516_end_0"), val = tensor([1, 256, 1, 1500])]; tensor var_2516_end_mask_0 = const()[name = tensor("op_2516_end_mask_0"), val = tensor([true, false, true, true])]; tensor var_2516_cast_fp16 = slice_by_index(begin = var_2516_begin_0, end = var_2516_end_0, end_mask = var_2516_end_mask_0, x = value_7_cast_fp16)[name = tensor("op_2516_cast_fp16")]; tensor var_2520_begin_0 = const()[name = tensor("op_2520_begin_0"), val = tensor([0, 256, 0, 0])]; tensor var_2520_end_0 = const()[name = tensor("op_2520_end_0"), val = tensor([1, 320, 1, 1500])]; tensor var_2520_end_mask_0 = const()[name = tensor("op_2520_end_mask_0"), val = tensor([true, false, true, true])]; tensor var_2520_cast_fp16 = slice_by_index(begin = var_2520_begin_0, end = var_2520_end_0, end_mask = var_2520_end_mask_0, x = value_7_cast_fp16)[name = tensor("op_2520_cast_fp16")]; tensor var_2524_begin_0 = const()[name = tensor("op_2524_begin_0"), val = tensor([0, 320, 0, 0])]; tensor var_2524_end_0 = const()[name = tensor("op_2524_end_0"), val = tensor([1, 384, 1, 1500])]; tensor var_2524_end_mask_0 = const()[name = tensor("op_2524_end_mask_0"), val = tensor([true, false, true, true])]; tensor var_2524_cast_fp16 = slice_by_index(begin = var_2524_begin_0, end = var_2524_end_0, end_mask = var_2524_end_mask_0, x = value_7_cast_fp16)[name = tensor("op_2524_cast_fp16")]; tensor var_2528_begin_0 = const()[name = tensor("op_2528_begin_0"), val = tensor([0, 384, 0, 0])]; tensor var_2528_end_0 = const()[name = tensor("op_2528_end_0"), val = tensor([1, 448, 1, 1500])]; tensor var_2528_end_mask_0 = const()[name = tensor("op_2528_end_mask_0"), val = tensor([true, false, true, true])]; tensor var_2528_cast_fp16 = slice_by_index(begin = var_2528_begin_0, end = var_2528_end_0, end_mask = var_2528_end_mask_0, x = value_7_cast_fp16)[name = tensor("op_2528_cast_fp16")]; tensor var_2532_begin_0 = const()[name = tensor("op_2532_begin_0"), val = tensor([0, 448, 0, 0])]; tensor var_2532_end_0 = const()[name = tensor("op_2532_end_0"), val = tensor([1, 512, 1, 1500])]; tensor var_2532_end_mask_0 = const()[name = tensor("op_2532_end_mask_0"), val = tensor([true, false, true, true])]; tensor var_2532_cast_fp16 = slice_by_index(begin = var_2532_begin_0, end = var_2532_end_0, end_mask = var_2532_end_mask_0, x = value_7_cast_fp16)[name = tensor("op_2532_cast_fp16")]; tensor var_2536_equation_0 = const()[name = tensor("op_2536_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_2536_cast_fp16 = einsum(equation = var_2536_equation_0, values = (var_2474_cast_fp16, var_2252_cast_fp16))[name = tensor("op_2536_cast_fp16")]; tensor var_2537_to_fp16 = const()[name = tensor("op_2537_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_193_cast_fp16 = mul(x = var_2536_cast_fp16, y = var_2537_to_fp16)[name = tensor("aw_chunk_193_cast_fp16")]; tensor var_2540_equation_0 = const()[name = tensor("op_2540_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_2540_cast_fp16 = einsum(equation = var_2540_equation_0, values = (var_2474_cast_fp16, var_2259_cast_fp16))[name = tensor("op_2540_cast_fp16")]; tensor var_2541_to_fp16 = const()[name = tensor("op_2541_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_195_cast_fp16 = mul(x = var_2540_cast_fp16, y = var_2541_to_fp16)[name = tensor("aw_chunk_195_cast_fp16")]; tensor var_2544_equation_0 = const()[name = tensor("op_2544_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_2544_cast_fp16 = einsum(equation = var_2544_equation_0, values = (var_2474_cast_fp16, var_2266_cast_fp16))[name = tensor("op_2544_cast_fp16")]; tensor var_2545_to_fp16 = const()[name = tensor("op_2545_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_197_cast_fp16 = mul(x = var_2544_cast_fp16, y = var_2545_to_fp16)[name = tensor("aw_chunk_197_cast_fp16")]; tensor var_2548_equation_0 = const()[name = tensor("op_2548_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_2548_cast_fp16 = einsum(equation = var_2548_equation_0, values = (var_2474_cast_fp16, var_2273_cast_fp16))[name = tensor("op_2548_cast_fp16")]; tensor var_2549_to_fp16 = const()[name = tensor("op_2549_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_199_cast_fp16 = mul(x = var_2548_cast_fp16, y = var_2549_to_fp16)[name = tensor("aw_chunk_199_cast_fp16")]; tensor var_2552_equation_0 = const()[name = tensor("op_2552_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_2552_cast_fp16 = einsum(equation = var_2552_equation_0, values = (var_2478_cast_fp16, var_2280_cast_fp16))[name = tensor("op_2552_cast_fp16")]; tensor var_2553_to_fp16 = const()[name = tensor("op_2553_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_201_cast_fp16 = mul(x = var_2552_cast_fp16, y = var_2553_to_fp16)[name = tensor("aw_chunk_201_cast_fp16")]; tensor var_2556_equation_0 = const()[name = tensor("op_2556_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_2556_cast_fp16 = einsum(equation = var_2556_equation_0, values = (var_2478_cast_fp16, var_2287_cast_fp16))[name = tensor("op_2556_cast_fp16")]; tensor var_2557_to_fp16 = const()[name = tensor("op_2557_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_203_cast_fp16 = mul(x = var_2556_cast_fp16, y = var_2557_to_fp16)[name = tensor("aw_chunk_203_cast_fp16")]; tensor var_2560_equation_0 = const()[name = tensor("op_2560_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_2560_cast_fp16 = einsum(equation = var_2560_equation_0, values = (var_2478_cast_fp16, var_2294_cast_fp16))[name = tensor("op_2560_cast_fp16")]; tensor var_2561_to_fp16 = const()[name = tensor("op_2561_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_205_cast_fp16 = mul(x = var_2560_cast_fp16, y = var_2561_to_fp16)[name = tensor("aw_chunk_205_cast_fp16")]; tensor var_2564_equation_0 = const()[name = tensor("op_2564_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_2564_cast_fp16 = einsum(equation = var_2564_equation_0, values = (var_2478_cast_fp16, var_2301_cast_fp16))[name = tensor("op_2564_cast_fp16")]; tensor var_2565_to_fp16 = const()[name = tensor("op_2565_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_207_cast_fp16 = mul(x = var_2564_cast_fp16, y = var_2565_to_fp16)[name = tensor("aw_chunk_207_cast_fp16")]; tensor var_2568_equation_0 = const()[name = tensor("op_2568_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_2568_cast_fp16 = einsum(equation = var_2568_equation_0, values = (var_2482_cast_fp16, var_2308_cast_fp16))[name = tensor("op_2568_cast_fp16")]; tensor var_2569_to_fp16 = const()[name = tensor("op_2569_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_209_cast_fp16 = mul(x = var_2568_cast_fp16, y = var_2569_to_fp16)[name = tensor("aw_chunk_209_cast_fp16")]; tensor var_2572_equation_0 = const()[name = tensor("op_2572_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_2572_cast_fp16 = einsum(equation = var_2572_equation_0, values = (var_2482_cast_fp16, var_2315_cast_fp16))[name = tensor("op_2572_cast_fp16")]; tensor var_2573_to_fp16 = const()[name = tensor("op_2573_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_211_cast_fp16 = mul(x = var_2572_cast_fp16, y = var_2573_to_fp16)[name = tensor("aw_chunk_211_cast_fp16")]; tensor var_2576_equation_0 = const()[name = tensor("op_2576_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_2576_cast_fp16 = einsum(equation = var_2576_equation_0, values = (var_2482_cast_fp16, var_2322_cast_fp16))[name = tensor("op_2576_cast_fp16")]; tensor var_2577_to_fp16 = const()[name = tensor("op_2577_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_213_cast_fp16 = mul(x = var_2576_cast_fp16, y = var_2577_to_fp16)[name = tensor("aw_chunk_213_cast_fp16")]; tensor var_2580_equation_0 = const()[name = tensor("op_2580_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_2580_cast_fp16 = einsum(equation = var_2580_equation_0, values = (var_2482_cast_fp16, var_2329_cast_fp16))[name = tensor("op_2580_cast_fp16")]; tensor var_2581_to_fp16 = const()[name = tensor("op_2581_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_215_cast_fp16 = mul(x = var_2580_cast_fp16, y = var_2581_to_fp16)[name = tensor("aw_chunk_215_cast_fp16")]; tensor var_2584_equation_0 = const()[name = tensor("op_2584_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_2584_cast_fp16 = einsum(equation = var_2584_equation_0, values = (var_2486_cast_fp16, var_2336_cast_fp16))[name = tensor("op_2584_cast_fp16")]; tensor var_2585_to_fp16 = const()[name = tensor("op_2585_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_217_cast_fp16 = mul(x = var_2584_cast_fp16, y = var_2585_to_fp16)[name = tensor("aw_chunk_217_cast_fp16")]; tensor var_2588_equation_0 = const()[name = tensor("op_2588_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_2588_cast_fp16 = einsum(equation = var_2588_equation_0, values = (var_2486_cast_fp16, var_2343_cast_fp16))[name = tensor("op_2588_cast_fp16")]; tensor var_2589_to_fp16 = const()[name = tensor("op_2589_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_219_cast_fp16 = mul(x = var_2588_cast_fp16, y = var_2589_to_fp16)[name = tensor("aw_chunk_219_cast_fp16")]; tensor var_2592_equation_0 = const()[name = tensor("op_2592_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_2592_cast_fp16 = einsum(equation = var_2592_equation_0, values = (var_2486_cast_fp16, var_2350_cast_fp16))[name = tensor("op_2592_cast_fp16")]; tensor var_2593_to_fp16 = const()[name = tensor("op_2593_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_221_cast_fp16 = mul(x = var_2592_cast_fp16, y = var_2593_to_fp16)[name = tensor("aw_chunk_221_cast_fp16")]; tensor var_2596_equation_0 = const()[name = tensor("op_2596_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_2596_cast_fp16 = einsum(equation = var_2596_equation_0, values = (var_2486_cast_fp16, var_2357_cast_fp16))[name = tensor("op_2596_cast_fp16")]; tensor var_2597_to_fp16 = const()[name = tensor("op_2597_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_223_cast_fp16 = mul(x = var_2596_cast_fp16, y = var_2597_to_fp16)[name = tensor("aw_chunk_223_cast_fp16")]; tensor var_2600_equation_0 = const()[name = tensor("op_2600_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_2600_cast_fp16 = einsum(equation = var_2600_equation_0, values = (var_2490_cast_fp16, var_2364_cast_fp16))[name = tensor("op_2600_cast_fp16")]; tensor var_2601_to_fp16 = const()[name = tensor("op_2601_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_225_cast_fp16 = mul(x = var_2600_cast_fp16, y = var_2601_to_fp16)[name = tensor("aw_chunk_225_cast_fp16")]; tensor var_2604_equation_0 = const()[name = tensor("op_2604_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_2604_cast_fp16 = einsum(equation = var_2604_equation_0, values = (var_2490_cast_fp16, var_2371_cast_fp16))[name = tensor("op_2604_cast_fp16")]; tensor var_2605_to_fp16 = const()[name = tensor("op_2605_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_227_cast_fp16 = mul(x = var_2604_cast_fp16, y = var_2605_to_fp16)[name = tensor("aw_chunk_227_cast_fp16")]; tensor var_2608_equation_0 = const()[name = tensor("op_2608_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_2608_cast_fp16 = einsum(equation = var_2608_equation_0, values = (var_2490_cast_fp16, var_2378_cast_fp16))[name = tensor("op_2608_cast_fp16")]; tensor var_2609_to_fp16 = const()[name = tensor("op_2609_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_229_cast_fp16 = mul(x = var_2608_cast_fp16, y = var_2609_to_fp16)[name = tensor("aw_chunk_229_cast_fp16")]; tensor var_2612_equation_0 = const()[name = tensor("op_2612_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_2612_cast_fp16 = einsum(equation = var_2612_equation_0, values = (var_2490_cast_fp16, var_2385_cast_fp16))[name = tensor("op_2612_cast_fp16")]; tensor var_2613_to_fp16 = const()[name = tensor("op_2613_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_231_cast_fp16 = mul(x = var_2612_cast_fp16, y = var_2613_to_fp16)[name = tensor("aw_chunk_231_cast_fp16")]; tensor var_2616_equation_0 = const()[name = tensor("op_2616_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_2616_cast_fp16 = einsum(equation = var_2616_equation_0, values = (var_2494_cast_fp16, var_2392_cast_fp16))[name = tensor("op_2616_cast_fp16")]; tensor var_2617_to_fp16 = const()[name = tensor("op_2617_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_233_cast_fp16 = mul(x = var_2616_cast_fp16, y = var_2617_to_fp16)[name = tensor("aw_chunk_233_cast_fp16")]; tensor var_2620_equation_0 = const()[name = tensor("op_2620_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_2620_cast_fp16 = einsum(equation = var_2620_equation_0, values = (var_2494_cast_fp16, var_2399_cast_fp16))[name = tensor("op_2620_cast_fp16")]; tensor var_2621_to_fp16 = const()[name = tensor("op_2621_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_235_cast_fp16 = mul(x = var_2620_cast_fp16, y = var_2621_to_fp16)[name = tensor("aw_chunk_235_cast_fp16")]; tensor var_2624_equation_0 = const()[name = tensor("op_2624_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_2624_cast_fp16 = einsum(equation = var_2624_equation_0, values = (var_2494_cast_fp16, var_2406_cast_fp16))[name = tensor("op_2624_cast_fp16")]; tensor var_2625_to_fp16 = const()[name = tensor("op_2625_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_237_cast_fp16 = mul(x = var_2624_cast_fp16, y = var_2625_to_fp16)[name = tensor("aw_chunk_237_cast_fp16")]; tensor var_2628_equation_0 = const()[name = tensor("op_2628_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_2628_cast_fp16 = einsum(equation = var_2628_equation_0, values = (var_2494_cast_fp16, var_2413_cast_fp16))[name = tensor("op_2628_cast_fp16")]; tensor var_2629_to_fp16 = const()[name = tensor("op_2629_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_239_cast_fp16 = mul(x = var_2628_cast_fp16, y = var_2629_to_fp16)[name = tensor("aw_chunk_239_cast_fp16")]; tensor var_2632_equation_0 = const()[name = tensor("op_2632_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_2632_cast_fp16 = einsum(equation = var_2632_equation_0, values = (var_2498_cast_fp16, var_2420_cast_fp16))[name = tensor("op_2632_cast_fp16")]; tensor var_2633_to_fp16 = const()[name = tensor("op_2633_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_241_cast_fp16 = mul(x = var_2632_cast_fp16, y = var_2633_to_fp16)[name = tensor("aw_chunk_241_cast_fp16")]; tensor var_2636_equation_0 = const()[name = tensor("op_2636_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_2636_cast_fp16 = einsum(equation = var_2636_equation_0, values = (var_2498_cast_fp16, var_2427_cast_fp16))[name = tensor("op_2636_cast_fp16")]; tensor var_2637_to_fp16 = const()[name = tensor("op_2637_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_243_cast_fp16 = mul(x = var_2636_cast_fp16, y = var_2637_to_fp16)[name = tensor("aw_chunk_243_cast_fp16")]; tensor var_2640_equation_0 = const()[name = tensor("op_2640_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_2640_cast_fp16 = einsum(equation = var_2640_equation_0, values = (var_2498_cast_fp16, var_2434_cast_fp16))[name = tensor("op_2640_cast_fp16")]; tensor var_2641_to_fp16 = const()[name = tensor("op_2641_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_245_cast_fp16 = mul(x = var_2640_cast_fp16, y = var_2641_to_fp16)[name = tensor("aw_chunk_245_cast_fp16")]; tensor var_2644_equation_0 = const()[name = tensor("op_2644_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_2644_cast_fp16 = einsum(equation = var_2644_equation_0, values = (var_2498_cast_fp16, var_2441_cast_fp16))[name = tensor("op_2644_cast_fp16")]; tensor var_2645_to_fp16 = const()[name = tensor("op_2645_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_247_cast_fp16 = mul(x = var_2644_cast_fp16, y = var_2645_to_fp16)[name = tensor("aw_chunk_247_cast_fp16")]; tensor var_2648_equation_0 = const()[name = tensor("op_2648_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_2648_cast_fp16 = einsum(equation = var_2648_equation_0, values = (var_2502_cast_fp16, var_2448_cast_fp16))[name = tensor("op_2648_cast_fp16")]; tensor var_2649_to_fp16 = const()[name = tensor("op_2649_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_249_cast_fp16 = mul(x = var_2648_cast_fp16, y = var_2649_to_fp16)[name = tensor("aw_chunk_249_cast_fp16")]; tensor var_2652_equation_0 = const()[name = tensor("op_2652_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_2652_cast_fp16 = einsum(equation = var_2652_equation_0, values = (var_2502_cast_fp16, var_2455_cast_fp16))[name = tensor("op_2652_cast_fp16")]; tensor var_2653_to_fp16 = const()[name = tensor("op_2653_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_251_cast_fp16 = mul(x = var_2652_cast_fp16, y = var_2653_to_fp16)[name = tensor("aw_chunk_251_cast_fp16")]; tensor var_2656_equation_0 = const()[name = tensor("op_2656_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_2656_cast_fp16 = einsum(equation = var_2656_equation_0, values = (var_2502_cast_fp16, var_2462_cast_fp16))[name = tensor("op_2656_cast_fp16")]; tensor var_2657_to_fp16 = const()[name = tensor("op_2657_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_253_cast_fp16 = mul(x = var_2656_cast_fp16, y = var_2657_to_fp16)[name = tensor("aw_chunk_253_cast_fp16")]; tensor var_2660_equation_0 = const()[name = tensor("op_2660_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_2660_cast_fp16 = einsum(equation = var_2660_equation_0, values = (var_2502_cast_fp16, var_2469_cast_fp16))[name = tensor("op_2660_cast_fp16")]; tensor var_2661_to_fp16 = const()[name = tensor("op_2661_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_255_cast_fp16 = mul(x = var_2660_cast_fp16, y = var_2661_to_fp16)[name = tensor("aw_chunk_255_cast_fp16")]; tensor var_2663_cast_fp16 = softmax(axis = var_2160, x = aw_chunk_193_cast_fp16)[name = tensor("op_2663_cast_fp16")]; tensor var_2664_cast_fp16 = softmax(axis = var_2160, x = aw_chunk_195_cast_fp16)[name = tensor("op_2664_cast_fp16")]; tensor var_2665_cast_fp16 = softmax(axis = var_2160, x = aw_chunk_197_cast_fp16)[name = tensor("op_2665_cast_fp16")]; tensor var_2666_cast_fp16 = softmax(axis = var_2160, x = aw_chunk_199_cast_fp16)[name = tensor("op_2666_cast_fp16")]; tensor var_2667_cast_fp16 = softmax(axis = var_2160, x = aw_chunk_201_cast_fp16)[name = tensor("op_2667_cast_fp16")]; tensor var_2668_cast_fp16 = softmax(axis = var_2160, x = aw_chunk_203_cast_fp16)[name = tensor("op_2668_cast_fp16")]; tensor var_2669_cast_fp16 = softmax(axis = var_2160, x = aw_chunk_205_cast_fp16)[name = tensor("op_2669_cast_fp16")]; tensor var_2670_cast_fp16 = softmax(axis = var_2160, x = aw_chunk_207_cast_fp16)[name = tensor("op_2670_cast_fp16")]; tensor var_2671_cast_fp16 = softmax(axis = var_2160, x = aw_chunk_209_cast_fp16)[name = tensor("op_2671_cast_fp16")]; tensor var_2672_cast_fp16 = softmax(axis = var_2160, x = aw_chunk_211_cast_fp16)[name = tensor("op_2672_cast_fp16")]; tensor var_2673_cast_fp16 = softmax(axis = var_2160, x = aw_chunk_213_cast_fp16)[name = tensor("op_2673_cast_fp16")]; tensor var_2674_cast_fp16 = softmax(axis = var_2160, x = aw_chunk_215_cast_fp16)[name = tensor("op_2674_cast_fp16")]; tensor var_2675_cast_fp16 = softmax(axis = var_2160, x = aw_chunk_217_cast_fp16)[name = tensor("op_2675_cast_fp16")]; tensor var_2676_cast_fp16 = softmax(axis = var_2160, x = aw_chunk_219_cast_fp16)[name = tensor("op_2676_cast_fp16")]; tensor var_2677_cast_fp16 = softmax(axis = var_2160, x = aw_chunk_221_cast_fp16)[name = tensor("op_2677_cast_fp16")]; tensor var_2678_cast_fp16 = softmax(axis = var_2160, x = aw_chunk_223_cast_fp16)[name = tensor("op_2678_cast_fp16")]; tensor var_2679_cast_fp16 = softmax(axis = var_2160, x = aw_chunk_225_cast_fp16)[name = tensor("op_2679_cast_fp16")]; tensor var_2680_cast_fp16 = softmax(axis = var_2160, x = aw_chunk_227_cast_fp16)[name = tensor("op_2680_cast_fp16")]; tensor var_2681_cast_fp16 = softmax(axis = var_2160, x = aw_chunk_229_cast_fp16)[name = tensor("op_2681_cast_fp16")]; tensor var_2682_cast_fp16 = softmax(axis = var_2160, x = aw_chunk_231_cast_fp16)[name = tensor("op_2682_cast_fp16")]; tensor var_2683_cast_fp16 = softmax(axis = var_2160, x = aw_chunk_233_cast_fp16)[name = tensor("op_2683_cast_fp16")]; tensor var_2684_cast_fp16 = softmax(axis = var_2160, x = aw_chunk_235_cast_fp16)[name = tensor("op_2684_cast_fp16")]; tensor var_2685_cast_fp16 = softmax(axis = var_2160, x = aw_chunk_237_cast_fp16)[name = tensor("op_2685_cast_fp16")]; tensor var_2686_cast_fp16 = softmax(axis = var_2160, x = aw_chunk_239_cast_fp16)[name = tensor("op_2686_cast_fp16")]; tensor var_2687_cast_fp16 = softmax(axis = var_2160, x = aw_chunk_241_cast_fp16)[name = tensor("op_2687_cast_fp16")]; tensor var_2688_cast_fp16 = softmax(axis = var_2160, x = aw_chunk_243_cast_fp16)[name = tensor("op_2688_cast_fp16")]; tensor var_2689_cast_fp16 = softmax(axis = var_2160, x = aw_chunk_245_cast_fp16)[name = tensor("op_2689_cast_fp16")]; tensor var_2690_cast_fp16 = softmax(axis = var_2160, x = aw_chunk_247_cast_fp16)[name = tensor("op_2690_cast_fp16")]; tensor var_2691_cast_fp16 = softmax(axis = var_2160, x = aw_chunk_249_cast_fp16)[name = tensor("op_2691_cast_fp16")]; tensor var_2692_cast_fp16 = softmax(axis = var_2160, x = aw_chunk_251_cast_fp16)[name = tensor("op_2692_cast_fp16")]; tensor var_2693_cast_fp16 = softmax(axis = var_2160, x = aw_chunk_253_cast_fp16)[name = tensor("op_2693_cast_fp16")]; tensor var_2694_cast_fp16 = softmax(axis = var_2160, x = aw_chunk_255_cast_fp16)[name = tensor("op_2694_cast_fp16")]; tensor var_2696_equation_0 = const()[name = tensor("op_2696_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_2696_cast_fp16 = einsum(equation = var_2696_equation_0, values = (var_2504_cast_fp16, var_2663_cast_fp16))[name = tensor("op_2696_cast_fp16")]; tensor var_2698_equation_0 = const()[name = tensor("op_2698_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_2698_cast_fp16 = einsum(equation = var_2698_equation_0, values = (var_2504_cast_fp16, var_2664_cast_fp16))[name = tensor("op_2698_cast_fp16")]; tensor var_2700_equation_0 = const()[name = tensor("op_2700_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_2700_cast_fp16 = einsum(equation = var_2700_equation_0, values = (var_2504_cast_fp16, var_2665_cast_fp16))[name = tensor("op_2700_cast_fp16")]; tensor var_2702_equation_0 = const()[name = tensor("op_2702_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_2702_cast_fp16 = einsum(equation = var_2702_equation_0, values = (var_2504_cast_fp16, var_2666_cast_fp16))[name = tensor("op_2702_cast_fp16")]; tensor var_2704_equation_0 = const()[name = tensor("op_2704_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_2704_cast_fp16 = einsum(equation = var_2704_equation_0, values = (var_2508_cast_fp16, var_2667_cast_fp16))[name = tensor("op_2704_cast_fp16")]; tensor var_2706_equation_0 = const()[name = tensor("op_2706_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_2706_cast_fp16 = einsum(equation = var_2706_equation_0, values = (var_2508_cast_fp16, var_2668_cast_fp16))[name = tensor("op_2706_cast_fp16")]; tensor var_2708_equation_0 = const()[name = tensor("op_2708_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_2708_cast_fp16 = einsum(equation = var_2708_equation_0, values = (var_2508_cast_fp16, var_2669_cast_fp16))[name = tensor("op_2708_cast_fp16")]; tensor var_2710_equation_0 = const()[name = tensor("op_2710_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_2710_cast_fp16 = einsum(equation = var_2710_equation_0, values = (var_2508_cast_fp16, var_2670_cast_fp16))[name = tensor("op_2710_cast_fp16")]; tensor var_2712_equation_0 = const()[name = tensor("op_2712_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_2712_cast_fp16 = einsum(equation = var_2712_equation_0, values = (var_2512_cast_fp16, var_2671_cast_fp16))[name = tensor("op_2712_cast_fp16")]; tensor var_2714_equation_0 = const()[name = tensor("op_2714_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_2714_cast_fp16 = einsum(equation = var_2714_equation_0, values = (var_2512_cast_fp16, var_2672_cast_fp16))[name = tensor("op_2714_cast_fp16")]; tensor var_2716_equation_0 = const()[name = tensor("op_2716_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_2716_cast_fp16 = einsum(equation = var_2716_equation_0, values = (var_2512_cast_fp16, var_2673_cast_fp16))[name = tensor("op_2716_cast_fp16")]; tensor var_2718_equation_0 = const()[name = tensor("op_2718_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_2718_cast_fp16 = einsum(equation = var_2718_equation_0, values = (var_2512_cast_fp16, var_2674_cast_fp16))[name = tensor("op_2718_cast_fp16")]; tensor var_2720_equation_0 = const()[name = tensor("op_2720_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_2720_cast_fp16 = einsum(equation = var_2720_equation_0, values = (var_2516_cast_fp16, var_2675_cast_fp16))[name = tensor("op_2720_cast_fp16")]; tensor var_2722_equation_0 = const()[name = tensor("op_2722_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_2722_cast_fp16 = einsum(equation = var_2722_equation_0, values = (var_2516_cast_fp16, var_2676_cast_fp16))[name = tensor("op_2722_cast_fp16")]; tensor var_2724_equation_0 = const()[name = tensor("op_2724_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_2724_cast_fp16 = einsum(equation = var_2724_equation_0, values = (var_2516_cast_fp16, var_2677_cast_fp16))[name = tensor("op_2724_cast_fp16")]; tensor var_2726_equation_0 = const()[name = tensor("op_2726_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_2726_cast_fp16 = einsum(equation = var_2726_equation_0, values = (var_2516_cast_fp16, var_2678_cast_fp16))[name = tensor("op_2726_cast_fp16")]; tensor var_2728_equation_0 = const()[name = tensor("op_2728_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_2728_cast_fp16 = einsum(equation = var_2728_equation_0, values = (var_2520_cast_fp16, var_2679_cast_fp16))[name = tensor("op_2728_cast_fp16")]; tensor var_2730_equation_0 = const()[name = tensor("op_2730_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_2730_cast_fp16 = einsum(equation = var_2730_equation_0, values = (var_2520_cast_fp16, var_2680_cast_fp16))[name = tensor("op_2730_cast_fp16")]; tensor var_2732_equation_0 = const()[name = tensor("op_2732_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_2732_cast_fp16 = einsum(equation = var_2732_equation_0, values = (var_2520_cast_fp16, var_2681_cast_fp16))[name = tensor("op_2732_cast_fp16")]; tensor var_2734_equation_0 = const()[name = tensor("op_2734_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_2734_cast_fp16 = einsum(equation = var_2734_equation_0, values = (var_2520_cast_fp16, var_2682_cast_fp16))[name = tensor("op_2734_cast_fp16")]; tensor var_2736_equation_0 = const()[name = tensor("op_2736_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_2736_cast_fp16 = einsum(equation = var_2736_equation_0, values = (var_2524_cast_fp16, var_2683_cast_fp16))[name = tensor("op_2736_cast_fp16")]; tensor var_2738_equation_0 = const()[name = tensor("op_2738_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_2738_cast_fp16 = einsum(equation = var_2738_equation_0, values = (var_2524_cast_fp16, var_2684_cast_fp16))[name = tensor("op_2738_cast_fp16")]; tensor var_2740_equation_0 = const()[name = tensor("op_2740_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_2740_cast_fp16 = einsum(equation = var_2740_equation_0, values = (var_2524_cast_fp16, var_2685_cast_fp16))[name = tensor("op_2740_cast_fp16")]; tensor var_2742_equation_0 = const()[name = tensor("op_2742_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_2742_cast_fp16 = einsum(equation = var_2742_equation_0, values = (var_2524_cast_fp16, var_2686_cast_fp16))[name = tensor("op_2742_cast_fp16")]; tensor var_2744_equation_0 = const()[name = tensor("op_2744_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_2744_cast_fp16 = einsum(equation = var_2744_equation_0, values = (var_2528_cast_fp16, var_2687_cast_fp16))[name = tensor("op_2744_cast_fp16")]; tensor var_2746_equation_0 = const()[name = tensor("op_2746_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_2746_cast_fp16 = einsum(equation = var_2746_equation_0, values = (var_2528_cast_fp16, var_2688_cast_fp16))[name = tensor("op_2746_cast_fp16")]; tensor var_2748_equation_0 = const()[name = tensor("op_2748_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_2748_cast_fp16 = einsum(equation = var_2748_equation_0, values = (var_2528_cast_fp16, var_2689_cast_fp16))[name = tensor("op_2748_cast_fp16")]; tensor var_2750_equation_0 = const()[name = tensor("op_2750_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_2750_cast_fp16 = einsum(equation = var_2750_equation_0, values = (var_2528_cast_fp16, var_2690_cast_fp16))[name = tensor("op_2750_cast_fp16")]; tensor var_2752_equation_0 = const()[name = tensor("op_2752_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_2752_cast_fp16 = einsum(equation = var_2752_equation_0, values = (var_2532_cast_fp16, var_2691_cast_fp16))[name = tensor("op_2752_cast_fp16")]; tensor var_2754_equation_0 = const()[name = tensor("op_2754_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_2754_cast_fp16 = einsum(equation = var_2754_equation_0, values = (var_2532_cast_fp16, var_2692_cast_fp16))[name = tensor("op_2754_cast_fp16")]; tensor var_2756_equation_0 = const()[name = tensor("op_2756_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_2756_cast_fp16 = einsum(equation = var_2756_equation_0, values = (var_2532_cast_fp16, var_2693_cast_fp16))[name = tensor("op_2756_cast_fp16")]; tensor var_2758_equation_0 = const()[name = tensor("op_2758_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_2758_cast_fp16 = einsum(equation = var_2758_equation_0, values = (var_2532_cast_fp16, var_2694_cast_fp16))[name = tensor("op_2758_cast_fp16")]; tensor var_2760_interleave_0 = const()[name = tensor("op_2760_interleave_0"), val = tensor(false)]; tensor var_2760_cast_fp16 = concat(axis = var_2147, interleave = var_2760_interleave_0, values = (var_2696_cast_fp16, var_2698_cast_fp16, var_2700_cast_fp16, var_2702_cast_fp16))[name = tensor("op_2760_cast_fp16")]; tensor var_2762_interleave_0 = const()[name = tensor("op_2762_interleave_0"), val = tensor(false)]; tensor var_2762_cast_fp16 = concat(axis = var_2147, interleave = var_2762_interleave_0, values = (var_2704_cast_fp16, var_2706_cast_fp16, var_2708_cast_fp16, var_2710_cast_fp16))[name = tensor("op_2762_cast_fp16")]; tensor var_2764_interleave_0 = const()[name = tensor("op_2764_interleave_0"), val = tensor(false)]; tensor var_2764_cast_fp16 = concat(axis = var_2147, interleave = var_2764_interleave_0, values = (var_2712_cast_fp16, var_2714_cast_fp16, var_2716_cast_fp16, var_2718_cast_fp16))[name = tensor("op_2764_cast_fp16")]; tensor var_2766_interleave_0 = const()[name = tensor("op_2766_interleave_0"), val = tensor(false)]; tensor var_2766_cast_fp16 = concat(axis = var_2147, interleave = var_2766_interleave_0, values = (var_2720_cast_fp16, var_2722_cast_fp16, var_2724_cast_fp16, var_2726_cast_fp16))[name = tensor("op_2766_cast_fp16")]; tensor var_2768_interleave_0 = const()[name = tensor("op_2768_interleave_0"), val = tensor(false)]; tensor var_2768_cast_fp16 = concat(axis = var_2147, interleave = var_2768_interleave_0, values = (var_2728_cast_fp16, var_2730_cast_fp16, var_2732_cast_fp16, var_2734_cast_fp16))[name = tensor("op_2768_cast_fp16")]; tensor var_2770_interleave_0 = const()[name = tensor("op_2770_interleave_0"), val = tensor(false)]; tensor var_2770_cast_fp16 = concat(axis = var_2147, interleave = var_2770_interleave_0, values = (var_2736_cast_fp16, var_2738_cast_fp16, var_2740_cast_fp16, var_2742_cast_fp16))[name = tensor("op_2770_cast_fp16")]; tensor var_2772_interleave_0 = const()[name = tensor("op_2772_interleave_0"), val = tensor(false)]; tensor var_2772_cast_fp16 = concat(axis = var_2147, interleave = var_2772_interleave_0, values = (var_2744_cast_fp16, var_2746_cast_fp16, var_2748_cast_fp16, var_2750_cast_fp16))[name = tensor("op_2772_cast_fp16")]; tensor var_2774_interleave_0 = const()[name = tensor("op_2774_interleave_0"), val = tensor(false)]; tensor var_2774_cast_fp16 = concat(axis = var_2147, interleave = var_2774_interleave_0, values = (var_2752_cast_fp16, var_2754_cast_fp16, var_2756_cast_fp16, var_2758_cast_fp16))[name = tensor("op_2774_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_2160, interleave = input_25_interleave_0, values = (var_2760_cast_fp16, var_2762_cast_fp16, var_2764_cast_fp16, var_2766_cast_fp16, var_2768_cast_fp16, var_2770_cast_fp16, var_2772_cast_fp16, var_2774_cast_fp16))[name = tensor("input_25_cast_fp16")]; tensor var_2779 = const()[name = tensor("op_2779"), val = tensor([1, 1])]; tensor var_2781 = const()[name = tensor("op_2781"), val = tensor([1, 1])]; tensor obj_15_pad_type_0 = const()[name = tensor("obj_15_pad_type_0"), val = tensor("custom")]; tensor obj_15_pad_0 = const()[name = tensor("obj_15_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(23850752)))]; 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(24375104)))]; tensor obj_15_cast_fp16 = conv(bias = layers_3_self_attn_o_proj_bias_to_fp16, dilations = var_2781, groups = var_2160, pad = obj_15_pad_0, pad_type = obj_15_pad_type_0, strides = var_2779, weight = layers_3_self_attn_o_proj_weight_to_fp16, x = input_25_cast_fp16)[name = tensor("obj_15_cast_fp16")]; tensor inputs_15_cast_fp16 = add(x = inputs_13_cast_fp16, y = obj_15_cast_fp16)[name = tensor("inputs_15_cast_fp16")]; tensor var_2787 = const()[name = tensor("op_2787"), val = tensor([1])]; tensor channels_mean_15_cast_fp16 = reduce_mean(axes = var_2787, keep_dims = var_2161, 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_2791 = const()[name = tensor("op_2791"), val = tensor([1])]; tensor var_2792_cast_fp16 = reduce_mean(axes = var_2791, keep_dims = var_2161, x = zero_mean_sq_15_cast_fp16)[name = tensor("op_2792_cast_fp16")]; tensor var_2793_to_fp16 = const()[name = tensor("op_2793_to_fp16"), val = tensor(0x1.5p-17)]; tensor var_2794_cast_fp16 = add(x = var_2792_cast_fp16, y = var_2793_to_fp16)[name = tensor("op_2794_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_2794_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(24376192)))]; 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(24377280)))]; 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_2805 = const()[name = tensor("op_2805"), val = tensor([1, 1])]; tensor var_2807 = const()[name = tensor("op_2807"), 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(24378368)))]; 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(26475584)))]; tensor input_29_cast_fp16 = conv(bias = layers_3_fc1_bias_to_fp16, dilations = var_2807, groups = var_2160, pad = input_29_pad_0, pad_type = input_29_pad_type_0, strides = var_2805, weight = layers_3_fc1_weight_to_fp16, x = input_27_cast_fp16)[name = tensor("input_29_cast_fp16")]; tensor input_31_mode_0 = const()[name = tensor("input_31_mode_0"), val = tensor("EXACT")]; tensor input_31_cast_fp16 = gelu(mode = input_31_mode_0, x = input_29_cast_fp16)[name = tensor("input_31_cast_fp16")]; tensor var_2813 = const()[name = tensor("op_2813"), val = tensor([1, 1])]; tensor var_2815 = const()[name = tensor("op_2815"), val = tensor([1, 1])]; tensor hidden_states_11_pad_type_0 = const()[name = tensor("hidden_states_11_pad_type_0"), val = tensor("custom")]; tensor hidden_states_11_pad_0 = const()[name = tensor("hidden_states_11_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(26479744)))]; 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(28576960)))]; tensor hidden_states_11_cast_fp16 = conv(bias = layers_3_fc2_bias_to_fp16, dilations = var_2815, groups = var_2160, pad = hidden_states_11_pad_0, pad_type = hidden_states_11_pad_type_0, strides = var_2813, weight = layers_3_fc2_weight_to_fp16, x = input_31_cast_fp16)[name = tensor("hidden_states_11_cast_fp16")]; tensor inputs_17_cast_fp16 = add(x = inputs_15_cast_fp16, y = hidden_states_11_cast_fp16)[name = tensor("inputs_17_cast_fp16")]; tensor var_2822 = const()[name = tensor("op_2822"), val = tensor(3)]; tensor var_2835 = const()[name = tensor("op_2835"), val = tensor(1)]; tensor var_2836 = const()[name = tensor("op_2836"), val = tensor(true)]; tensor var_2846 = const()[name = tensor("op_2846"), val = tensor([1])]; tensor channels_mean_17_cast_fp16 = reduce_mean(axes = var_2846, keep_dims = var_2836, x = inputs_17_cast_fp16)[name = tensor("channels_mean_17_cast_fp16")]; tensor zero_mean_17_cast_fp16 = sub(x = inputs_17_cast_fp16, y = channels_mean_17_cast_fp16)[name = tensor("zero_mean_17_cast_fp16")]; tensor zero_mean_sq_17_cast_fp16 = mul(x = zero_mean_17_cast_fp16, y = zero_mean_17_cast_fp16)[name = tensor("zero_mean_sq_17_cast_fp16")]; tensor var_2850 = const()[name = tensor("op_2850"), val = tensor([1])]; tensor var_2851_cast_fp16 = reduce_mean(axes = var_2850, keep_dims = var_2836, x = zero_mean_sq_17_cast_fp16)[name = tensor("op_2851_cast_fp16")]; tensor var_2852_to_fp16 = const()[name = tensor("op_2852_to_fp16"), val = tensor(0x1.5p-17)]; tensor var_2853_cast_fp16 = add(x = var_2851_cast_fp16, y = var_2852_to_fp16)[name = tensor("op_2853_cast_fp16")]; tensor denom_17_epsilon_0_to_fp16 = const()[name = tensor("denom_17_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; tensor denom_17_cast_fp16 = rsqrt(epsilon = denom_17_epsilon_0_to_fp16, x = var_2853_cast_fp16)[name = tensor("denom_17_cast_fp16")]; tensor out_17_cast_fp16 = mul(x = zero_mean_17_cast_fp16, y = denom_17_cast_fp16)[name = tensor("out_17_cast_fp16")]; tensor obj_17_gamma_0_to_fp16 = const()[name = tensor("obj_17_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28578048)))]; tensor obj_17_beta_0_to_fp16 = const()[name = tensor("obj_17_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28579136)))]; tensor obj_17_epsilon_0_to_fp16 = const()[name = tensor("obj_17_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; tensor obj_17_cast_fp16 = batch_norm(beta = obj_17_beta_0_to_fp16, epsilon = obj_17_epsilon_0_to_fp16, gamma = obj_17_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_17_cast_fp16)[name = tensor("obj_17_cast_fp16")]; tensor var_2868 = const()[name = tensor("op_2868"), val = tensor([1, 1])]; tensor var_2870 = const()[name = tensor("op_2870"), val = tensor([1, 1])]; tensor query_9_pad_type_0 = const()[name = tensor("query_9_pad_type_0"), val = tensor("custom")]; tensor query_9_pad_0 = const()[name = tensor("query_9_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_4_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_4_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28580224)))]; tensor layers_4_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_4_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29104576)))]; tensor query_9_cast_fp16 = conv(bias = layers_4_self_attn_q_proj_bias_to_fp16, dilations = var_2870, groups = var_2835, pad = query_9_pad_0, pad_type = query_9_pad_type_0, strides = var_2868, weight = layers_4_self_attn_q_proj_weight_to_fp16, x = obj_17_cast_fp16)[name = tensor("query_9_cast_fp16")]; tensor var_2874 = const()[name = tensor("op_2874"), val = tensor([1, 1])]; tensor var_2876 = const()[name = tensor("op_2876"), val = tensor([1, 1])]; tensor key_9_pad_type_0 = const()[name = tensor("key_9_pad_type_0"), val = tensor("custom")]; tensor key_9_pad_0 = const()[name = tensor("key_9_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_4_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_4_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29105664)))]; tensor key_9_cast_fp16 = conv(dilations = var_2876, groups = var_2835, pad = key_9_pad_0, pad_type = key_9_pad_type_0, strides = var_2874, weight = layers_4_self_attn_k_proj_weight_to_fp16, x = obj_17_cast_fp16)[name = tensor("key_9_cast_fp16")]; tensor var_2881 = const()[name = tensor("op_2881"), val = tensor([1, 1])]; tensor var_2883 = const()[name = tensor("op_2883"), val = tensor([1, 1])]; tensor value_9_pad_type_0 = const()[name = tensor("value_9_pad_type_0"), val = tensor("custom")]; tensor value_9_pad_0 = const()[name = tensor("value_9_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_4_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_4_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29630016)))]; tensor layers_4_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_4_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30154368)))]; tensor value_9_cast_fp16 = conv(bias = layers_4_self_attn_v_proj_bias_to_fp16, dilations = var_2883, groups = var_2835, pad = value_9_pad_0, pad_type = value_9_pad_type_0, strides = var_2881, weight = layers_4_self_attn_v_proj_weight_to_fp16, x = obj_17_cast_fp16)[name = tensor("value_9_cast_fp16")]; tensor var_2890_begin_0 = const()[name = tensor("op_2890_begin_0"), val = tensor([0, 0, 0, 0])]; tensor var_2890_end_0 = const()[name = tensor("op_2890_end_0"), val = tensor([1, 64, 1, 1500])]; tensor var_2890_end_mask_0 = const()[name = tensor("op_2890_end_mask_0"), val = tensor([true, false, true, true])]; tensor var_2890_cast_fp16 = slice_by_index(begin = var_2890_begin_0, end = var_2890_end_0, end_mask = var_2890_end_mask_0, x = query_9_cast_fp16)[name = tensor("op_2890_cast_fp16")]; tensor var_2894_begin_0 = const()[name = tensor("op_2894_begin_0"), val = tensor([0, 64, 0, 0])]; tensor var_2894_end_0 = const()[name = tensor("op_2894_end_0"), val = tensor([1, 128, 1, 1500])]; tensor var_2894_end_mask_0 = const()[name = tensor("op_2894_end_mask_0"), val = tensor([true, false, true, true])]; tensor var_2894_cast_fp16 = slice_by_index(begin = var_2894_begin_0, end = var_2894_end_0, end_mask = var_2894_end_mask_0, x = query_9_cast_fp16)[name = tensor("op_2894_cast_fp16")]; tensor var_2898_begin_0 = const()[name = tensor("op_2898_begin_0"), val = tensor([0, 128, 0, 0])]; tensor var_2898_end_0 = const()[name = tensor("op_2898_end_0"), val = tensor([1, 192, 1, 1500])]; tensor var_2898_end_mask_0 = const()[name = tensor("op_2898_end_mask_0"), val = tensor([true, false, true, true])]; tensor var_2898_cast_fp16 = slice_by_index(begin = var_2898_begin_0, end = var_2898_end_0, end_mask = var_2898_end_mask_0, x = query_9_cast_fp16)[name = tensor("op_2898_cast_fp16")]; tensor var_2902_begin_0 = const()[name = tensor("op_2902_begin_0"), val = tensor([0, 192, 0, 0])]; tensor var_2902_end_0 = const()[name = tensor("op_2902_end_0"), val = tensor([1, 256, 1, 1500])]; tensor var_2902_end_mask_0 = const()[name = tensor("op_2902_end_mask_0"), val = tensor([true, false, true, true])]; tensor var_2902_cast_fp16 = slice_by_index(begin = var_2902_begin_0, end = var_2902_end_0, end_mask = var_2902_end_mask_0, x = query_9_cast_fp16)[name = tensor("op_2902_cast_fp16")]; tensor var_2906_begin_0 = const()[name = tensor("op_2906_begin_0"), val = tensor([0, 256, 0, 0])]; tensor var_2906_end_0 = const()[name = tensor("op_2906_end_0"), val = tensor([1, 320, 1, 1500])]; tensor var_2906_end_mask_0 = const()[name = tensor("op_2906_end_mask_0"), val = tensor([true, false, true, true])]; tensor var_2906_cast_fp16 = slice_by_index(begin = var_2906_begin_0, end = var_2906_end_0, end_mask = var_2906_end_mask_0, x = query_9_cast_fp16)[name = tensor("op_2906_cast_fp16")]; tensor var_2910_begin_0 = const()[name = tensor("op_2910_begin_0"), val = tensor([0, 320, 0, 0])]; tensor var_2910_end_0 = const()[name = tensor("op_2910_end_0"), val = tensor([1, 384, 1, 1500])]; tensor var_2910_end_mask_0 = const()[name = tensor("op_2910_end_mask_0"), val = tensor([true, false, true, true])]; tensor var_2910_cast_fp16 = slice_by_index(begin = var_2910_begin_0, end = var_2910_end_0, end_mask = var_2910_end_mask_0, x = query_9_cast_fp16)[name = tensor("op_2910_cast_fp16")]; tensor var_2914_begin_0 = const()[name = tensor("op_2914_begin_0"), val = tensor([0, 384, 0, 0])]; tensor var_2914_end_0 = const()[name = tensor("op_2914_end_0"), val = tensor([1, 448, 1, 1500])]; tensor var_2914_end_mask_0 = const()[name = tensor("op_2914_end_mask_0"), val = tensor([true, false, true, true])]; tensor var_2914_cast_fp16 = slice_by_index(begin = var_2914_begin_0, end = var_2914_end_0, end_mask = var_2914_end_mask_0, x = query_9_cast_fp16)[name = tensor("op_2914_cast_fp16")]; tensor var_2918_begin_0 = const()[name = tensor("op_2918_begin_0"), val = tensor([0, 448, 0, 0])]; tensor var_2918_end_0 = const()[name = tensor("op_2918_end_0"), val = tensor([1, 512, 1, 1500])]; tensor var_2918_end_mask_0 = const()[name = tensor("op_2918_end_mask_0"), val = tensor([true, false, true, true])]; tensor var_2918_cast_fp16 = slice_by_index(begin = var_2918_begin_0, end = var_2918_end_0, end_mask = var_2918_end_mask_0, x = query_9_cast_fp16)[name = tensor("op_2918_cast_fp16")]; tensor var_2927_begin_0 = const()[name = tensor("op_2927_begin_0"), val = tensor([0, 0, 0, 0])]; tensor var_2927_end_0 = const()[name = tensor("op_2927_end_0"), val = tensor([1, 64, 1, 375])]; tensor var_2927_end_mask_0 = const()[name = tensor("op_2927_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_2927_cast_fp16 = slice_by_index(begin = var_2927_begin_0, end = var_2927_end_0, end_mask = var_2927_end_mask_0, x = var_2890_cast_fp16)[name = tensor("op_2927_cast_fp16")]; tensor var_2934_begin_0 = const()[name = tensor("op_2934_begin_0"), val = tensor([0, 0, 0, 375])]; tensor var_2934_end_0 = const()[name = tensor("op_2934_end_0"), val = tensor([1, 64, 1, 750])]; tensor var_2934_end_mask_0 = const()[name = tensor("op_2934_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_2934_cast_fp16 = slice_by_index(begin = var_2934_begin_0, end = var_2934_end_0, end_mask = var_2934_end_mask_0, x = var_2890_cast_fp16)[name = tensor("op_2934_cast_fp16")]; tensor var_2941_begin_0 = const()[name = tensor("op_2941_begin_0"), val = tensor([0, 0, 0, 750])]; tensor var_2941_end_0 = const()[name = tensor("op_2941_end_0"), val = tensor([1, 64, 1, 1125])]; tensor var_2941_end_mask_0 = const()[name = tensor("op_2941_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_2941_cast_fp16 = slice_by_index(begin = var_2941_begin_0, end = var_2941_end_0, end_mask = var_2941_end_mask_0, x = var_2890_cast_fp16)[name = tensor("op_2941_cast_fp16")]; tensor var_2948_begin_0 = const()[name = tensor("op_2948_begin_0"), val = tensor([0, 0, 0, 1125])]; tensor var_2948_end_0 = const()[name = tensor("op_2948_end_0"), val = tensor([1, 64, 1, 1500])]; tensor var_2948_end_mask_0 = const()[name = tensor("op_2948_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_2948_cast_fp16 = slice_by_index(begin = var_2948_begin_0, end = var_2948_end_0, end_mask = var_2948_end_mask_0, x = var_2890_cast_fp16)[name = tensor("op_2948_cast_fp16")]; tensor var_2955_begin_0 = const()[name = tensor("op_2955_begin_0"), val = tensor([0, 0, 0, 0])]; tensor var_2955_end_0 = const()[name = tensor("op_2955_end_0"), val = tensor([1, 64, 1, 375])]; tensor var_2955_end_mask_0 = const()[name = tensor("op_2955_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_2955_cast_fp16 = slice_by_index(begin = var_2955_begin_0, end = var_2955_end_0, end_mask = var_2955_end_mask_0, x = var_2894_cast_fp16)[name = tensor("op_2955_cast_fp16")]; tensor var_2962_begin_0 = const()[name = tensor("op_2962_begin_0"), val = tensor([0, 0, 0, 375])]; tensor var_2962_end_0 = const()[name = tensor("op_2962_end_0"), val = tensor([1, 64, 1, 750])]; tensor var_2962_end_mask_0 = const()[name = tensor("op_2962_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_2962_cast_fp16 = slice_by_index(begin = var_2962_begin_0, end = var_2962_end_0, end_mask = var_2962_end_mask_0, x = var_2894_cast_fp16)[name = tensor("op_2962_cast_fp16")]; tensor var_2969_begin_0 = const()[name = tensor("op_2969_begin_0"), val = tensor([0, 0, 0, 750])]; tensor var_2969_end_0 = const()[name = tensor("op_2969_end_0"), val = tensor([1, 64, 1, 1125])]; tensor var_2969_end_mask_0 = const()[name = tensor("op_2969_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_2969_cast_fp16 = slice_by_index(begin = var_2969_begin_0, end = var_2969_end_0, end_mask = var_2969_end_mask_0, x = var_2894_cast_fp16)[name = tensor("op_2969_cast_fp16")]; tensor var_2976_begin_0 = const()[name = tensor("op_2976_begin_0"), val = tensor([0, 0, 0, 1125])]; tensor var_2976_end_0 = const()[name = tensor("op_2976_end_0"), val = tensor([1, 64, 1, 1500])]; tensor var_2976_end_mask_0 = const()[name = tensor("op_2976_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_2976_cast_fp16 = slice_by_index(begin = var_2976_begin_0, end = var_2976_end_0, end_mask = var_2976_end_mask_0, x = var_2894_cast_fp16)[name = tensor("op_2976_cast_fp16")]; tensor var_2983_begin_0 = const()[name = tensor("op_2983_begin_0"), val = tensor([0, 0, 0, 0])]; tensor var_2983_end_0 = const()[name = tensor("op_2983_end_0"), val = tensor([1, 64, 1, 375])]; tensor var_2983_end_mask_0 = const()[name = tensor("op_2983_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_2983_cast_fp16 = slice_by_index(begin = var_2983_begin_0, end = var_2983_end_0, end_mask = var_2983_end_mask_0, x = var_2898_cast_fp16)[name = tensor("op_2983_cast_fp16")]; tensor var_2990_begin_0 = const()[name = tensor("op_2990_begin_0"), val = tensor([0, 0, 0, 375])]; tensor var_2990_end_0 = const()[name = tensor("op_2990_end_0"), val = tensor([1, 64, 1, 750])]; tensor var_2990_end_mask_0 = const()[name = tensor("op_2990_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_2990_cast_fp16 = slice_by_index(begin = var_2990_begin_0, end = var_2990_end_0, end_mask = var_2990_end_mask_0, x = var_2898_cast_fp16)[name = tensor("op_2990_cast_fp16")]; tensor var_2997_begin_0 = const()[name = tensor("op_2997_begin_0"), val = tensor([0, 0, 0, 750])]; tensor var_2997_end_0 = const()[name = tensor("op_2997_end_0"), val = tensor([1, 64, 1, 1125])]; tensor var_2997_end_mask_0 = const()[name = tensor("op_2997_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_2997_cast_fp16 = slice_by_index(begin = var_2997_begin_0, end = var_2997_end_0, end_mask = var_2997_end_mask_0, x = var_2898_cast_fp16)[name = tensor("op_2997_cast_fp16")]; tensor var_3004_begin_0 = const()[name = tensor("op_3004_begin_0"), val = tensor([0, 0, 0, 1125])]; tensor var_3004_end_0 = const()[name = tensor("op_3004_end_0"), val = tensor([1, 64, 1, 1500])]; tensor var_3004_end_mask_0 = const()[name = tensor("op_3004_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_3004_cast_fp16 = slice_by_index(begin = var_3004_begin_0, end = var_3004_end_0, end_mask = var_3004_end_mask_0, x = var_2898_cast_fp16)[name = tensor("op_3004_cast_fp16")]; tensor var_3011_begin_0 = const()[name = tensor("op_3011_begin_0"), val = tensor([0, 0, 0, 0])]; tensor var_3011_end_0 = const()[name = tensor("op_3011_end_0"), val = tensor([1, 64, 1, 375])]; tensor var_3011_end_mask_0 = const()[name = tensor("op_3011_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_3011_cast_fp16 = slice_by_index(begin = var_3011_begin_0, end = var_3011_end_0, end_mask = var_3011_end_mask_0, x = var_2902_cast_fp16)[name = tensor("op_3011_cast_fp16")]; tensor var_3018_begin_0 = const()[name = tensor("op_3018_begin_0"), val = tensor([0, 0, 0, 375])]; tensor var_3018_end_0 = const()[name = tensor("op_3018_end_0"), val = tensor([1, 64, 1, 750])]; tensor var_3018_end_mask_0 = const()[name = tensor("op_3018_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_3018_cast_fp16 = slice_by_index(begin = var_3018_begin_0, end = var_3018_end_0, end_mask = var_3018_end_mask_0, x = var_2902_cast_fp16)[name = tensor("op_3018_cast_fp16")]; tensor var_3025_begin_0 = const()[name = tensor("op_3025_begin_0"), val = tensor([0, 0, 0, 750])]; tensor var_3025_end_0 = const()[name = tensor("op_3025_end_0"), val = tensor([1, 64, 1, 1125])]; tensor var_3025_end_mask_0 = const()[name = tensor("op_3025_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_3025_cast_fp16 = slice_by_index(begin = var_3025_begin_0, end = var_3025_end_0, end_mask = var_3025_end_mask_0, x = var_2902_cast_fp16)[name = tensor("op_3025_cast_fp16")]; tensor var_3032_begin_0 = const()[name = tensor("op_3032_begin_0"), val = tensor([0, 0, 0, 1125])]; tensor var_3032_end_0 = const()[name = tensor("op_3032_end_0"), val = tensor([1, 64, 1, 1500])]; tensor var_3032_end_mask_0 = const()[name = tensor("op_3032_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_3032_cast_fp16 = slice_by_index(begin = var_3032_begin_0, end = var_3032_end_0, end_mask = var_3032_end_mask_0, x = var_2902_cast_fp16)[name = tensor("op_3032_cast_fp16")]; tensor var_3039_begin_0 = const()[name = tensor("op_3039_begin_0"), val = tensor([0, 0, 0, 0])]; tensor var_3039_end_0 = const()[name = tensor("op_3039_end_0"), val = tensor([1, 64, 1, 375])]; tensor var_3039_end_mask_0 = const()[name = tensor("op_3039_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_3039_cast_fp16 = slice_by_index(begin = var_3039_begin_0, end = var_3039_end_0, end_mask = var_3039_end_mask_0, x = var_2906_cast_fp16)[name = tensor("op_3039_cast_fp16")]; tensor var_3046_begin_0 = const()[name = tensor("op_3046_begin_0"), val = tensor([0, 0, 0, 375])]; tensor var_3046_end_0 = const()[name = tensor("op_3046_end_0"), val = tensor([1, 64, 1, 750])]; tensor var_3046_end_mask_0 = const()[name = tensor("op_3046_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_3046_cast_fp16 = slice_by_index(begin = var_3046_begin_0, end = var_3046_end_0, end_mask = var_3046_end_mask_0, x = var_2906_cast_fp16)[name = tensor("op_3046_cast_fp16")]; tensor var_3053_begin_0 = const()[name = tensor("op_3053_begin_0"), val = tensor([0, 0, 0, 750])]; tensor var_3053_end_0 = const()[name = tensor("op_3053_end_0"), val = tensor([1, 64, 1, 1125])]; tensor var_3053_end_mask_0 = const()[name = tensor("op_3053_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_3053_cast_fp16 = slice_by_index(begin = var_3053_begin_0, end = var_3053_end_0, end_mask = var_3053_end_mask_0, x = var_2906_cast_fp16)[name = tensor("op_3053_cast_fp16")]; tensor var_3060_begin_0 = const()[name = tensor("op_3060_begin_0"), val = tensor([0, 0, 0, 1125])]; tensor var_3060_end_0 = const()[name = tensor("op_3060_end_0"), val = tensor([1, 64, 1, 1500])]; tensor var_3060_end_mask_0 = const()[name = tensor("op_3060_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_3060_cast_fp16 = slice_by_index(begin = var_3060_begin_0, end = var_3060_end_0, end_mask = var_3060_end_mask_0, x = var_2906_cast_fp16)[name = tensor("op_3060_cast_fp16")]; tensor var_3067_begin_0 = const()[name = tensor("op_3067_begin_0"), val = tensor([0, 0, 0, 0])]; tensor var_3067_end_0 = const()[name = tensor("op_3067_end_0"), val = tensor([1, 64, 1, 375])]; tensor var_3067_end_mask_0 = const()[name = tensor("op_3067_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_3067_cast_fp16 = slice_by_index(begin = var_3067_begin_0, end = var_3067_end_0, end_mask = var_3067_end_mask_0, x = var_2910_cast_fp16)[name = tensor("op_3067_cast_fp16")]; tensor var_3074_begin_0 = const()[name = tensor("op_3074_begin_0"), val = tensor([0, 0, 0, 375])]; tensor var_3074_end_0 = const()[name = tensor("op_3074_end_0"), val = tensor([1, 64, 1, 750])]; tensor var_3074_end_mask_0 = const()[name = tensor("op_3074_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_3074_cast_fp16 = slice_by_index(begin = var_3074_begin_0, end = var_3074_end_0, end_mask = var_3074_end_mask_0, x = var_2910_cast_fp16)[name = tensor("op_3074_cast_fp16")]; tensor var_3081_begin_0 = const()[name = tensor("op_3081_begin_0"), val = tensor([0, 0, 0, 750])]; tensor var_3081_end_0 = const()[name = tensor("op_3081_end_0"), val = tensor([1, 64, 1, 1125])]; tensor var_3081_end_mask_0 = const()[name = tensor("op_3081_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_3081_cast_fp16 = slice_by_index(begin = var_3081_begin_0, end = var_3081_end_0, end_mask = var_3081_end_mask_0, x = var_2910_cast_fp16)[name = tensor("op_3081_cast_fp16")]; tensor var_3088_begin_0 = const()[name = tensor("op_3088_begin_0"), val = tensor([0, 0, 0, 1125])]; tensor var_3088_end_0 = const()[name = tensor("op_3088_end_0"), val = tensor([1, 64, 1, 1500])]; tensor var_3088_end_mask_0 = const()[name = tensor("op_3088_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_3088_cast_fp16 = slice_by_index(begin = var_3088_begin_0, end = var_3088_end_0, end_mask = var_3088_end_mask_0, x = var_2910_cast_fp16)[name = tensor("op_3088_cast_fp16")]; tensor var_3095_begin_0 = const()[name = tensor("op_3095_begin_0"), val = tensor([0, 0, 0, 0])]; tensor var_3095_end_0 = const()[name = tensor("op_3095_end_0"), val = tensor([1, 64, 1, 375])]; tensor var_3095_end_mask_0 = const()[name = tensor("op_3095_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_3095_cast_fp16 = slice_by_index(begin = var_3095_begin_0, end = var_3095_end_0, end_mask = var_3095_end_mask_0, x = var_2914_cast_fp16)[name = tensor("op_3095_cast_fp16")]; tensor var_3102_begin_0 = const()[name = tensor("op_3102_begin_0"), val = tensor([0, 0, 0, 375])]; tensor var_3102_end_0 = const()[name = tensor("op_3102_end_0"), val = tensor([1, 64, 1, 750])]; tensor var_3102_end_mask_0 = const()[name = tensor("op_3102_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_3102_cast_fp16 = slice_by_index(begin = var_3102_begin_0, end = var_3102_end_0, end_mask = var_3102_end_mask_0, x = var_2914_cast_fp16)[name = tensor("op_3102_cast_fp16")]; tensor var_3109_begin_0 = const()[name = tensor("op_3109_begin_0"), val = tensor([0, 0, 0, 750])]; tensor var_3109_end_0 = const()[name = tensor("op_3109_end_0"), val = tensor([1, 64, 1, 1125])]; tensor var_3109_end_mask_0 = const()[name = tensor("op_3109_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_3109_cast_fp16 = slice_by_index(begin = var_3109_begin_0, end = var_3109_end_0, end_mask = var_3109_end_mask_0, x = var_2914_cast_fp16)[name = tensor("op_3109_cast_fp16")]; tensor var_3116_begin_0 = const()[name = tensor("op_3116_begin_0"), val = tensor([0, 0, 0, 1125])]; tensor var_3116_end_0 = const()[name = tensor("op_3116_end_0"), val = tensor([1, 64, 1, 1500])]; tensor var_3116_end_mask_0 = const()[name = tensor("op_3116_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_3116_cast_fp16 = slice_by_index(begin = var_3116_begin_0, end = var_3116_end_0, end_mask = var_3116_end_mask_0, x = var_2914_cast_fp16)[name = tensor("op_3116_cast_fp16")]; tensor var_3123_begin_0 = const()[name = tensor("op_3123_begin_0"), val = tensor([0, 0, 0, 0])]; tensor var_3123_end_0 = const()[name = tensor("op_3123_end_0"), val = tensor([1, 64, 1, 375])]; tensor var_3123_end_mask_0 = const()[name = tensor("op_3123_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_3123_cast_fp16 = slice_by_index(begin = var_3123_begin_0, end = var_3123_end_0, end_mask = var_3123_end_mask_0, x = var_2918_cast_fp16)[name = tensor("op_3123_cast_fp16")]; tensor var_3130_begin_0 = const()[name = tensor("op_3130_begin_0"), val = tensor([0, 0, 0, 375])]; tensor var_3130_end_0 = const()[name = tensor("op_3130_end_0"), val = tensor([1, 64, 1, 750])]; tensor var_3130_end_mask_0 = const()[name = tensor("op_3130_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_3130_cast_fp16 = slice_by_index(begin = var_3130_begin_0, end = var_3130_end_0, end_mask = var_3130_end_mask_0, x = var_2918_cast_fp16)[name = tensor("op_3130_cast_fp16")]; tensor var_3137_begin_0 = const()[name = tensor("op_3137_begin_0"), val = tensor([0, 0, 0, 750])]; tensor var_3137_end_0 = const()[name = tensor("op_3137_end_0"), val = tensor([1, 64, 1, 1125])]; tensor var_3137_end_mask_0 = const()[name = tensor("op_3137_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_3137_cast_fp16 = slice_by_index(begin = var_3137_begin_0, end = var_3137_end_0, end_mask = var_3137_end_mask_0, x = var_2918_cast_fp16)[name = tensor("op_3137_cast_fp16")]; tensor var_3144_begin_0 = const()[name = tensor("op_3144_begin_0"), val = tensor([0, 0, 0, 1125])]; tensor var_3144_end_0 = const()[name = tensor("op_3144_end_0"), val = tensor([1, 64, 1, 1500])]; tensor var_3144_end_mask_0 = const()[name = tensor("op_3144_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_3144_cast_fp16 = slice_by_index(begin = var_3144_begin_0, end = var_3144_end_0, end_mask = var_3144_end_mask_0, x = var_2918_cast_fp16)[name = tensor("op_3144_cast_fp16")]; tensor k_9_perm_0 = const()[name = tensor("k_9_perm_0"), val = tensor([0, 3, 2, 1])]; tensor var_3149_begin_0 = const()[name = tensor("op_3149_begin_0"), val = tensor([0, 0, 0, 0])]; tensor var_3149_end_0 = const()[name = tensor("op_3149_end_0"), val = tensor([1, 1500, 1, 64])]; tensor var_3149_end_mask_0 = const()[name = tensor("op_3149_end_mask_0"), val = tensor([true, true, true, false])]; tensor transpose_1 = transpose(perm = k_9_perm_0, x = key_9_cast_fp16)[name = tensor("transpose_1")]; tensor var_3149_cast_fp16 = slice_by_index(begin = var_3149_begin_0, end = var_3149_end_0, end_mask = var_3149_end_mask_0, x = transpose_1)[name = tensor("op_3149_cast_fp16")]; tensor var_3153_begin_0 = const()[name = tensor("op_3153_begin_0"), val = tensor([0, 0, 0, 64])]; tensor var_3153_end_0 = const()[name = tensor("op_3153_end_0"), val = tensor([1, 1500, 1, 128])]; tensor var_3153_end_mask_0 = const()[name = tensor("op_3153_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_3153_cast_fp16 = slice_by_index(begin = var_3153_begin_0, end = var_3153_end_0, end_mask = var_3153_end_mask_0, x = transpose_1)[name = tensor("op_3153_cast_fp16")]; tensor var_3157_begin_0 = const()[name = tensor("op_3157_begin_0"), val = tensor([0, 0, 0, 128])]; tensor var_3157_end_0 = const()[name = tensor("op_3157_end_0"), val = tensor([1, 1500, 1, 192])]; tensor var_3157_end_mask_0 = const()[name = tensor("op_3157_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_3157_cast_fp16 = slice_by_index(begin = var_3157_begin_0, end = var_3157_end_0, end_mask = var_3157_end_mask_0, x = transpose_1)[name = tensor("op_3157_cast_fp16")]; tensor var_3161_begin_0 = const()[name = tensor("op_3161_begin_0"), val = tensor([0, 0, 0, 192])]; tensor var_3161_end_0 = const()[name = tensor("op_3161_end_0"), val = tensor([1, 1500, 1, 256])]; tensor var_3161_end_mask_0 = const()[name = tensor("op_3161_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_3161_cast_fp16 = slice_by_index(begin = var_3161_begin_0, end = var_3161_end_0, end_mask = var_3161_end_mask_0, x = transpose_1)[name = tensor("op_3161_cast_fp16")]; tensor var_3165_begin_0 = const()[name = tensor("op_3165_begin_0"), val = tensor([0, 0, 0, 256])]; tensor var_3165_end_0 = const()[name = tensor("op_3165_end_0"), val = tensor([1, 1500, 1, 320])]; tensor var_3165_end_mask_0 = const()[name = tensor("op_3165_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_3165_cast_fp16 = slice_by_index(begin = var_3165_begin_0, end = var_3165_end_0, end_mask = var_3165_end_mask_0, x = transpose_1)[name = tensor("op_3165_cast_fp16")]; tensor var_3169_begin_0 = const()[name = tensor("op_3169_begin_0"), val = tensor([0, 0, 0, 320])]; tensor var_3169_end_0 = const()[name = tensor("op_3169_end_0"), val = tensor([1, 1500, 1, 384])]; tensor var_3169_end_mask_0 = const()[name = tensor("op_3169_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_3169_cast_fp16 = slice_by_index(begin = var_3169_begin_0, end = var_3169_end_0, end_mask = var_3169_end_mask_0, x = transpose_1)[name = tensor("op_3169_cast_fp16")]; tensor var_3173_begin_0 = const()[name = tensor("op_3173_begin_0"), val = tensor([0, 0, 0, 384])]; tensor var_3173_end_0 = const()[name = tensor("op_3173_end_0"), val = tensor([1, 1500, 1, 448])]; tensor var_3173_end_mask_0 = const()[name = tensor("op_3173_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_3173_cast_fp16 = slice_by_index(begin = var_3173_begin_0, end = var_3173_end_0, end_mask = var_3173_end_mask_0, x = transpose_1)[name = tensor("op_3173_cast_fp16")]; tensor var_3177_begin_0 = const()[name = tensor("op_3177_begin_0"), val = tensor([0, 0, 0, 448])]; tensor var_3177_end_0 = const()[name = tensor("op_3177_end_0"), val = tensor([1, 1500, 1, 512])]; tensor var_3177_end_mask_0 = const()[name = tensor("op_3177_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_3177_cast_fp16 = slice_by_index(begin = var_3177_begin_0, end = var_3177_end_0, end_mask = var_3177_end_mask_0, x = transpose_1)[name = tensor("op_3177_cast_fp16")]; tensor var_3179_begin_0 = const()[name = tensor("op_3179_begin_0"), val = tensor([0, 0, 0, 0])]; tensor var_3179_end_0 = const()[name = tensor("op_3179_end_0"), val = tensor([1, 64, 1, 1500])]; tensor var_3179_end_mask_0 = const()[name = tensor("op_3179_end_mask_0"), val = tensor([true, false, true, true])]; tensor var_3179_cast_fp16 = slice_by_index(begin = var_3179_begin_0, end = var_3179_end_0, end_mask = var_3179_end_mask_0, x = value_9_cast_fp16)[name = tensor("op_3179_cast_fp16")]; tensor var_3183_begin_0 = const()[name = tensor("op_3183_begin_0"), val = tensor([0, 64, 0, 0])]; tensor var_3183_end_0 = const()[name = tensor("op_3183_end_0"), val = tensor([1, 128, 1, 1500])]; tensor var_3183_end_mask_0 = const()[name = tensor("op_3183_end_mask_0"), val = tensor([true, false, true, true])]; tensor var_3183_cast_fp16 = slice_by_index(begin = var_3183_begin_0, end = var_3183_end_0, end_mask = var_3183_end_mask_0, x = value_9_cast_fp16)[name = tensor("op_3183_cast_fp16")]; tensor var_3187_begin_0 = const()[name = tensor("op_3187_begin_0"), val = tensor([0, 128, 0, 0])]; tensor var_3187_end_0 = const()[name = tensor("op_3187_end_0"), val = tensor([1, 192, 1, 1500])]; tensor var_3187_end_mask_0 = const()[name = tensor("op_3187_end_mask_0"), val = tensor([true, false, true, true])]; tensor var_3187_cast_fp16 = slice_by_index(begin = var_3187_begin_0, end = var_3187_end_0, end_mask = var_3187_end_mask_0, x = value_9_cast_fp16)[name = tensor("op_3187_cast_fp16")]; tensor var_3191_begin_0 = const()[name = tensor("op_3191_begin_0"), val = tensor([0, 192, 0, 0])]; tensor var_3191_end_0 = const()[name = tensor("op_3191_end_0"), val = tensor([1, 256, 1, 1500])]; tensor var_3191_end_mask_0 = const()[name = tensor("op_3191_end_mask_0"), val = tensor([true, false, true, true])]; tensor var_3191_cast_fp16 = slice_by_index(begin = var_3191_begin_0, end = var_3191_end_0, end_mask = var_3191_end_mask_0, x = value_9_cast_fp16)[name = tensor("op_3191_cast_fp16")]; tensor var_3195_begin_0 = const()[name = tensor("op_3195_begin_0"), val = tensor([0, 256, 0, 0])]; tensor var_3195_end_0 = const()[name = tensor("op_3195_end_0"), val = tensor([1, 320, 1, 1500])]; tensor var_3195_end_mask_0 = const()[name = tensor("op_3195_end_mask_0"), val = tensor([true, false, true, true])]; tensor var_3195_cast_fp16 = slice_by_index(begin = var_3195_begin_0, end = var_3195_end_0, end_mask = var_3195_end_mask_0, x = value_9_cast_fp16)[name = tensor("op_3195_cast_fp16")]; tensor var_3199_begin_0 = const()[name = tensor("op_3199_begin_0"), val = tensor([0, 320, 0, 0])]; tensor var_3199_end_0 = const()[name = tensor("op_3199_end_0"), val = tensor([1, 384, 1, 1500])]; tensor var_3199_end_mask_0 = const()[name = tensor("op_3199_end_mask_0"), val = tensor([true, false, true, true])]; tensor var_3199_cast_fp16 = slice_by_index(begin = var_3199_begin_0, end = var_3199_end_0, end_mask = var_3199_end_mask_0, x = value_9_cast_fp16)[name = tensor("op_3199_cast_fp16")]; tensor var_3203_begin_0 = const()[name = tensor("op_3203_begin_0"), val = tensor([0, 384, 0, 0])]; tensor var_3203_end_0 = const()[name = tensor("op_3203_end_0"), val = tensor([1, 448, 1, 1500])]; tensor var_3203_end_mask_0 = const()[name = tensor("op_3203_end_mask_0"), val = tensor([true, false, true, true])]; tensor var_3203_cast_fp16 = slice_by_index(begin = var_3203_begin_0, end = var_3203_end_0, end_mask = var_3203_end_mask_0, x = value_9_cast_fp16)[name = tensor("op_3203_cast_fp16")]; tensor var_3207_begin_0 = const()[name = tensor("op_3207_begin_0"), val = tensor([0, 448, 0, 0])]; tensor var_3207_end_0 = const()[name = tensor("op_3207_end_0"), val = tensor([1, 512, 1, 1500])]; tensor var_3207_end_mask_0 = const()[name = tensor("op_3207_end_mask_0"), val = tensor([true, false, true, true])]; tensor var_3207_cast_fp16 = slice_by_index(begin = var_3207_begin_0, end = var_3207_end_0, end_mask = var_3207_end_mask_0, x = value_9_cast_fp16)[name = tensor("op_3207_cast_fp16")]; tensor var_3211_equation_0 = const()[name = tensor("op_3211_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_3211_cast_fp16 = einsum(equation = var_3211_equation_0, values = (var_3149_cast_fp16, var_2927_cast_fp16))[name = tensor("op_3211_cast_fp16")]; tensor var_3212_to_fp16 = const()[name = tensor("op_3212_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_257_cast_fp16 = mul(x = var_3211_cast_fp16, y = var_3212_to_fp16)[name = tensor("aw_chunk_257_cast_fp16")]; tensor var_3215_equation_0 = const()[name = tensor("op_3215_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_3215_cast_fp16 = einsum(equation = var_3215_equation_0, values = (var_3149_cast_fp16, var_2934_cast_fp16))[name = tensor("op_3215_cast_fp16")]; tensor var_3216_to_fp16 = const()[name = tensor("op_3216_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_259_cast_fp16 = mul(x = var_3215_cast_fp16, y = var_3216_to_fp16)[name = tensor("aw_chunk_259_cast_fp16")]; tensor var_3219_equation_0 = const()[name = tensor("op_3219_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_3219_cast_fp16 = einsum(equation = var_3219_equation_0, values = (var_3149_cast_fp16, var_2941_cast_fp16))[name = tensor("op_3219_cast_fp16")]; tensor var_3220_to_fp16 = const()[name = tensor("op_3220_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_261_cast_fp16 = mul(x = var_3219_cast_fp16, y = var_3220_to_fp16)[name = tensor("aw_chunk_261_cast_fp16")]; tensor var_3223_equation_0 = const()[name = tensor("op_3223_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_3223_cast_fp16 = einsum(equation = var_3223_equation_0, values = (var_3149_cast_fp16, var_2948_cast_fp16))[name = tensor("op_3223_cast_fp16")]; tensor var_3224_to_fp16 = const()[name = tensor("op_3224_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_263_cast_fp16 = mul(x = var_3223_cast_fp16, y = var_3224_to_fp16)[name = tensor("aw_chunk_263_cast_fp16")]; tensor var_3227_equation_0 = const()[name = tensor("op_3227_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_3227_cast_fp16 = einsum(equation = var_3227_equation_0, values = (var_3153_cast_fp16, var_2955_cast_fp16))[name = tensor("op_3227_cast_fp16")]; tensor var_3228_to_fp16 = const()[name = tensor("op_3228_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_265_cast_fp16 = mul(x = var_3227_cast_fp16, y = var_3228_to_fp16)[name = tensor("aw_chunk_265_cast_fp16")]; tensor var_3231_equation_0 = const()[name = tensor("op_3231_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_3231_cast_fp16 = einsum(equation = var_3231_equation_0, values = (var_3153_cast_fp16, var_2962_cast_fp16))[name = tensor("op_3231_cast_fp16")]; tensor var_3232_to_fp16 = const()[name = tensor("op_3232_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_267_cast_fp16 = mul(x = var_3231_cast_fp16, y = var_3232_to_fp16)[name = tensor("aw_chunk_267_cast_fp16")]; tensor var_3235_equation_0 = const()[name = tensor("op_3235_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_3235_cast_fp16 = einsum(equation = var_3235_equation_0, values = (var_3153_cast_fp16, var_2969_cast_fp16))[name = tensor("op_3235_cast_fp16")]; tensor var_3236_to_fp16 = const()[name = tensor("op_3236_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_269_cast_fp16 = mul(x = var_3235_cast_fp16, y = var_3236_to_fp16)[name = tensor("aw_chunk_269_cast_fp16")]; tensor var_3239_equation_0 = const()[name = tensor("op_3239_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_3239_cast_fp16 = einsum(equation = var_3239_equation_0, values = (var_3153_cast_fp16, var_2976_cast_fp16))[name = tensor("op_3239_cast_fp16")]; tensor var_3240_to_fp16 = const()[name = tensor("op_3240_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_271_cast_fp16 = mul(x = var_3239_cast_fp16, y = var_3240_to_fp16)[name = tensor("aw_chunk_271_cast_fp16")]; tensor var_3243_equation_0 = const()[name = tensor("op_3243_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_3243_cast_fp16 = einsum(equation = var_3243_equation_0, values = (var_3157_cast_fp16, var_2983_cast_fp16))[name = tensor("op_3243_cast_fp16")]; tensor var_3244_to_fp16 = const()[name = tensor("op_3244_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_273_cast_fp16 = mul(x = var_3243_cast_fp16, y = var_3244_to_fp16)[name = tensor("aw_chunk_273_cast_fp16")]; tensor var_3247_equation_0 = const()[name = tensor("op_3247_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_3247_cast_fp16 = einsum(equation = var_3247_equation_0, values = (var_3157_cast_fp16, var_2990_cast_fp16))[name = tensor("op_3247_cast_fp16")]; tensor var_3248_to_fp16 = const()[name = tensor("op_3248_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_275_cast_fp16 = mul(x = var_3247_cast_fp16, y = var_3248_to_fp16)[name = tensor("aw_chunk_275_cast_fp16")]; tensor var_3251_equation_0 = const()[name = tensor("op_3251_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_3251_cast_fp16 = einsum(equation = var_3251_equation_0, values = (var_3157_cast_fp16, var_2997_cast_fp16))[name = tensor("op_3251_cast_fp16")]; tensor var_3252_to_fp16 = const()[name = tensor("op_3252_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_277_cast_fp16 = mul(x = var_3251_cast_fp16, y = var_3252_to_fp16)[name = tensor("aw_chunk_277_cast_fp16")]; tensor var_3255_equation_0 = const()[name = tensor("op_3255_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_3255_cast_fp16 = einsum(equation = var_3255_equation_0, values = (var_3157_cast_fp16, var_3004_cast_fp16))[name = tensor("op_3255_cast_fp16")]; tensor var_3256_to_fp16 = const()[name = tensor("op_3256_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_279_cast_fp16 = mul(x = var_3255_cast_fp16, y = var_3256_to_fp16)[name = tensor("aw_chunk_279_cast_fp16")]; tensor var_3259_equation_0 = const()[name = tensor("op_3259_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_3259_cast_fp16 = einsum(equation = var_3259_equation_0, values = (var_3161_cast_fp16, var_3011_cast_fp16))[name = tensor("op_3259_cast_fp16")]; tensor var_3260_to_fp16 = const()[name = tensor("op_3260_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_281_cast_fp16 = mul(x = var_3259_cast_fp16, y = var_3260_to_fp16)[name = tensor("aw_chunk_281_cast_fp16")]; tensor var_3263_equation_0 = const()[name = tensor("op_3263_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_3263_cast_fp16 = einsum(equation = var_3263_equation_0, values = (var_3161_cast_fp16, var_3018_cast_fp16))[name = tensor("op_3263_cast_fp16")]; tensor var_3264_to_fp16 = const()[name = tensor("op_3264_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_283_cast_fp16 = mul(x = var_3263_cast_fp16, y = var_3264_to_fp16)[name = tensor("aw_chunk_283_cast_fp16")]; tensor var_3267_equation_0 = const()[name = tensor("op_3267_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_3267_cast_fp16 = einsum(equation = var_3267_equation_0, values = (var_3161_cast_fp16, var_3025_cast_fp16))[name = tensor("op_3267_cast_fp16")]; tensor var_3268_to_fp16 = const()[name = tensor("op_3268_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_285_cast_fp16 = mul(x = var_3267_cast_fp16, y = var_3268_to_fp16)[name = tensor("aw_chunk_285_cast_fp16")]; tensor var_3271_equation_0 = const()[name = tensor("op_3271_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_3271_cast_fp16 = einsum(equation = var_3271_equation_0, values = (var_3161_cast_fp16, var_3032_cast_fp16))[name = tensor("op_3271_cast_fp16")]; tensor var_3272_to_fp16 = const()[name = tensor("op_3272_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_287_cast_fp16 = mul(x = var_3271_cast_fp16, y = var_3272_to_fp16)[name = tensor("aw_chunk_287_cast_fp16")]; tensor var_3275_equation_0 = const()[name = tensor("op_3275_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_3275_cast_fp16 = einsum(equation = var_3275_equation_0, values = (var_3165_cast_fp16, var_3039_cast_fp16))[name = tensor("op_3275_cast_fp16")]; tensor var_3276_to_fp16 = const()[name = tensor("op_3276_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_289_cast_fp16 = mul(x = var_3275_cast_fp16, y = var_3276_to_fp16)[name = tensor("aw_chunk_289_cast_fp16")]; tensor var_3279_equation_0 = const()[name = tensor("op_3279_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_3279_cast_fp16 = einsum(equation = var_3279_equation_0, values = (var_3165_cast_fp16, var_3046_cast_fp16))[name = tensor("op_3279_cast_fp16")]; tensor var_3280_to_fp16 = const()[name = tensor("op_3280_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_291_cast_fp16 = mul(x = var_3279_cast_fp16, y = var_3280_to_fp16)[name = tensor("aw_chunk_291_cast_fp16")]; tensor var_3283_equation_0 = const()[name = tensor("op_3283_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_3283_cast_fp16 = einsum(equation = var_3283_equation_0, values = (var_3165_cast_fp16, var_3053_cast_fp16))[name = tensor("op_3283_cast_fp16")]; tensor var_3284_to_fp16 = const()[name = tensor("op_3284_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_293_cast_fp16 = mul(x = var_3283_cast_fp16, y = var_3284_to_fp16)[name = tensor("aw_chunk_293_cast_fp16")]; tensor var_3287_equation_0 = const()[name = tensor("op_3287_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_3287_cast_fp16 = einsum(equation = var_3287_equation_0, values = (var_3165_cast_fp16, var_3060_cast_fp16))[name = tensor("op_3287_cast_fp16")]; tensor var_3288_to_fp16 = const()[name = tensor("op_3288_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_295_cast_fp16 = mul(x = var_3287_cast_fp16, y = var_3288_to_fp16)[name = tensor("aw_chunk_295_cast_fp16")]; tensor var_3291_equation_0 = const()[name = tensor("op_3291_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_3291_cast_fp16 = einsum(equation = var_3291_equation_0, values = (var_3169_cast_fp16, var_3067_cast_fp16))[name = tensor("op_3291_cast_fp16")]; tensor var_3292_to_fp16 = const()[name = tensor("op_3292_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_297_cast_fp16 = mul(x = var_3291_cast_fp16, y = var_3292_to_fp16)[name = tensor("aw_chunk_297_cast_fp16")]; tensor var_3295_equation_0 = const()[name = tensor("op_3295_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_3295_cast_fp16 = einsum(equation = var_3295_equation_0, values = (var_3169_cast_fp16, var_3074_cast_fp16))[name = tensor("op_3295_cast_fp16")]; tensor var_3296_to_fp16 = const()[name = tensor("op_3296_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_299_cast_fp16 = mul(x = var_3295_cast_fp16, y = var_3296_to_fp16)[name = tensor("aw_chunk_299_cast_fp16")]; tensor var_3299_equation_0 = const()[name = tensor("op_3299_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_3299_cast_fp16 = einsum(equation = var_3299_equation_0, values = (var_3169_cast_fp16, var_3081_cast_fp16))[name = tensor("op_3299_cast_fp16")]; tensor var_3300_to_fp16 = const()[name = tensor("op_3300_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_301_cast_fp16 = mul(x = var_3299_cast_fp16, y = var_3300_to_fp16)[name = tensor("aw_chunk_301_cast_fp16")]; tensor var_3303_equation_0 = const()[name = tensor("op_3303_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_3303_cast_fp16 = einsum(equation = var_3303_equation_0, values = (var_3169_cast_fp16, var_3088_cast_fp16))[name = tensor("op_3303_cast_fp16")]; tensor var_3304_to_fp16 = const()[name = tensor("op_3304_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_303_cast_fp16 = mul(x = var_3303_cast_fp16, y = var_3304_to_fp16)[name = tensor("aw_chunk_303_cast_fp16")]; tensor var_3307_equation_0 = const()[name = tensor("op_3307_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_3307_cast_fp16 = einsum(equation = var_3307_equation_0, values = (var_3173_cast_fp16, var_3095_cast_fp16))[name = tensor("op_3307_cast_fp16")]; tensor var_3308_to_fp16 = const()[name = tensor("op_3308_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_305_cast_fp16 = mul(x = var_3307_cast_fp16, y = var_3308_to_fp16)[name = tensor("aw_chunk_305_cast_fp16")]; tensor var_3311_equation_0 = const()[name = tensor("op_3311_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_3311_cast_fp16 = einsum(equation = var_3311_equation_0, values = (var_3173_cast_fp16, var_3102_cast_fp16))[name = tensor("op_3311_cast_fp16")]; tensor var_3312_to_fp16 = const()[name = tensor("op_3312_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_307_cast_fp16 = mul(x = var_3311_cast_fp16, y = var_3312_to_fp16)[name = tensor("aw_chunk_307_cast_fp16")]; tensor var_3315_equation_0 = const()[name = tensor("op_3315_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_3315_cast_fp16 = einsum(equation = var_3315_equation_0, values = (var_3173_cast_fp16, var_3109_cast_fp16))[name = tensor("op_3315_cast_fp16")]; tensor var_3316_to_fp16 = const()[name = tensor("op_3316_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_309_cast_fp16 = mul(x = var_3315_cast_fp16, y = var_3316_to_fp16)[name = tensor("aw_chunk_309_cast_fp16")]; tensor var_3319_equation_0 = const()[name = tensor("op_3319_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_3319_cast_fp16 = einsum(equation = var_3319_equation_0, values = (var_3173_cast_fp16, var_3116_cast_fp16))[name = tensor("op_3319_cast_fp16")]; tensor var_3320_to_fp16 = const()[name = tensor("op_3320_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_311_cast_fp16 = mul(x = var_3319_cast_fp16, y = var_3320_to_fp16)[name = tensor("aw_chunk_311_cast_fp16")]; tensor var_3323_equation_0 = const()[name = tensor("op_3323_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_3323_cast_fp16 = einsum(equation = var_3323_equation_0, values = (var_3177_cast_fp16, var_3123_cast_fp16))[name = tensor("op_3323_cast_fp16")]; tensor var_3324_to_fp16 = const()[name = tensor("op_3324_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_313_cast_fp16 = mul(x = var_3323_cast_fp16, y = var_3324_to_fp16)[name = tensor("aw_chunk_313_cast_fp16")]; tensor var_3327_equation_0 = const()[name = tensor("op_3327_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_3327_cast_fp16 = einsum(equation = var_3327_equation_0, values = (var_3177_cast_fp16, var_3130_cast_fp16))[name = tensor("op_3327_cast_fp16")]; tensor var_3328_to_fp16 = const()[name = tensor("op_3328_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_315_cast_fp16 = mul(x = var_3327_cast_fp16, y = var_3328_to_fp16)[name = tensor("aw_chunk_315_cast_fp16")]; tensor var_3331_equation_0 = const()[name = tensor("op_3331_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_3331_cast_fp16 = einsum(equation = var_3331_equation_0, values = (var_3177_cast_fp16, var_3137_cast_fp16))[name = tensor("op_3331_cast_fp16")]; tensor var_3332_to_fp16 = const()[name = tensor("op_3332_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_317_cast_fp16 = mul(x = var_3331_cast_fp16, y = var_3332_to_fp16)[name = tensor("aw_chunk_317_cast_fp16")]; tensor var_3335_equation_0 = const()[name = tensor("op_3335_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_3335_cast_fp16 = einsum(equation = var_3335_equation_0, values = (var_3177_cast_fp16, var_3144_cast_fp16))[name = tensor("op_3335_cast_fp16")]; tensor var_3336_to_fp16 = const()[name = tensor("op_3336_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_319_cast_fp16 = mul(x = var_3335_cast_fp16, y = var_3336_to_fp16)[name = tensor("aw_chunk_319_cast_fp16")]; tensor var_3338_cast_fp16 = softmax(axis = var_2835, x = aw_chunk_257_cast_fp16)[name = tensor("op_3338_cast_fp16")]; tensor var_3339_cast_fp16 = softmax(axis = var_2835, x = aw_chunk_259_cast_fp16)[name = tensor("op_3339_cast_fp16")]; tensor var_3340_cast_fp16 = softmax(axis = var_2835, x = aw_chunk_261_cast_fp16)[name = tensor("op_3340_cast_fp16")]; tensor var_3341_cast_fp16 = softmax(axis = var_2835, x = aw_chunk_263_cast_fp16)[name = tensor("op_3341_cast_fp16")]; tensor var_3342_cast_fp16 = softmax(axis = var_2835, x = aw_chunk_265_cast_fp16)[name = tensor("op_3342_cast_fp16")]; tensor var_3343_cast_fp16 = softmax(axis = var_2835, x = aw_chunk_267_cast_fp16)[name = tensor("op_3343_cast_fp16")]; tensor var_3344_cast_fp16 = softmax(axis = var_2835, x = aw_chunk_269_cast_fp16)[name = tensor("op_3344_cast_fp16")]; tensor var_3345_cast_fp16 = softmax(axis = var_2835, x = aw_chunk_271_cast_fp16)[name = tensor("op_3345_cast_fp16")]; tensor var_3346_cast_fp16 = softmax(axis = var_2835, x = aw_chunk_273_cast_fp16)[name = tensor("op_3346_cast_fp16")]; tensor var_3347_cast_fp16 = softmax(axis = var_2835, x = aw_chunk_275_cast_fp16)[name = tensor("op_3347_cast_fp16")]; tensor var_3348_cast_fp16 = softmax(axis = var_2835, x = aw_chunk_277_cast_fp16)[name = tensor("op_3348_cast_fp16")]; tensor var_3349_cast_fp16 = softmax(axis = var_2835, x = aw_chunk_279_cast_fp16)[name = tensor("op_3349_cast_fp16")]; tensor var_3350_cast_fp16 = softmax(axis = var_2835, x = aw_chunk_281_cast_fp16)[name = tensor("op_3350_cast_fp16")]; tensor var_3351_cast_fp16 = softmax(axis = var_2835, x = aw_chunk_283_cast_fp16)[name = tensor("op_3351_cast_fp16")]; tensor var_3352_cast_fp16 = softmax(axis = var_2835, x = aw_chunk_285_cast_fp16)[name = tensor("op_3352_cast_fp16")]; tensor var_3353_cast_fp16 = softmax(axis = var_2835, x = aw_chunk_287_cast_fp16)[name = tensor("op_3353_cast_fp16")]; tensor var_3354_cast_fp16 = softmax(axis = var_2835, x = aw_chunk_289_cast_fp16)[name = tensor("op_3354_cast_fp16")]; tensor var_3355_cast_fp16 = softmax(axis = var_2835, x = aw_chunk_291_cast_fp16)[name = tensor("op_3355_cast_fp16")]; tensor var_3356_cast_fp16 = softmax(axis = var_2835, x = aw_chunk_293_cast_fp16)[name = tensor("op_3356_cast_fp16")]; tensor var_3357_cast_fp16 = softmax(axis = var_2835, x = aw_chunk_295_cast_fp16)[name = tensor("op_3357_cast_fp16")]; tensor var_3358_cast_fp16 = softmax(axis = var_2835, x = aw_chunk_297_cast_fp16)[name = tensor("op_3358_cast_fp16")]; tensor var_3359_cast_fp16 = softmax(axis = var_2835, x = aw_chunk_299_cast_fp16)[name = tensor("op_3359_cast_fp16")]; tensor var_3360_cast_fp16 = softmax(axis = var_2835, x = aw_chunk_301_cast_fp16)[name = tensor("op_3360_cast_fp16")]; tensor var_3361_cast_fp16 = softmax(axis = var_2835, x = aw_chunk_303_cast_fp16)[name = tensor("op_3361_cast_fp16")]; tensor var_3362_cast_fp16 = softmax(axis = var_2835, x = aw_chunk_305_cast_fp16)[name = tensor("op_3362_cast_fp16")]; tensor var_3363_cast_fp16 = softmax(axis = var_2835, x = aw_chunk_307_cast_fp16)[name = tensor("op_3363_cast_fp16")]; tensor var_3364_cast_fp16 = softmax(axis = var_2835, x = aw_chunk_309_cast_fp16)[name = tensor("op_3364_cast_fp16")]; tensor var_3365_cast_fp16 = softmax(axis = var_2835, x = aw_chunk_311_cast_fp16)[name = tensor("op_3365_cast_fp16")]; tensor var_3366_cast_fp16 = softmax(axis = var_2835, x = aw_chunk_313_cast_fp16)[name = tensor("op_3366_cast_fp16")]; tensor var_3367_cast_fp16 = softmax(axis = var_2835, x = aw_chunk_315_cast_fp16)[name = tensor("op_3367_cast_fp16")]; tensor var_3368_cast_fp16 = softmax(axis = var_2835, x = aw_chunk_317_cast_fp16)[name = tensor("op_3368_cast_fp16")]; tensor var_3369_cast_fp16 = softmax(axis = var_2835, x = aw_chunk_319_cast_fp16)[name = tensor("op_3369_cast_fp16")]; tensor var_3371_equation_0 = const()[name = tensor("op_3371_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_3371_cast_fp16 = einsum(equation = var_3371_equation_0, values = (var_3179_cast_fp16, var_3338_cast_fp16))[name = tensor("op_3371_cast_fp16")]; tensor var_3373_equation_0 = const()[name = tensor("op_3373_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_3373_cast_fp16 = einsum(equation = var_3373_equation_0, values = (var_3179_cast_fp16, var_3339_cast_fp16))[name = tensor("op_3373_cast_fp16")]; tensor var_3375_equation_0 = const()[name = tensor("op_3375_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_3375_cast_fp16 = einsum(equation = var_3375_equation_0, values = (var_3179_cast_fp16, var_3340_cast_fp16))[name = tensor("op_3375_cast_fp16")]; tensor var_3377_equation_0 = const()[name = tensor("op_3377_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_3377_cast_fp16 = einsum(equation = var_3377_equation_0, values = (var_3179_cast_fp16, var_3341_cast_fp16))[name = tensor("op_3377_cast_fp16")]; tensor var_3379_equation_0 = const()[name = tensor("op_3379_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_3379_cast_fp16 = einsum(equation = var_3379_equation_0, values = (var_3183_cast_fp16, var_3342_cast_fp16))[name = tensor("op_3379_cast_fp16")]; tensor var_3381_equation_0 = const()[name = tensor("op_3381_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_3381_cast_fp16 = einsum(equation = var_3381_equation_0, values = (var_3183_cast_fp16, var_3343_cast_fp16))[name = tensor("op_3381_cast_fp16")]; tensor var_3383_equation_0 = const()[name = tensor("op_3383_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_3383_cast_fp16 = einsum(equation = var_3383_equation_0, values = (var_3183_cast_fp16, var_3344_cast_fp16))[name = tensor("op_3383_cast_fp16")]; tensor var_3385_equation_0 = const()[name = tensor("op_3385_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_3385_cast_fp16 = einsum(equation = var_3385_equation_0, values = (var_3183_cast_fp16, var_3345_cast_fp16))[name = tensor("op_3385_cast_fp16")]; tensor var_3387_equation_0 = const()[name = tensor("op_3387_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_3387_cast_fp16 = einsum(equation = var_3387_equation_0, values = (var_3187_cast_fp16, var_3346_cast_fp16))[name = tensor("op_3387_cast_fp16")]; tensor var_3389_equation_0 = const()[name = tensor("op_3389_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_3389_cast_fp16 = einsum(equation = var_3389_equation_0, values = (var_3187_cast_fp16, var_3347_cast_fp16))[name = tensor("op_3389_cast_fp16")]; tensor var_3391_equation_0 = const()[name = tensor("op_3391_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_3391_cast_fp16 = einsum(equation = var_3391_equation_0, values = (var_3187_cast_fp16, var_3348_cast_fp16))[name = tensor("op_3391_cast_fp16")]; tensor var_3393_equation_0 = const()[name = tensor("op_3393_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_3393_cast_fp16 = einsum(equation = var_3393_equation_0, values = (var_3187_cast_fp16, var_3349_cast_fp16))[name = tensor("op_3393_cast_fp16")]; tensor var_3395_equation_0 = const()[name = tensor("op_3395_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_3395_cast_fp16 = einsum(equation = var_3395_equation_0, values = (var_3191_cast_fp16, var_3350_cast_fp16))[name = tensor("op_3395_cast_fp16")]; tensor var_3397_equation_0 = const()[name = tensor("op_3397_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_3397_cast_fp16 = einsum(equation = var_3397_equation_0, values = (var_3191_cast_fp16, var_3351_cast_fp16))[name = tensor("op_3397_cast_fp16")]; tensor var_3399_equation_0 = const()[name = tensor("op_3399_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_3399_cast_fp16 = einsum(equation = var_3399_equation_0, values = (var_3191_cast_fp16, var_3352_cast_fp16))[name = tensor("op_3399_cast_fp16")]; tensor var_3401_equation_0 = const()[name = tensor("op_3401_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_3401_cast_fp16 = einsum(equation = var_3401_equation_0, values = (var_3191_cast_fp16, var_3353_cast_fp16))[name = tensor("op_3401_cast_fp16")]; tensor var_3403_equation_0 = const()[name = tensor("op_3403_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_3403_cast_fp16 = einsum(equation = var_3403_equation_0, values = (var_3195_cast_fp16, var_3354_cast_fp16))[name = tensor("op_3403_cast_fp16")]; tensor var_3405_equation_0 = const()[name = tensor("op_3405_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_3405_cast_fp16 = einsum(equation = var_3405_equation_0, values = (var_3195_cast_fp16, var_3355_cast_fp16))[name = tensor("op_3405_cast_fp16")]; tensor var_3407_equation_0 = const()[name = tensor("op_3407_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_3407_cast_fp16 = einsum(equation = var_3407_equation_0, values = (var_3195_cast_fp16, var_3356_cast_fp16))[name = tensor("op_3407_cast_fp16")]; tensor var_3409_equation_0 = const()[name = tensor("op_3409_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_3409_cast_fp16 = einsum(equation = var_3409_equation_0, values = (var_3195_cast_fp16, var_3357_cast_fp16))[name = tensor("op_3409_cast_fp16")]; tensor var_3411_equation_0 = const()[name = tensor("op_3411_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_3411_cast_fp16 = einsum(equation = var_3411_equation_0, values = (var_3199_cast_fp16, var_3358_cast_fp16))[name = tensor("op_3411_cast_fp16")]; tensor var_3413_equation_0 = const()[name = tensor("op_3413_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_3413_cast_fp16 = einsum(equation = var_3413_equation_0, values = (var_3199_cast_fp16, var_3359_cast_fp16))[name = tensor("op_3413_cast_fp16")]; tensor var_3415_equation_0 = const()[name = tensor("op_3415_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_3415_cast_fp16 = einsum(equation = var_3415_equation_0, values = (var_3199_cast_fp16, var_3360_cast_fp16))[name = tensor("op_3415_cast_fp16")]; tensor var_3417_equation_0 = const()[name = tensor("op_3417_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_3417_cast_fp16 = einsum(equation = var_3417_equation_0, values = (var_3199_cast_fp16, var_3361_cast_fp16))[name = tensor("op_3417_cast_fp16")]; tensor var_3419_equation_0 = const()[name = tensor("op_3419_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_3419_cast_fp16 = einsum(equation = var_3419_equation_0, values = (var_3203_cast_fp16, var_3362_cast_fp16))[name = tensor("op_3419_cast_fp16")]; tensor var_3421_equation_0 = const()[name = tensor("op_3421_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_3421_cast_fp16 = einsum(equation = var_3421_equation_0, values = (var_3203_cast_fp16, var_3363_cast_fp16))[name = tensor("op_3421_cast_fp16")]; tensor var_3423_equation_0 = const()[name = tensor("op_3423_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_3423_cast_fp16 = einsum(equation = var_3423_equation_0, values = (var_3203_cast_fp16, var_3364_cast_fp16))[name = tensor("op_3423_cast_fp16")]; tensor var_3425_equation_0 = const()[name = tensor("op_3425_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_3425_cast_fp16 = einsum(equation = var_3425_equation_0, values = (var_3203_cast_fp16, var_3365_cast_fp16))[name = tensor("op_3425_cast_fp16")]; tensor var_3427_equation_0 = const()[name = tensor("op_3427_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_3427_cast_fp16 = einsum(equation = var_3427_equation_0, values = (var_3207_cast_fp16, var_3366_cast_fp16))[name = tensor("op_3427_cast_fp16")]; tensor var_3429_equation_0 = const()[name = tensor("op_3429_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_3429_cast_fp16 = einsum(equation = var_3429_equation_0, values = (var_3207_cast_fp16, var_3367_cast_fp16))[name = tensor("op_3429_cast_fp16")]; tensor var_3431_equation_0 = const()[name = tensor("op_3431_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_3431_cast_fp16 = einsum(equation = var_3431_equation_0, values = (var_3207_cast_fp16, var_3368_cast_fp16))[name = tensor("op_3431_cast_fp16")]; tensor var_3433_equation_0 = const()[name = tensor("op_3433_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_3433_cast_fp16 = einsum(equation = var_3433_equation_0, values = (var_3207_cast_fp16, var_3369_cast_fp16))[name = tensor("op_3433_cast_fp16")]; tensor var_3435_interleave_0 = const()[name = tensor("op_3435_interleave_0"), val = tensor(false)]; tensor var_3435_cast_fp16 = concat(axis = var_2822, interleave = var_3435_interleave_0, values = (var_3371_cast_fp16, var_3373_cast_fp16, var_3375_cast_fp16, var_3377_cast_fp16))[name = tensor("op_3435_cast_fp16")]; tensor var_3437_interleave_0 = const()[name = tensor("op_3437_interleave_0"), val = tensor(false)]; tensor var_3437_cast_fp16 = concat(axis = var_2822, interleave = var_3437_interleave_0, values = (var_3379_cast_fp16, var_3381_cast_fp16, var_3383_cast_fp16, var_3385_cast_fp16))[name = tensor("op_3437_cast_fp16")]; tensor var_3439_interleave_0 = const()[name = tensor("op_3439_interleave_0"), val = tensor(false)]; tensor var_3439_cast_fp16 = concat(axis = var_2822, interleave = var_3439_interleave_0, values = (var_3387_cast_fp16, var_3389_cast_fp16, var_3391_cast_fp16, var_3393_cast_fp16))[name = tensor("op_3439_cast_fp16")]; tensor var_3441_interleave_0 = const()[name = tensor("op_3441_interleave_0"), val = tensor(false)]; tensor var_3441_cast_fp16 = concat(axis = var_2822, interleave = var_3441_interleave_0, values = (var_3395_cast_fp16, var_3397_cast_fp16, var_3399_cast_fp16, var_3401_cast_fp16))[name = tensor("op_3441_cast_fp16")]; tensor var_3443_interleave_0 = const()[name = tensor("op_3443_interleave_0"), val = tensor(false)]; tensor var_3443_cast_fp16 = concat(axis = var_2822, interleave = var_3443_interleave_0, values = (var_3403_cast_fp16, var_3405_cast_fp16, var_3407_cast_fp16, var_3409_cast_fp16))[name = tensor("op_3443_cast_fp16")]; tensor var_3445_interleave_0 = const()[name = tensor("op_3445_interleave_0"), val = tensor(false)]; tensor var_3445_cast_fp16 = concat(axis = var_2822, interleave = var_3445_interleave_0, values = (var_3411_cast_fp16, var_3413_cast_fp16, var_3415_cast_fp16, var_3417_cast_fp16))[name = tensor("op_3445_cast_fp16")]; tensor var_3447_interleave_0 = const()[name = tensor("op_3447_interleave_0"), val = tensor(false)]; tensor var_3447_cast_fp16 = concat(axis = var_2822, interleave = var_3447_interleave_0, values = (var_3419_cast_fp16, var_3421_cast_fp16, var_3423_cast_fp16, var_3425_cast_fp16))[name = tensor("op_3447_cast_fp16")]; tensor var_3449_interleave_0 = const()[name = tensor("op_3449_interleave_0"), val = tensor(false)]; tensor var_3449_cast_fp16 = concat(axis = var_2822, interleave = var_3449_interleave_0, values = (var_3427_cast_fp16, var_3429_cast_fp16, var_3431_cast_fp16, var_3433_cast_fp16))[name = tensor("op_3449_cast_fp16")]; tensor input_33_interleave_0 = const()[name = tensor("input_33_interleave_0"), val = tensor(false)]; tensor input_33_cast_fp16 = concat(axis = var_2835, interleave = input_33_interleave_0, values = (var_3435_cast_fp16, var_3437_cast_fp16, var_3439_cast_fp16, var_3441_cast_fp16, var_3443_cast_fp16, var_3445_cast_fp16, var_3447_cast_fp16, var_3449_cast_fp16))[name = tensor("input_33_cast_fp16")]; tensor var_3454 = const()[name = tensor("op_3454"), val = tensor([1, 1])]; tensor var_3456 = const()[name = tensor("op_3456"), val = tensor([1, 1])]; tensor obj_19_pad_type_0 = const()[name = tensor("obj_19_pad_type_0"), val = tensor("custom")]; tensor obj_19_pad_0 = const()[name = tensor("obj_19_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_4_self_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_4_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30155456)))]; tensor layers_4_self_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_4_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30679808)))]; tensor obj_19_cast_fp16 = conv(bias = layers_4_self_attn_o_proj_bias_to_fp16, dilations = var_3456, groups = var_2835, pad = obj_19_pad_0, pad_type = obj_19_pad_type_0, strides = var_3454, weight = layers_4_self_attn_o_proj_weight_to_fp16, x = input_33_cast_fp16)[name = tensor("obj_19_cast_fp16")]; tensor inputs_19_cast_fp16 = add(x = inputs_17_cast_fp16, y = obj_19_cast_fp16)[name = tensor("inputs_19_cast_fp16")]; tensor var_3462 = const()[name = tensor("op_3462"), val = tensor([1])]; tensor channels_mean_19_cast_fp16 = reduce_mean(axes = var_3462, keep_dims = var_2836, x = inputs_19_cast_fp16)[name = tensor("channels_mean_19_cast_fp16")]; tensor zero_mean_19_cast_fp16 = sub(x = inputs_19_cast_fp16, y = channels_mean_19_cast_fp16)[name = tensor("zero_mean_19_cast_fp16")]; tensor zero_mean_sq_19_cast_fp16 = mul(x = zero_mean_19_cast_fp16, y = zero_mean_19_cast_fp16)[name = tensor("zero_mean_sq_19_cast_fp16")]; tensor var_3466 = const()[name = tensor("op_3466"), val = tensor([1])]; tensor var_3467_cast_fp16 = reduce_mean(axes = var_3466, keep_dims = var_2836, x = zero_mean_sq_19_cast_fp16)[name = tensor("op_3467_cast_fp16")]; tensor var_3468_to_fp16 = const()[name = tensor("op_3468_to_fp16"), val = tensor(0x1.5p-17)]; tensor var_3469_cast_fp16 = add(x = var_3467_cast_fp16, y = var_3468_to_fp16)[name = tensor("op_3469_cast_fp16")]; tensor denom_19_epsilon_0_to_fp16 = const()[name = tensor("denom_19_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; tensor denom_19_cast_fp16 = rsqrt(epsilon = denom_19_epsilon_0_to_fp16, x = var_3469_cast_fp16)[name = tensor("denom_19_cast_fp16")]; tensor out_19_cast_fp16 = mul(x = zero_mean_19_cast_fp16, y = denom_19_cast_fp16)[name = tensor("out_19_cast_fp16")]; tensor input_35_gamma_0_to_fp16 = const()[name = tensor("input_35_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30680896)))]; tensor input_35_beta_0_to_fp16 = const()[name = tensor("input_35_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30681984)))]; tensor input_35_epsilon_0_to_fp16 = const()[name = tensor("input_35_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; tensor input_35_cast_fp16 = batch_norm(beta = input_35_beta_0_to_fp16, epsilon = input_35_epsilon_0_to_fp16, gamma = input_35_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_19_cast_fp16)[name = tensor("input_35_cast_fp16")]; tensor var_3480 = const()[name = tensor("op_3480"), val = tensor([1, 1])]; tensor var_3482 = const()[name = tensor("op_3482"), val = tensor([1, 1])]; tensor input_37_pad_type_0 = const()[name = tensor("input_37_pad_type_0"), val = tensor("custom")]; tensor input_37_pad_0 = const()[name = tensor("input_37_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_4_fc1_weight_to_fp16 = const()[name = tensor("layers_4_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30683072)))]; tensor layers_4_fc1_bias_to_fp16 = const()[name = tensor("layers_4_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32780288)))]; tensor input_37_cast_fp16 = conv(bias = layers_4_fc1_bias_to_fp16, dilations = var_3482, groups = var_2835, pad = input_37_pad_0, pad_type = input_37_pad_type_0, strides = var_3480, weight = layers_4_fc1_weight_to_fp16, x = input_35_cast_fp16)[name = tensor("input_37_cast_fp16")]; tensor input_39_mode_0 = const()[name = tensor("input_39_mode_0"), val = tensor("EXACT")]; tensor input_39_cast_fp16 = gelu(mode = input_39_mode_0, x = input_37_cast_fp16)[name = tensor("input_39_cast_fp16")]; tensor var_3488 = const()[name = tensor("op_3488"), val = tensor([1, 1])]; tensor var_3490 = const()[name = tensor("op_3490"), val = tensor([1, 1])]; tensor hidden_states_13_pad_type_0 = const()[name = tensor("hidden_states_13_pad_type_0"), val = tensor("custom")]; tensor hidden_states_13_pad_0 = const()[name = tensor("hidden_states_13_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_4_fc2_weight_to_fp16 = const()[name = tensor("layers_4_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32784448)))]; tensor layers_4_fc2_bias_to_fp16 = const()[name = tensor("layers_4_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(34881664)))]; tensor hidden_states_13_cast_fp16 = conv(bias = layers_4_fc2_bias_to_fp16, dilations = var_3490, groups = var_2835, pad = hidden_states_13_pad_0, pad_type = hidden_states_13_pad_type_0, strides = var_3488, weight = layers_4_fc2_weight_to_fp16, x = input_39_cast_fp16)[name = tensor("hidden_states_13_cast_fp16")]; tensor inputs_21_cast_fp16 = add(x = inputs_19_cast_fp16, y = hidden_states_13_cast_fp16)[name = tensor("inputs_21_cast_fp16")]; tensor var_3497 = const()[name = tensor("op_3497"), val = tensor(3)]; tensor var_3510 = const()[name = tensor("op_3510"), val = tensor(1)]; tensor var_3511 = const()[name = tensor("op_3511"), val = tensor(true)]; tensor var_3521 = const()[name = tensor("op_3521"), val = tensor([1])]; tensor channels_mean_21_cast_fp16 = reduce_mean(axes = var_3521, keep_dims = var_3511, x = inputs_21_cast_fp16)[name = tensor("channels_mean_21_cast_fp16")]; tensor zero_mean_21_cast_fp16 = sub(x = inputs_21_cast_fp16, y = channels_mean_21_cast_fp16)[name = tensor("zero_mean_21_cast_fp16")]; tensor zero_mean_sq_21_cast_fp16 = mul(x = zero_mean_21_cast_fp16, y = zero_mean_21_cast_fp16)[name = tensor("zero_mean_sq_21_cast_fp16")]; tensor var_3525 = const()[name = tensor("op_3525"), val = tensor([1])]; tensor var_3526_cast_fp16 = reduce_mean(axes = var_3525, keep_dims = var_3511, x = zero_mean_sq_21_cast_fp16)[name = tensor("op_3526_cast_fp16")]; tensor var_3527_to_fp16 = const()[name = tensor("op_3527_to_fp16"), val = tensor(0x1.5p-17)]; tensor var_3528_cast_fp16 = add(x = var_3526_cast_fp16, y = var_3527_to_fp16)[name = tensor("op_3528_cast_fp16")]; tensor denom_21_epsilon_0_to_fp16 = const()[name = tensor("denom_21_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; tensor denom_21_cast_fp16 = rsqrt(epsilon = denom_21_epsilon_0_to_fp16, x = var_3528_cast_fp16)[name = tensor("denom_21_cast_fp16")]; tensor out_21_cast_fp16 = mul(x = zero_mean_21_cast_fp16, y = denom_21_cast_fp16)[name = tensor("out_21_cast_fp16")]; tensor obj_21_gamma_0_to_fp16 = const()[name = tensor("obj_21_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(34882752)))]; tensor obj_21_beta_0_to_fp16 = const()[name = tensor("obj_21_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(34883840)))]; tensor obj_21_epsilon_0_to_fp16 = const()[name = tensor("obj_21_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; tensor obj_21_cast_fp16 = batch_norm(beta = obj_21_beta_0_to_fp16, epsilon = obj_21_epsilon_0_to_fp16, gamma = obj_21_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_21_cast_fp16)[name = tensor("obj_21_cast_fp16")]; tensor var_3543 = const()[name = tensor("op_3543"), val = tensor([1, 1])]; tensor var_3545 = const()[name = tensor("op_3545"), 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_5_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_5_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(34884928)))]; tensor layers_5_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_5_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35409280)))]; tensor query_cast_fp16 = conv(bias = layers_5_self_attn_q_proj_bias_to_fp16, dilations = var_3545, groups = var_3510, pad = query_pad_0, pad_type = query_pad_type_0, strides = var_3543, weight = layers_5_self_attn_q_proj_weight_to_fp16, x = obj_21_cast_fp16)[name = tensor("query_cast_fp16")]; tensor var_3549 = const()[name = tensor("op_3549"), val = tensor([1, 1])]; tensor var_3551 = const()[name = tensor("op_3551"), 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_5_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_5_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35410368)))]; tensor key_cast_fp16 = conv(dilations = var_3551, groups = var_3510, pad = key_pad_0, pad_type = key_pad_type_0, strides = var_3549, weight = layers_5_self_attn_k_proj_weight_to_fp16, x = obj_21_cast_fp16)[name = tensor("key_cast_fp16")]; tensor var_3556 = const()[name = tensor("op_3556"), val = tensor([1, 1])]; tensor var_3558 = const()[name = tensor("op_3558"), 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_5_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_5_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35934720)))]; tensor layers_5_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_5_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(36459072)))]; tensor value_cast_fp16 = conv(bias = layers_5_self_attn_v_proj_bias_to_fp16, dilations = var_3558, groups = var_3510, pad = value_pad_0, pad_type = value_pad_type_0, strides = var_3556, weight = layers_5_self_attn_v_proj_weight_to_fp16, x = obj_21_cast_fp16)[name = tensor("value_cast_fp16")]; tensor var_3565_begin_0 = const()[name = tensor("op_3565_begin_0"), val = tensor([0, 0, 0, 0])]; tensor var_3565_end_0 = const()[name = tensor("op_3565_end_0"), val = tensor([1, 64, 1, 1500])]; tensor var_3565_end_mask_0 = const()[name = tensor("op_3565_end_mask_0"), val = tensor([true, false, true, true])]; tensor var_3565_cast_fp16 = slice_by_index(begin = var_3565_begin_0, end = var_3565_end_0, end_mask = var_3565_end_mask_0, x = query_cast_fp16)[name = tensor("op_3565_cast_fp16")]; tensor var_3569_begin_0 = const()[name = tensor("op_3569_begin_0"), val = tensor([0, 64, 0, 0])]; tensor var_3569_end_0 = const()[name = tensor("op_3569_end_0"), val = tensor([1, 128, 1, 1500])]; tensor var_3569_end_mask_0 = const()[name = tensor("op_3569_end_mask_0"), val = tensor([true, false, true, true])]; tensor var_3569_cast_fp16 = slice_by_index(begin = var_3569_begin_0, end = var_3569_end_0, end_mask = var_3569_end_mask_0, x = query_cast_fp16)[name = tensor("op_3569_cast_fp16")]; tensor var_3573_begin_0 = const()[name = tensor("op_3573_begin_0"), val = tensor([0, 128, 0, 0])]; tensor var_3573_end_0 = const()[name = tensor("op_3573_end_0"), val = tensor([1, 192, 1, 1500])]; tensor var_3573_end_mask_0 = const()[name = tensor("op_3573_end_mask_0"), val = tensor([true, false, true, true])]; tensor var_3573_cast_fp16 = slice_by_index(begin = var_3573_begin_0, end = var_3573_end_0, end_mask = var_3573_end_mask_0, x = query_cast_fp16)[name = tensor("op_3573_cast_fp16")]; tensor var_3577_begin_0 = const()[name = tensor("op_3577_begin_0"), val = tensor([0, 192, 0, 0])]; tensor var_3577_end_0 = const()[name = tensor("op_3577_end_0"), val = tensor([1, 256, 1, 1500])]; tensor var_3577_end_mask_0 = const()[name = tensor("op_3577_end_mask_0"), val = tensor([true, false, true, true])]; tensor var_3577_cast_fp16 = slice_by_index(begin = var_3577_begin_0, end = var_3577_end_0, end_mask = var_3577_end_mask_0, x = query_cast_fp16)[name = tensor("op_3577_cast_fp16")]; tensor var_3581_begin_0 = const()[name = tensor("op_3581_begin_0"), val = tensor([0, 256, 0, 0])]; tensor var_3581_end_0 = const()[name = tensor("op_3581_end_0"), val = tensor([1, 320, 1, 1500])]; tensor var_3581_end_mask_0 = const()[name = tensor("op_3581_end_mask_0"), val = tensor([true, false, true, true])]; tensor var_3581_cast_fp16 = slice_by_index(begin = var_3581_begin_0, end = var_3581_end_0, end_mask = var_3581_end_mask_0, x = query_cast_fp16)[name = tensor("op_3581_cast_fp16")]; tensor var_3585_begin_0 = const()[name = tensor("op_3585_begin_0"), val = tensor([0, 320, 0, 0])]; tensor var_3585_end_0 = const()[name = tensor("op_3585_end_0"), val = tensor([1, 384, 1, 1500])]; tensor var_3585_end_mask_0 = const()[name = tensor("op_3585_end_mask_0"), val = tensor([true, false, true, true])]; tensor var_3585_cast_fp16 = slice_by_index(begin = var_3585_begin_0, end = var_3585_end_0, end_mask = var_3585_end_mask_0, x = query_cast_fp16)[name = tensor("op_3585_cast_fp16")]; tensor var_3589_begin_0 = const()[name = tensor("op_3589_begin_0"), val = tensor([0, 384, 0, 0])]; tensor var_3589_end_0 = const()[name = tensor("op_3589_end_0"), val = tensor([1, 448, 1, 1500])]; tensor var_3589_end_mask_0 = const()[name = tensor("op_3589_end_mask_0"), val = tensor([true, false, true, true])]; tensor var_3589_cast_fp16 = slice_by_index(begin = var_3589_begin_0, end = var_3589_end_0, end_mask = var_3589_end_mask_0, x = query_cast_fp16)[name = tensor("op_3589_cast_fp16")]; tensor var_3593_begin_0 = const()[name = tensor("op_3593_begin_0"), val = tensor([0, 448, 0, 0])]; tensor var_3593_end_0 = const()[name = tensor("op_3593_end_0"), val = tensor([1, 512, 1, 1500])]; tensor var_3593_end_mask_0 = const()[name = tensor("op_3593_end_mask_0"), val = tensor([true, false, true, true])]; tensor var_3593_cast_fp16 = slice_by_index(begin = var_3593_begin_0, end = var_3593_end_0, end_mask = var_3593_end_mask_0, x = query_cast_fp16)[name = tensor("op_3593_cast_fp16")]; tensor var_3602_begin_0 = const()[name = tensor("op_3602_begin_0"), val = tensor([0, 0, 0, 0])]; tensor var_3602_end_0 = const()[name = tensor("op_3602_end_0"), val = tensor([1, 64, 1, 375])]; tensor var_3602_end_mask_0 = const()[name = tensor("op_3602_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_3602_cast_fp16 = slice_by_index(begin = var_3602_begin_0, end = var_3602_end_0, end_mask = var_3602_end_mask_0, x = var_3565_cast_fp16)[name = tensor("op_3602_cast_fp16")]; tensor var_3609_begin_0 = const()[name = tensor("op_3609_begin_0"), val = tensor([0, 0, 0, 375])]; tensor var_3609_end_0 = const()[name = tensor("op_3609_end_0"), val = tensor([1, 64, 1, 750])]; tensor var_3609_end_mask_0 = const()[name = tensor("op_3609_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_3609_cast_fp16 = slice_by_index(begin = var_3609_begin_0, end = var_3609_end_0, end_mask = var_3609_end_mask_0, x = var_3565_cast_fp16)[name = tensor("op_3609_cast_fp16")]; tensor var_3616_begin_0 = const()[name = tensor("op_3616_begin_0"), val = tensor([0, 0, 0, 750])]; tensor var_3616_end_0 = const()[name = tensor("op_3616_end_0"), val = tensor([1, 64, 1, 1125])]; tensor var_3616_end_mask_0 = const()[name = tensor("op_3616_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_3616_cast_fp16 = slice_by_index(begin = var_3616_begin_0, end = var_3616_end_0, end_mask = var_3616_end_mask_0, x = var_3565_cast_fp16)[name = tensor("op_3616_cast_fp16")]; tensor var_3623_begin_0 = const()[name = tensor("op_3623_begin_0"), val = tensor([0, 0, 0, 1125])]; tensor var_3623_end_0 = const()[name = tensor("op_3623_end_0"), val = tensor([1, 64, 1, 1500])]; tensor var_3623_end_mask_0 = const()[name = tensor("op_3623_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_3623_cast_fp16 = slice_by_index(begin = var_3623_begin_0, end = var_3623_end_0, end_mask = var_3623_end_mask_0, x = var_3565_cast_fp16)[name = tensor("op_3623_cast_fp16")]; tensor var_3630_begin_0 = const()[name = tensor("op_3630_begin_0"), val = tensor([0, 0, 0, 0])]; tensor var_3630_end_0 = const()[name = tensor("op_3630_end_0"), val = tensor([1, 64, 1, 375])]; tensor var_3630_end_mask_0 = const()[name = tensor("op_3630_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_3630_cast_fp16 = slice_by_index(begin = var_3630_begin_0, end = var_3630_end_0, end_mask = var_3630_end_mask_0, x = var_3569_cast_fp16)[name = tensor("op_3630_cast_fp16")]; tensor var_3637_begin_0 = const()[name = tensor("op_3637_begin_0"), val = tensor([0, 0, 0, 375])]; tensor var_3637_end_0 = const()[name = tensor("op_3637_end_0"), val = tensor([1, 64, 1, 750])]; tensor var_3637_end_mask_0 = const()[name = tensor("op_3637_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_3637_cast_fp16 = slice_by_index(begin = var_3637_begin_0, end = var_3637_end_0, end_mask = var_3637_end_mask_0, x = var_3569_cast_fp16)[name = tensor("op_3637_cast_fp16")]; tensor var_3644_begin_0 = const()[name = tensor("op_3644_begin_0"), val = tensor([0, 0, 0, 750])]; tensor var_3644_end_0 = const()[name = tensor("op_3644_end_0"), val = tensor([1, 64, 1, 1125])]; tensor var_3644_end_mask_0 = const()[name = tensor("op_3644_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_3644_cast_fp16 = slice_by_index(begin = var_3644_begin_0, end = var_3644_end_0, end_mask = var_3644_end_mask_0, x = var_3569_cast_fp16)[name = tensor("op_3644_cast_fp16")]; tensor var_3651_begin_0 = const()[name = tensor("op_3651_begin_0"), val = tensor([0, 0, 0, 1125])]; tensor var_3651_end_0 = const()[name = tensor("op_3651_end_0"), val = tensor([1, 64, 1, 1500])]; tensor var_3651_end_mask_0 = const()[name = tensor("op_3651_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_3651_cast_fp16 = slice_by_index(begin = var_3651_begin_0, end = var_3651_end_0, end_mask = var_3651_end_mask_0, x = var_3569_cast_fp16)[name = tensor("op_3651_cast_fp16")]; tensor var_3658_begin_0 = const()[name = tensor("op_3658_begin_0"), val = tensor([0, 0, 0, 0])]; tensor var_3658_end_0 = const()[name = tensor("op_3658_end_0"), val = tensor([1, 64, 1, 375])]; tensor var_3658_end_mask_0 = const()[name = tensor("op_3658_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_3658_cast_fp16 = slice_by_index(begin = var_3658_begin_0, end = var_3658_end_0, end_mask = var_3658_end_mask_0, x = var_3573_cast_fp16)[name = tensor("op_3658_cast_fp16")]; tensor var_3665_begin_0 = const()[name = tensor("op_3665_begin_0"), val = tensor([0, 0, 0, 375])]; tensor var_3665_end_0 = const()[name = tensor("op_3665_end_0"), val = tensor([1, 64, 1, 750])]; tensor var_3665_end_mask_0 = const()[name = tensor("op_3665_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_3665_cast_fp16 = slice_by_index(begin = var_3665_begin_0, end = var_3665_end_0, end_mask = var_3665_end_mask_0, x = var_3573_cast_fp16)[name = tensor("op_3665_cast_fp16")]; tensor var_3672_begin_0 = const()[name = tensor("op_3672_begin_0"), val = tensor([0, 0, 0, 750])]; tensor var_3672_end_0 = const()[name = tensor("op_3672_end_0"), val = tensor([1, 64, 1, 1125])]; tensor var_3672_end_mask_0 = const()[name = tensor("op_3672_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_3672_cast_fp16 = slice_by_index(begin = var_3672_begin_0, end = var_3672_end_0, end_mask = var_3672_end_mask_0, x = var_3573_cast_fp16)[name = tensor("op_3672_cast_fp16")]; tensor var_3679_begin_0 = const()[name = tensor("op_3679_begin_0"), val = tensor([0, 0, 0, 1125])]; tensor var_3679_end_0 = const()[name = tensor("op_3679_end_0"), val = tensor([1, 64, 1, 1500])]; tensor var_3679_end_mask_0 = const()[name = tensor("op_3679_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_3679_cast_fp16 = slice_by_index(begin = var_3679_begin_0, end = var_3679_end_0, end_mask = var_3679_end_mask_0, x = var_3573_cast_fp16)[name = tensor("op_3679_cast_fp16")]; tensor var_3686_begin_0 = const()[name = tensor("op_3686_begin_0"), val = tensor([0, 0, 0, 0])]; tensor var_3686_end_0 = const()[name = tensor("op_3686_end_0"), val = tensor([1, 64, 1, 375])]; tensor var_3686_end_mask_0 = const()[name = tensor("op_3686_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_3686_cast_fp16 = slice_by_index(begin = var_3686_begin_0, end = var_3686_end_0, end_mask = var_3686_end_mask_0, x = var_3577_cast_fp16)[name = tensor("op_3686_cast_fp16")]; tensor var_3693_begin_0 = const()[name = tensor("op_3693_begin_0"), val = tensor([0, 0, 0, 375])]; tensor var_3693_end_0 = const()[name = tensor("op_3693_end_0"), val = tensor([1, 64, 1, 750])]; tensor var_3693_end_mask_0 = const()[name = tensor("op_3693_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_3693_cast_fp16 = slice_by_index(begin = var_3693_begin_0, end = var_3693_end_0, end_mask = var_3693_end_mask_0, x = var_3577_cast_fp16)[name = tensor("op_3693_cast_fp16")]; tensor var_3700_begin_0 = const()[name = tensor("op_3700_begin_0"), val = tensor([0, 0, 0, 750])]; tensor var_3700_end_0 = const()[name = tensor("op_3700_end_0"), val = tensor([1, 64, 1, 1125])]; tensor var_3700_end_mask_0 = const()[name = tensor("op_3700_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_3700_cast_fp16 = slice_by_index(begin = var_3700_begin_0, end = var_3700_end_0, end_mask = var_3700_end_mask_0, x = var_3577_cast_fp16)[name = tensor("op_3700_cast_fp16")]; tensor var_3707_begin_0 = const()[name = tensor("op_3707_begin_0"), val = tensor([0, 0, 0, 1125])]; tensor var_3707_end_0 = const()[name = tensor("op_3707_end_0"), val = tensor([1, 64, 1, 1500])]; tensor var_3707_end_mask_0 = const()[name = tensor("op_3707_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_3707_cast_fp16 = slice_by_index(begin = var_3707_begin_0, end = var_3707_end_0, end_mask = var_3707_end_mask_0, x = var_3577_cast_fp16)[name = tensor("op_3707_cast_fp16")]; tensor var_3714_begin_0 = const()[name = tensor("op_3714_begin_0"), val = tensor([0, 0, 0, 0])]; tensor var_3714_end_0 = const()[name = tensor("op_3714_end_0"), val = tensor([1, 64, 1, 375])]; tensor var_3714_end_mask_0 = const()[name = tensor("op_3714_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_3714_cast_fp16 = slice_by_index(begin = var_3714_begin_0, end = var_3714_end_0, end_mask = var_3714_end_mask_0, x = var_3581_cast_fp16)[name = tensor("op_3714_cast_fp16")]; tensor var_3721_begin_0 = const()[name = tensor("op_3721_begin_0"), val = tensor([0, 0, 0, 375])]; tensor var_3721_end_0 = const()[name = tensor("op_3721_end_0"), val = tensor([1, 64, 1, 750])]; tensor var_3721_end_mask_0 = const()[name = tensor("op_3721_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_3721_cast_fp16 = slice_by_index(begin = var_3721_begin_0, end = var_3721_end_0, end_mask = var_3721_end_mask_0, x = var_3581_cast_fp16)[name = tensor("op_3721_cast_fp16")]; tensor var_3728_begin_0 = const()[name = tensor("op_3728_begin_0"), val = tensor([0, 0, 0, 750])]; tensor var_3728_end_0 = const()[name = tensor("op_3728_end_0"), val = tensor([1, 64, 1, 1125])]; tensor var_3728_end_mask_0 = const()[name = tensor("op_3728_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_3728_cast_fp16 = slice_by_index(begin = var_3728_begin_0, end = var_3728_end_0, end_mask = var_3728_end_mask_0, x = var_3581_cast_fp16)[name = tensor("op_3728_cast_fp16")]; tensor var_3735_begin_0 = const()[name = tensor("op_3735_begin_0"), val = tensor([0, 0, 0, 1125])]; tensor var_3735_end_0 = const()[name = tensor("op_3735_end_0"), val = tensor([1, 64, 1, 1500])]; tensor var_3735_end_mask_0 = const()[name = tensor("op_3735_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_3735_cast_fp16 = slice_by_index(begin = var_3735_begin_0, end = var_3735_end_0, end_mask = var_3735_end_mask_0, x = var_3581_cast_fp16)[name = tensor("op_3735_cast_fp16")]; tensor var_3742_begin_0 = const()[name = tensor("op_3742_begin_0"), val = tensor([0, 0, 0, 0])]; tensor var_3742_end_0 = const()[name = tensor("op_3742_end_0"), val = tensor([1, 64, 1, 375])]; tensor var_3742_end_mask_0 = const()[name = tensor("op_3742_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_3742_cast_fp16 = slice_by_index(begin = var_3742_begin_0, end = var_3742_end_0, end_mask = var_3742_end_mask_0, x = var_3585_cast_fp16)[name = tensor("op_3742_cast_fp16")]; tensor var_3749_begin_0 = const()[name = tensor("op_3749_begin_0"), val = tensor([0, 0, 0, 375])]; tensor var_3749_end_0 = const()[name = tensor("op_3749_end_0"), val = tensor([1, 64, 1, 750])]; tensor var_3749_end_mask_0 = const()[name = tensor("op_3749_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_3749_cast_fp16 = slice_by_index(begin = var_3749_begin_0, end = var_3749_end_0, end_mask = var_3749_end_mask_0, x = var_3585_cast_fp16)[name = tensor("op_3749_cast_fp16")]; tensor var_3756_begin_0 = const()[name = tensor("op_3756_begin_0"), val = tensor([0, 0, 0, 750])]; tensor var_3756_end_0 = const()[name = tensor("op_3756_end_0"), val = tensor([1, 64, 1, 1125])]; tensor var_3756_end_mask_0 = const()[name = tensor("op_3756_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_3756_cast_fp16 = slice_by_index(begin = var_3756_begin_0, end = var_3756_end_0, end_mask = var_3756_end_mask_0, x = var_3585_cast_fp16)[name = tensor("op_3756_cast_fp16")]; tensor var_3763_begin_0 = const()[name = tensor("op_3763_begin_0"), val = tensor([0, 0, 0, 1125])]; tensor var_3763_end_0 = const()[name = tensor("op_3763_end_0"), val = tensor([1, 64, 1, 1500])]; tensor var_3763_end_mask_0 = const()[name = tensor("op_3763_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_3763_cast_fp16 = slice_by_index(begin = var_3763_begin_0, end = var_3763_end_0, end_mask = var_3763_end_mask_0, x = var_3585_cast_fp16)[name = tensor("op_3763_cast_fp16")]; tensor var_3770_begin_0 = const()[name = tensor("op_3770_begin_0"), val = tensor([0, 0, 0, 0])]; tensor var_3770_end_0 = const()[name = tensor("op_3770_end_0"), val = tensor([1, 64, 1, 375])]; tensor var_3770_end_mask_0 = const()[name = tensor("op_3770_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_3770_cast_fp16 = slice_by_index(begin = var_3770_begin_0, end = var_3770_end_0, end_mask = var_3770_end_mask_0, x = var_3589_cast_fp16)[name = tensor("op_3770_cast_fp16")]; tensor var_3777_begin_0 = const()[name = tensor("op_3777_begin_0"), val = tensor([0, 0, 0, 375])]; tensor var_3777_end_0 = const()[name = tensor("op_3777_end_0"), val = tensor([1, 64, 1, 750])]; tensor var_3777_end_mask_0 = const()[name = tensor("op_3777_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_3777_cast_fp16 = slice_by_index(begin = var_3777_begin_0, end = var_3777_end_0, end_mask = var_3777_end_mask_0, x = var_3589_cast_fp16)[name = tensor("op_3777_cast_fp16")]; tensor var_3784_begin_0 = const()[name = tensor("op_3784_begin_0"), val = tensor([0, 0, 0, 750])]; tensor var_3784_end_0 = const()[name = tensor("op_3784_end_0"), val = tensor([1, 64, 1, 1125])]; tensor var_3784_end_mask_0 = const()[name = tensor("op_3784_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_3784_cast_fp16 = slice_by_index(begin = var_3784_begin_0, end = var_3784_end_0, end_mask = var_3784_end_mask_0, x = var_3589_cast_fp16)[name = tensor("op_3784_cast_fp16")]; tensor var_3791_begin_0 = const()[name = tensor("op_3791_begin_0"), val = tensor([0, 0, 0, 1125])]; tensor var_3791_end_0 = const()[name = tensor("op_3791_end_0"), val = tensor([1, 64, 1, 1500])]; tensor var_3791_end_mask_0 = const()[name = tensor("op_3791_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_3791_cast_fp16 = slice_by_index(begin = var_3791_begin_0, end = var_3791_end_0, end_mask = var_3791_end_mask_0, x = var_3589_cast_fp16)[name = tensor("op_3791_cast_fp16")]; tensor var_3798_begin_0 = const()[name = tensor("op_3798_begin_0"), val = tensor([0, 0, 0, 0])]; tensor var_3798_end_0 = const()[name = tensor("op_3798_end_0"), val = tensor([1, 64, 1, 375])]; tensor var_3798_end_mask_0 = const()[name = tensor("op_3798_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_3798_cast_fp16 = slice_by_index(begin = var_3798_begin_0, end = var_3798_end_0, end_mask = var_3798_end_mask_0, x = var_3593_cast_fp16)[name = tensor("op_3798_cast_fp16")]; tensor var_3805_begin_0 = const()[name = tensor("op_3805_begin_0"), val = tensor([0, 0, 0, 375])]; tensor var_3805_end_0 = const()[name = tensor("op_3805_end_0"), val = tensor([1, 64, 1, 750])]; tensor var_3805_end_mask_0 = const()[name = tensor("op_3805_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_3805_cast_fp16 = slice_by_index(begin = var_3805_begin_0, end = var_3805_end_0, end_mask = var_3805_end_mask_0, x = var_3593_cast_fp16)[name = tensor("op_3805_cast_fp16")]; tensor var_3812_begin_0 = const()[name = tensor("op_3812_begin_0"), val = tensor([0, 0, 0, 750])]; tensor var_3812_end_0 = const()[name = tensor("op_3812_end_0"), val = tensor([1, 64, 1, 1125])]; tensor var_3812_end_mask_0 = const()[name = tensor("op_3812_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_3812_cast_fp16 = slice_by_index(begin = var_3812_begin_0, end = var_3812_end_0, end_mask = var_3812_end_mask_0, x = var_3593_cast_fp16)[name = tensor("op_3812_cast_fp16")]; tensor var_3819_begin_0 = const()[name = tensor("op_3819_begin_0"), val = tensor([0, 0, 0, 1125])]; tensor var_3819_end_0 = const()[name = tensor("op_3819_end_0"), val = tensor([1, 64, 1, 1500])]; tensor var_3819_end_mask_0 = const()[name = tensor("op_3819_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_3819_cast_fp16 = slice_by_index(begin = var_3819_begin_0, end = var_3819_end_0, end_mask = var_3819_end_mask_0, x = var_3593_cast_fp16)[name = tensor("op_3819_cast_fp16")]; tensor k_perm_0 = const()[name = tensor("k_perm_0"), val = tensor([0, 3, 2, 1])]; tensor var_3824_begin_0 = const()[name = tensor("op_3824_begin_0"), val = tensor([0, 0, 0, 0])]; tensor var_3824_end_0 = const()[name = tensor("op_3824_end_0"), val = tensor([1, 1500, 1, 64])]; tensor var_3824_end_mask_0 = const()[name = tensor("op_3824_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_3824_cast_fp16 = slice_by_index(begin = var_3824_begin_0, end = var_3824_end_0, end_mask = var_3824_end_mask_0, x = transpose_0)[name = tensor("op_3824_cast_fp16")]; tensor var_3828_begin_0 = const()[name = tensor("op_3828_begin_0"), val = tensor([0, 0, 0, 64])]; tensor var_3828_end_0 = const()[name = tensor("op_3828_end_0"), val = tensor([1, 1500, 1, 128])]; tensor var_3828_end_mask_0 = const()[name = tensor("op_3828_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_3828_cast_fp16 = slice_by_index(begin = var_3828_begin_0, end = var_3828_end_0, end_mask = var_3828_end_mask_0, x = transpose_0)[name = tensor("op_3828_cast_fp16")]; tensor var_3832_begin_0 = const()[name = tensor("op_3832_begin_0"), val = tensor([0, 0, 0, 128])]; tensor var_3832_end_0 = const()[name = tensor("op_3832_end_0"), val = tensor([1, 1500, 1, 192])]; tensor var_3832_end_mask_0 = const()[name = tensor("op_3832_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_3832_cast_fp16 = slice_by_index(begin = var_3832_begin_0, end = var_3832_end_0, end_mask = var_3832_end_mask_0, x = transpose_0)[name = tensor("op_3832_cast_fp16")]; tensor var_3836_begin_0 = const()[name = tensor("op_3836_begin_0"), val = tensor([0, 0, 0, 192])]; tensor var_3836_end_0 = const()[name = tensor("op_3836_end_0"), val = tensor([1, 1500, 1, 256])]; tensor var_3836_end_mask_0 = const()[name = tensor("op_3836_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_3836_cast_fp16 = slice_by_index(begin = var_3836_begin_0, end = var_3836_end_0, end_mask = var_3836_end_mask_0, x = transpose_0)[name = tensor("op_3836_cast_fp16")]; tensor var_3840_begin_0 = const()[name = tensor("op_3840_begin_0"), val = tensor([0, 0, 0, 256])]; tensor var_3840_end_0 = const()[name = tensor("op_3840_end_0"), val = tensor([1, 1500, 1, 320])]; tensor var_3840_end_mask_0 = const()[name = tensor("op_3840_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_3840_cast_fp16 = slice_by_index(begin = var_3840_begin_0, end = var_3840_end_0, end_mask = var_3840_end_mask_0, x = transpose_0)[name = tensor("op_3840_cast_fp16")]; tensor var_3844_begin_0 = const()[name = tensor("op_3844_begin_0"), val = tensor([0, 0, 0, 320])]; tensor var_3844_end_0 = const()[name = tensor("op_3844_end_0"), val = tensor([1, 1500, 1, 384])]; tensor var_3844_end_mask_0 = const()[name = tensor("op_3844_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_3844_cast_fp16 = slice_by_index(begin = var_3844_begin_0, end = var_3844_end_0, end_mask = var_3844_end_mask_0, x = transpose_0)[name = tensor("op_3844_cast_fp16")]; tensor var_3848_begin_0 = const()[name = tensor("op_3848_begin_0"), val = tensor([0, 0, 0, 384])]; tensor var_3848_end_0 = const()[name = tensor("op_3848_end_0"), val = tensor([1, 1500, 1, 448])]; tensor var_3848_end_mask_0 = const()[name = tensor("op_3848_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_3848_cast_fp16 = slice_by_index(begin = var_3848_begin_0, end = var_3848_end_0, end_mask = var_3848_end_mask_0, x = transpose_0)[name = tensor("op_3848_cast_fp16")]; tensor var_3852_begin_0 = const()[name = tensor("op_3852_begin_0"), val = tensor([0, 0, 0, 448])]; tensor var_3852_end_0 = const()[name = tensor("op_3852_end_0"), val = tensor([1, 1500, 1, 512])]; tensor var_3852_end_mask_0 = const()[name = tensor("op_3852_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_3852_cast_fp16 = slice_by_index(begin = var_3852_begin_0, end = var_3852_end_0, end_mask = var_3852_end_mask_0, x = transpose_0)[name = tensor("op_3852_cast_fp16")]; tensor var_3854_begin_0 = const()[name = tensor("op_3854_begin_0"), val = tensor([0, 0, 0, 0])]; tensor var_3854_end_0 = const()[name = tensor("op_3854_end_0"), val = tensor([1, 64, 1, 1500])]; tensor var_3854_end_mask_0 = const()[name = tensor("op_3854_end_mask_0"), val = tensor([true, false, true, true])]; tensor var_3854_cast_fp16 = slice_by_index(begin = var_3854_begin_0, end = var_3854_end_0, end_mask = var_3854_end_mask_0, x = value_cast_fp16)[name = tensor("op_3854_cast_fp16")]; tensor var_3858_begin_0 = const()[name = tensor("op_3858_begin_0"), val = tensor([0, 64, 0, 0])]; tensor var_3858_end_0 = const()[name = tensor("op_3858_end_0"), val = tensor([1, 128, 1, 1500])]; tensor var_3858_end_mask_0 = const()[name = tensor("op_3858_end_mask_0"), val = tensor([true, false, true, true])]; tensor var_3858_cast_fp16 = slice_by_index(begin = var_3858_begin_0, end = var_3858_end_0, end_mask = var_3858_end_mask_0, x = value_cast_fp16)[name = tensor("op_3858_cast_fp16")]; tensor var_3862_begin_0 = const()[name = tensor("op_3862_begin_0"), val = tensor([0, 128, 0, 0])]; tensor var_3862_end_0 = const()[name = tensor("op_3862_end_0"), val = tensor([1, 192, 1, 1500])]; tensor var_3862_end_mask_0 = const()[name = tensor("op_3862_end_mask_0"), val = tensor([true, false, true, true])]; tensor var_3862_cast_fp16 = slice_by_index(begin = var_3862_begin_0, end = var_3862_end_0, end_mask = var_3862_end_mask_0, x = value_cast_fp16)[name = tensor("op_3862_cast_fp16")]; tensor var_3866_begin_0 = const()[name = tensor("op_3866_begin_0"), val = tensor([0, 192, 0, 0])]; tensor var_3866_end_0 = const()[name = tensor("op_3866_end_0"), val = tensor([1, 256, 1, 1500])]; tensor var_3866_end_mask_0 = const()[name = tensor("op_3866_end_mask_0"), val = tensor([true, false, true, true])]; tensor var_3866_cast_fp16 = slice_by_index(begin = var_3866_begin_0, end = var_3866_end_0, end_mask = var_3866_end_mask_0, x = value_cast_fp16)[name = tensor("op_3866_cast_fp16")]; tensor var_3870_begin_0 = const()[name = tensor("op_3870_begin_0"), val = tensor([0, 256, 0, 0])]; tensor var_3870_end_0 = const()[name = tensor("op_3870_end_0"), val = tensor([1, 320, 1, 1500])]; tensor var_3870_end_mask_0 = const()[name = tensor("op_3870_end_mask_0"), val = tensor([true, false, true, true])]; tensor var_3870_cast_fp16 = slice_by_index(begin = var_3870_begin_0, end = var_3870_end_0, end_mask = var_3870_end_mask_0, x = value_cast_fp16)[name = tensor("op_3870_cast_fp16")]; tensor var_3874_begin_0 = const()[name = tensor("op_3874_begin_0"), val = tensor([0, 320, 0, 0])]; tensor var_3874_end_0 = const()[name = tensor("op_3874_end_0"), val = tensor([1, 384, 1, 1500])]; tensor var_3874_end_mask_0 = const()[name = tensor("op_3874_end_mask_0"), val = tensor([true, false, true, true])]; tensor var_3874_cast_fp16 = slice_by_index(begin = var_3874_begin_0, end = var_3874_end_0, end_mask = var_3874_end_mask_0, x = value_cast_fp16)[name = tensor("op_3874_cast_fp16")]; tensor var_3878_begin_0 = const()[name = tensor("op_3878_begin_0"), val = tensor([0, 384, 0, 0])]; tensor var_3878_end_0 = const()[name = tensor("op_3878_end_0"), val = tensor([1, 448, 1, 1500])]; tensor var_3878_end_mask_0 = const()[name = tensor("op_3878_end_mask_0"), val = tensor([true, false, true, true])]; tensor var_3878_cast_fp16 = slice_by_index(begin = var_3878_begin_0, end = var_3878_end_0, end_mask = var_3878_end_mask_0, x = value_cast_fp16)[name = tensor("op_3878_cast_fp16")]; tensor var_3882_begin_0 = const()[name = tensor("op_3882_begin_0"), val = tensor([0, 448, 0, 0])]; tensor var_3882_end_0 = const()[name = tensor("op_3882_end_0"), val = tensor([1, 512, 1, 1500])]; tensor var_3882_end_mask_0 = const()[name = tensor("op_3882_end_mask_0"), val = tensor([true, false, true, true])]; tensor var_3882_cast_fp16 = slice_by_index(begin = var_3882_begin_0, end = var_3882_end_0, end_mask = var_3882_end_mask_0, x = value_cast_fp16)[name = tensor("op_3882_cast_fp16")]; tensor var_3886_equation_0 = const()[name = tensor("op_3886_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_3886_cast_fp16 = einsum(equation = var_3886_equation_0, values = (var_3824_cast_fp16, var_3602_cast_fp16))[name = tensor("op_3886_cast_fp16")]; tensor var_3887_to_fp16 = const()[name = tensor("op_3887_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_321_cast_fp16 = mul(x = var_3886_cast_fp16, y = var_3887_to_fp16)[name = tensor("aw_chunk_321_cast_fp16")]; tensor var_3890_equation_0 = const()[name = tensor("op_3890_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_3890_cast_fp16 = einsum(equation = var_3890_equation_0, values = (var_3824_cast_fp16, var_3609_cast_fp16))[name = tensor("op_3890_cast_fp16")]; tensor var_3891_to_fp16 = const()[name = tensor("op_3891_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_323_cast_fp16 = mul(x = var_3890_cast_fp16, y = var_3891_to_fp16)[name = tensor("aw_chunk_323_cast_fp16")]; tensor var_3894_equation_0 = const()[name = tensor("op_3894_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_3894_cast_fp16 = einsum(equation = var_3894_equation_0, values = (var_3824_cast_fp16, var_3616_cast_fp16))[name = tensor("op_3894_cast_fp16")]; tensor var_3895_to_fp16 = const()[name = tensor("op_3895_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_325_cast_fp16 = mul(x = var_3894_cast_fp16, y = var_3895_to_fp16)[name = tensor("aw_chunk_325_cast_fp16")]; tensor var_3898_equation_0 = const()[name = tensor("op_3898_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_3898_cast_fp16 = einsum(equation = var_3898_equation_0, values = (var_3824_cast_fp16, var_3623_cast_fp16))[name = tensor("op_3898_cast_fp16")]; tensor var_3899_to_fp16 = const()[name = tensor("op_3899_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_327_cast_fp16 = mul(x = var_3898_cast_fp16, y = var_3899_to_fp16)[name = tensor("aw_chunk_327_cast_fp16")]; tensor var_3902_equation_0 = const()[name = tensor("op_3902_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_3902_cast_fp16 = einsum(equation = var_3902_equation_0, values = (var_3828_cast_fp16, var_3630_cast_fp16))[name = tensor("op_3902_cast_fp16")]; tensor var_3903_to_fp16 = const()[name = tensor("op_3903_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_329_cast_fp16 = mul(x = var_3902_cast_fp16, y = var_3903_to_fp16)[name = tensor("aw_chunk_329_cast_fp16")]; tensor var_3906_equation_0 = const()[name = tensor("op_3906_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_3906_cast_fp16 = einsum(equation = var_3906_equation_0, values = (var_3828_cast_fp16, var_3637_cast_fp16))[name = tensor("op_3906_cast_fp16")]; tensor var_3907_to_fp16 = const()[name = tensor("op_3907_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_331_cast_fp16 = mul(x = var_3906_cast_fp16, y = var_3907_to_fp16)[name = tensor("aw_chunk_331_cast_fp16")]; tensor var_3910_equation_0 = const()[name = tensor("op_3910_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_3910_cast_fp16 = einsum(equation = var_3910_equation_0, values = (var_3828_cast_fp16, var_3644_cast_fp16))[name = tensor("op_3910_cast_fp16")]; tensor var_3911_to_fp16 = const()[name = tensor("op_3911_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_333_cast_fp16 = mul(x = var_3910_cast_fp16, y = var_3911_to_fp16)[name = tensor("aw_chunk_333_cast_fp16")]; tensor var_3914_equation_0 = const()[name = tensor("op_3914_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_3914_cast_fp16 = einsum(equation = var_3914_equation_0, values = (var_3828_cast_fp16, var_3651_cast_fp16))[name = tensor("op_3914_cast_fp16")]; tensor var_3915_to_fp16 = const()[name = tensor("op_3915_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_335_cast_fp16 = mul(x = var_3914_cast_fp16, y = var_3915_to_fp16)[name = tensor("aw_chunk_335_cast_fp16")]; tensor var_3918_equation_0 = const()[name = tensor("op_3918_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_3918_cast_fp16 = einsum(equation = var_3918_equation_0, values = (var_3832_cast_fp16, var_3658_cast_fp16))[name = tensor("op_3918_cast_fp16")]; tensor var_3919_to_fp16 = const()[name = tensor("op_3919_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_337_cast_fp16 = mul(x = var_3918_cast_fp16, y = var_3919_to_fp16)[name = tensor("aw_chunk_337_cast_fp16")]; tensor var_3922_equation_0 = const()[name = tensor("op_3922_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_3922_cast_fp16 = einsum(equation = var_3922_equation_0, values = (var_3832_cast_fp16, var_3665_cast_fp16))[name = tensor("op_3922_cast_fp16")]; tensor var_3923_to_fp16 = const()[name = tensor("op_3923_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_339_cast_fp16 = mul(x = var_3922_cast_fp16, y = var_3923_to_fp16)[name = tensor("aw_chunk_339_cast_fp16")]; tensor var_3926_equation_0 = const()[name = tensor("op_3926_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_3926_cast_fp16 = einsum(equation = var_3926_equation_0, values = (var_3832_cast_fp16, var_3672_cast_fp16))[name = tensor("op_3926_cast_fp16")]; tensor var_3927_to_fp16 = const()[name = tensor("op_3927_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_341_cast_fp16 = mul(x = var_3926_cast_fp16, y = var_3927_to_fp16)[name = tensor("aw_chunk_341_cast_fp16")]; tensor var_3930_equation_0 = const()[name = tensor("op_3930_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_3930_cast_fp16 = einsum(equation = var_3930_equation_0, values = (var_3832_cast_fp16, var_3679_cast_fp16))[name = tensor("op_3930_cast_fp16")]; tensor var_3931_to_fp16 = const()[name = tensor("op_3931_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_343_cast_fp16 = mul(x = var_3930_cast_fp16, y = var_3931_to_fp16)[name = tensor("aw_chunk_343_cast_fp16")]; tensor var_3934_equation_0 = const()[name = tensor("op_3934_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_3934_cast_fp16 = einsum(equation = var_3934_equation_0, values = (var_3836_cast_fp16, var_3686_cast_fp16))[name = tensor("op_3934_cast_fp16")]; tensor var_3935_to_fp16 = const()[name = tensor("op_3935_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_345_cast_fp16 = mul(x = var_3934_cast_fp16, y = var_3935_to_fp16)[name = tensor("aw_chunk_345_cast_fp16")]; tensor var_3938_equation_0 = const()[name = tensor("op_3938_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_3938_cast_fp16 = einsum(equation = var_3938_equation_0, values = (var_3836_cast_fp16, var_3693_cast_fp16))[name = tensor("op_3938_cast_fp16")]; tensor var_3939_to_fp16 = const()[name = tensor("op_3939_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_347_cast_fp16 = mul(x = var_3938_cast_fp16, y = var_3939_to_fp16)[name = tensor("aw_chunk_347_cast_fp16")]; tensor var_3942_equation_0 = const()[name = tensor("op_3942_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_3942_cast_fp16 = einsum(equation = var_3942_equation_0, values = (var_3836_cast_fp16, var_3700_cast_fp16))[name = tensor("op_3942_cast_fp16")]; tensor var_3943_to_fp16 = const()[name = tensor("op_3943_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_349_cast_fp16 = mul(x = var_3942_cast_fp16, y = var_3943_to_fp16)[name = tensor("aw_chunk_349_cast_fp16")]; tensor var_3946_equation_0 = const()[name = tensor("op_3946_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_3946_cast_fp16 = einsum(equation = var_3946_equation_0, values = (var_3836_cast_fp16, var_3707_cast_fp16))[name = tensor("op_3946_cast_fp16")]; tensor var_3947_to_fp16 = const()[name = tensor("op_3947_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_351_cast_fp16 = mul(x = var_3946_cast_fp16, y = var_3947_to_fp16)[name = tensor("aw_chunk_351_cast_fp16")]; tensor var_3950_equation_0 = const()[name = tensor("op_3950_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_3950_cast_fp16 = einsum(equation = var_3950_equation_0, values = (var_3840_cast_fp16, var_3714_cast_fp16))[name = tensor("op_3950_cast_fp16")]; tensor var_3951_to_fp16 = const()[name = tensor("op_3951_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_353_cast_fp16 = mul(x = var_3950_cast_fp16, y = var_3951_to_fp16)[name = tensor("aw_chunk_353_cast_fp16")]; tensor var_3954_equation_0 = const()[name = tensor("op_3954_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_3954_cast_fp16 = einsum(equation = var_3954_equation_0, values = (var_3840_cast_fp16, var_3721_cast_fp16))[name = tensor("op_3954_cast_fp16")]; tensor var_3955_to_fp16 = const()[name = tensor("op_3955_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_355_cast_fp16 = mul(x = var_3954_cast_fp16, y = var_3955_to_fp16)[name = tensor("aw_chunk_355_cast_fp16")]; tensor var_3958_equation_0 = const()[name = tensor("op_3958_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_3958_cast_fp16 = einsum(equation = var_3958_equation_0, values = (var_3840_cast_fp16, var_3728_cast_fp16))[name = tensor("op_3958_cast_fp16")]; tensor var_3959_to_fp16 = const()[name = tensor("op_3959_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_357_cast_fp16 = mul(x = var_3958_cast_fp16, y = var_3959_to_fp16)[name = tensor("aw_chunk_357_cast_fp16")]; tensor var_3962_equation_0 = const()[name = tensor("op_3962_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_3962_cast_fp16 = einsum(equation = var_3962_equation_0, values = (var_3840_cast_fp16, var_3735_cast_fp16))[name = tensor("op_3962_cast_fp16")]; tensor var_3963_to_fp16 = const()[name = tensor("op_3963_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_359_cast_fp16 = mul(x = var_3962_cast_fp16, y = var_3963_to_fp16)[name = tensor("aw_chunk_359_cast_fp16")]; tensor var_3966_equation_0 = const()[name = tensor("op_3966_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_3966_cast_fp16 = einsum(equation = var_3966_equation_0, values = (var_3844_cast_fp16, var_3742_cast_fp16))[name = tensor("op_3966_cast_fp16")]; tensor var_3967_to_fp16 = const()[name = tensor("op_3967_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_361_cast_fp16 = mul(x = var_3966_cast_fp16, y = var_3967_to_fp16)[name = tensor("aw_chunk_361_cast_fp16")]; tensor var_3970_equation_0 = const()[name = tensor("op_3970_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_3970_cast_fp16 = einsum(equation = var_3970_equation_0, values = (var_3844_cast_fp16, var_3749_cast_fp16))[name = tensor("op_3970_cast_fp16")]; tensor var_3971_to_fp16 = const()[name = tensor("op_3971_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_363_cast_fp16 = mul(x = var_3970_cast_fp16, y = var_3971_to_fp16)[name = tensor("aw_chunk_363_cast_fp16")]; tensor var_3974_equation_0 = const()[name = tensor("op_3974_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_3974_cast_fp16 = einsum(equation = var_3974_equation_0, values = (var_3844_cast_fp16, var_3756_cast_fp16))[name = tensor("op_3974_cast_fp16")]; tensor var_3975_to_fp16 = const()[name = tensor("op_3975_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_365_cast_fp16 = mul(x = var_3974_cast_fp16, y = var_3975_to_fp16)[name = tensor("aw_chunk_365_cast_fp16")]; tensor var_3978_equation_0 = const()[name = tensor("op_3978_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_3978_cast_fp16 = einsum(equation = var_3978_equation_0, values = (var_3844_cast_fp16, var_3763_cast_fp16))[name = tensor("op_3978_cast_fp16")]; tensor var_3979_to_fp16 = const()[name = tensor("op_3979_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_367_cast_fp16 = mul(x = var_3978_cast_fp16, y = var_3979_to_fp16)[name = tensor("aw_chunk_367_cast_fp16")]; tensor var_3982_equation_0 = const()[name = tensor("op_3982_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_3982_cast_fp16 = einsum(equation = var_3982_equation_0, values = (var_3848_cast_fp16, var_3770_cast_fp16))[name = tensor("op_3982_cast_fp16")]; tensor var_3983_to_fp16 = const()[name = tensor("op_3983_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_369_cast_fp16 = mul(x = var_3982_cast_fp16, y = var_3983_to_fp16)[name = tensor("aw_chunk_369_cast_fp16")]; tensor var_3986_equation_0 = const()[name = tensor("op_3986_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_3986_cast_fp16 = einsum(equation = var_3986_equation_0, values = (var_3848_cast_fp16, var_3777_cast_fp16))[name = tensor("op_3986_cast_fp16")]; tensor var_3987_to_fp16 = const()[name = tensor("op_3987_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_371_cast_fp16 = mul(x = var_3986_cast_fp16, y = var_3987_to_fp16)[name = tensor("aw_chunk_371_cast_fp16")]; tensor var_3990_equation_0 = const()[name = tensor("op_3990_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_3990_cast_fp16 = einsum(equation = var_3990_equation_0, values = (var_3848_cast_fp16, var_3784_cast_fp16))[name = tensor("op_3990_cast_fp16")]; tensor var_3991_to_fp16 = const()[name = tensor("op_3991_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_373_cast_fp16 = mul(x = var_3990_cast_fp16, y = var_3991_to_fp16)[name = tensor("aw_chunk_373_cast_fp16")]; tensor var_3994_equation_0 = const()[name = tensor("op_3994_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_3994_cast_fp16 = einsum(equation = var_3994_equation_0, values = (var_3848_cast_fp16, var_3791_cast_fp16))[name = tensor("op_3994_cast_fp16")]; tensor var_3995_to_fp16 = const()[name = tensor("op_3995_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_375_cast_fp16 = mul(x = var_3994_cast_fp16, y = var_3995_to_fp16)[name = tensor("aw_chunk_375_cast_fp16")]; tensor var_3998_equation_0 = const()[name = tensor("op_3998_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_3998_cast_fp16 = einsum(equation = var_3998_equation_0, values = (var_3852_cast_fp16, var_3798_cast_fp16))[name = tensor("op_3998_cast_fp16")]; tensor var_3999_to_fp16 = const()[name = tensor("op_3999_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_377_cast_fp16 = mul(x = var_3998_cast_fp16, y = var_3999_to_fp16)[name = tensor("aw_chunk_377_cast_fp16")]; tensor var_4002_equation_0 = const()[name = tensor("op_4002_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_4002_cast_fp16 = einsum(equation = var_4002_equation_0, values = (var_3852_cast_fp16, var_3805_cast_fp16))[name = tensor("op_4002_cast_fp16")]; tensor var_4003_to_fp16 = const()[name = tensor("op_4003_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_379_cast_fp16 = mul(x = var_4002_cast_fp16, y = var_4003_to_fp16)[name = tensor("aw_chunk_379_cast_fp16")]; tensor var_4006_equation_0 = const()[name = tensor("op_4006_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_4006_cast_fp16 = einsum(equation = var_4006_equation_0, values = (var_3852_cast_fp16, var_3812_cast_fp16))[name = tensor("op_4006_cast_fp16")]; tensor var_4007_to_fp16 = const()[name = tensor("op_4007_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_381_cast_fp16 = mul(x = var_4006_cast_fp16, y = var_4007_to_fp16)[name = tensor("aw_chunk_381_cast_fp16")]; tensor var_4010_equation_0 = const()[name = tensor("op_4010_equation_0"), val = tensor("bkhc,bchq->bkhq")]; tensor var_4010_cast_fp16 = einsum(equation = var_4010_equation_0, values = (var_3852_cast_fp16, var_3819_cast_fp16))[name = tensor("op_4010_cast_fp16")]; tensor var_4011_to_fp16 = const()[name = tensor("op_4011_to_fp16"), val = tensor(0x1p-3)]; tensor aw_chunk_cast_fp16 = mul(x = var_4010_cast_fp16, y = var_4011_to_fp16)[name = tensor("aw_chunk_cast_fp16")]; tensor var_4013_cast_fp16 = softmax(axis = var_3510, x = aw_chunk_321_cast_fp16)[name = tensor("op_4013_cast_fp16")]; tensor var_4014_cast_fp16 = softmax(axis = var_3510, x = aw_chunk_323_cast_fp16)[name = tensor("op_4014_cast_fp16")]; tensor var_4015_cast_fp16 = softmax(axis = var_3510, x = aw_chunk_325_cast_fp16)[name = tensor("op_4015_cast_fp16")]; tensor var_4016_cast_fp16 = softmax(axis = var_3510, x = aw_chunk_327_cast_fp16)[name = tensor("op_4016_cast_fp16")]; tensor var_4017_cast_fp16 = softmax(axis = var_3510, x = aw_chunk_329_cast_fp16)[name = tensor("op_4017_cast_fp16")]; tensor var_4018_cast_fp16 = softmax(axis = var_3510, x = aw_chunk_331_cast_fp16)[name = tensor("op_4018_cast_fp16")]; tensor var_4019_cast_fp16 = softmax(axis = var_3510, x = aw_chunk_333_cast_fp16)[name = tensor("op_4019_cast_fp16")]; tensor var_4020_cast_fp16 = softmax(axis = var_3510, x = aw_chunk_335_cast_fp16)[name = tensor("op_4020_cast_fp16")]; tensor var_4021_cast_fp16 = softmax(axis = var_3510, x = aw_chunk_337_cast_fp16)[name = tensor("op_4021_cast_fp16")]; tensor var_4022_cast_fp16 = softmax(axis = var_3510, x = aw_chunk_339_cast_fp16)[name = tensor("op_4022_cast_fp16")]; tensor var_4023_cast_fp16 = softmax(axis = var_3510, x = aw_chunk_341_cast_fp16)[name = tensor("op_4023_cast_fp16")]; tensor var_4024_cast_fp16 = softmax(axis = var_3510, x = aw_chunk_343_cast_fp16)[name = tensor("op_4024_cast_fp16")]; tensor var_4025_cast_fp16 = softmax(axis = var_3510, x = aw_chunk_345_cast_fp16)[name = tensor("op_4025_cast_fp16")]; tensor var_4026_cast_fp16 = softmax(axis = var_3510, x = aw_chunk_347_cast_fp16)[name = tensor("op_4026_cast_fp16")]; tensor var_4027_cast_fp16 = softmax(axis = var_3510, x = aw_chunk_349_cast_fp16)[name = tensor("op_4027_cast_fp16")]; tensor var_4028_cast_fp16 = softmax(axis = var_3510, x = aw_chunk_351_cast_fp16)[name = tensor("op_4028_cast_fp16")]; tensor var_4029_cast_fp16 = softmax(axis = var_3510, x = aw_chunk_353_cast_fp16)[name = tensor("op_4029_cast_fp16")]; tensor var_4030_cast_fp16 = softmax(axis = var_3510, x = aw_chunk_355_cast_fp16)[name = tensor("op_4030_cast_fp16")]; tensor var_4031_cast_fp16 = softmax(axis = var_3510, x = aw_chunk_357_cast_fp16)[name = tensor("op_4031_cast_fp16")]; tensor var_4032_cast_fp16 = softmax(axis = var_3510, x = aw_chunk_359_cast_fp16)[name = tensor("op_4032_cast_fp16")]; tensor var_4033_cast_fp16 = softmax(axis = var_3510, x = aw_chunk_361_cast_fp16)[name = tensor("op_4033_cast_fp16")]; tensor var_4034_cast_fp16 = softmax(axis = var_3510, x = aw_chunk_363_cast_fp16)[name = tensor("op_4034_cast_fp16")]; tensor var_4035_cast_fp16 = softmax(axis = var_3510, x = aw_chunk_365_cast_fp16)[name = tensor("op_4035_cast_fp16")]; tensor var_4036_cast_fp16 = softmax(axis = var_3510, x = aw_chunk_367_cast_fp16)[name = tensor("op_4036_cast_fp16")]; tensor var_4037_cast_fp16 = softmax(axis = var_3510, x = aw_chunk_369_cast_fp16)[name = tensor("op_4037_cast_fp16")]; tensor var_4038_cast_fp16 = softmax(axis = var_3510, x = aw_chunk_371_cast_fp16)[name = tensor("op_4038_cast_fp16")]; tensor var_4039_cast_fp16 = softmax(axis = var_3510, x = aw_chunk_373_cast_fp16)[name = tensor("op_4039_cast_fp16")]; tensor var_4040_cast_fp16 = softmax(axis = var_3510, x = aw_chunk_375_cast_fp16)[name = tensor("op_4040_cast_fp16")]; tensor var_4041_cast_fp16 = softmax(axis = var_3510, x = aw_chunk_377_cast_fp16)[name = tensor("op_4041_cast_fp16")]; tensor var_4042_cast_fp16 = softmax(axis = var_3510, x = aw_chunk_379_cast_fp16)[name = tensor("op_4042_cast_fp16")]; tensor var_4043_cast_fp16 = softmax(axis = var_3510, x = aw_chunk_381_cast_fp16)[name = tensor("op_4043_cast_fp16")]; tensor var_4044_cast_fp16 = softmax(axis = var_3510, x = aw_chunk_cast_fp16)[name = tensor("op_4044_cast_fp16")]; tensor var_4046_equation_0 = const()[name = tensor("op_4046_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_4046_cast_fp16 = einsum(equation = var_4046_equation_0, values = (var_3854_cast_fp16, var_4013_cast_fp16))[name = tensor("op_4046_cast_fp16")]; tensor var_4048_equation_0 = const()[name = tensor("op_4048_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_4048_cast_fp16 = einsum(equation = var_4048_equation_0, values = (var_3854_cast_fp16, var_4014_cast_fp16))[name = tensor("op_4048_cast_fp16")]; tensor var_4050_equation_0 = const()[name = tensor("op_4050_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_4050_cast_fp16 = einsum(equation = var_4050_equation_0, values = (var_3854_cast_fp16, var_4015_cast_fp16))[name = tensor("op_4050_cast_fp16")]; tensor var_4052_equation_0 = const()[name = tensor("op_4052_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_4052_cast_fp16 = einsum(equation = var_4052_equation_0, values = (var_3854_cast_fp16, var_4016_cast_fp16))[name = tensor("op_4052_cast_fp16")]; tensor var_4054_equation_0 = const()[name = tensor("op_4054_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_4054_cast_fp16 = einsum(equation = var_4054_equation_0, values = (var_3858_cast_fp16, var_4017_cast_fp16))[name = tensor("op_4054_cast_fp16")]; tensor var_4056_equation_0 = const()[name = tensor("op_4056_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_4056_cast_fp16 = einsum(equation = var_4056_equation_0, values = (var_3858_cast_fp16, var_4018_cast_fp16))[name = tensor("op_4056_cast_fp16")]; tensor var_4058_equation_0 = const()[name = tensor("op_4058_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_4058_cast_fp16 = einsum(equation = var_4058_equation_0, values = (var_3858_cast_fp16, var_4019_cast_fp16))[name = tensor("op_4058_cast_fp16")]; tensor var_4060_equation_0 = const()[name = tensor("op_4060_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_4060_cast_fp16 = einsum(equation = var_4060_equation_0, values = (var_3858_cast_fp16, var_4020_cast_fp16))[name = tensor("op_4060_cast_fp16")]; tensor var_4062_equation_0 = const()[name = tensor("op_4062_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_4062_cast_fp16 = einsum(equation = var_4062_equation_0, values = (var_3862_cast_fp16, var_4021_cast_fp16))[name = tensor("op_4062_cast_fp16")]; tensor var_4064_equation_0 = const()[name = tensor("op_4064_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_4064_cast_fp16 = einsum(equation = var_4064_equation_0, values = (var_3862_cast_fp16, var_4022_cast_fp16))[name = tensor("op_4064_cast_fp16")]; tensor var_4066_equation_0 = const()[name = tensor("op_4066_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_4066_cast_fp16 = einsum(equation = var_4066_equation_0, values = (var_3862_cast_fp16, var_4023_cast_fp16))[name = tensor("op_4066_cast_fp16")]; tensor var_4068_equation_0 = const()[name = tensor("op_4068_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_4068_cast_fp16 = einsum(equation = var_4068_equation_0, values = (var_3862_cast_fp16, var_4024_cast_fp16))[name = tensor("op_4068_cast_fp16")]; tensor var_4070_equation_0 = const()[name = tensor("op_4070_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_4070_cast_fp16 = einsum(equation = var_4070_equation_0, values = (var_3866_cast_fp16, var_4025_cast_fp16))[name = tensor("op_4070_cast_fp16")]; tensor var_4072_equation_0 = const()[name = tensor("op_4072_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_4072_cast_fp16 = einsum(equation = var_4072_equation_0, values = (var_3866_cast_fp16, var_4026_cast_fp16))[name = tensor("op_4072_cast_fp16")]; tensor var_4074_equation_0 = const()[name = tensor("op_4074_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_4074_cast_fp16 = einsum(equation = var_4074_equation_0, values = (var_3866_cast_fp16, var_4027_cast_fp16))[name = tensor("op_4074_cast_fp16")]; tensor var_4076_equation_0 = const()[name = tensor("op_4076_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_4076_cast_fp16 = einsum(equation = var_4076_equation_0, values = (var_3866_cast_fp16, var_4028_cast_fp16))[name = tensor("op_4076_cast_fp16")]; tensor var_4078_equation_0 = const()[name = tensor("op_4078_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_4078_cast_fp16 = einsum(equation = var_4078_equation_0, values = (var_3870_cast_fp16, var_4029_cast_fp16))[name = tensor("op_4078_cast_fp16")]; tensor var_4080_equation_0 = const()[name = tensor("op_4080_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_4080_cast_fp16 = einsum(equation = var_4080_equation_0, values = (var_3870_cast_fp16, var_4030_cast_fp16))[name = tensor("op_4080_cast_fp16")]; tensor var_4082_equation_0 = const()[name = tensor("op_4082_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_4082_cast_fp16 = einsum(equation = var_4082_equation_0, values = (var_3870_cast_fp16, var_4031_cast_fp16))[name = tensor("op_4082_cast_fp16")]; tensor var_4084_equation_0 = const()[name = tensor("op_4084_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_4084_cast_fp16 = einsum(equation = var_4084_equation_0, values = (var_3870_cast_fp16, var_4032_cast_fp16))[name = tensor("op_4084_cast_fp16")]; tensor var_4086_equation_0 = const()[name = tensor("op_4086_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_4086_cast_fp16 = einsum(equation = var_4086_equation_0, values = (var_3874_cast_fp16, var_4033_cast_fp16))[name = tensor("op_4086_cast_fp16")]; tensor var_4088_equation_0 = const()[name = tensor("op_4088_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_4088_cast_fp16 = einsum(equation = var_4088_equation_0, values = (var_3874_cast_fp16, var_4034_cast_fp16))[name = tensor("op_4088_cast_fp16")]; tensor var_4090_equation_0 = const()[name = tensor("op_4090_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_4090_cast_fp16 = einsum(equation = var_4090_equation_0, values = (var_3874_cast_fp16, var_4035_cast_fp16))[name = tensor("op_4090_cast_fp16")]; tensor var_4092_equation_0 = const()[name = tensor("op_4092_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_4092_cast_fp16 = einsum(equation = var_4092_equation_0, values = (var_3874_cast_fp16, var_4036_cast_fp16))[name = tensor("op_4092_cast_fp16")]; tensor var_4094_equation_0 = const()[name = tensor("op_4094_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_4094_cast_fp16 = einsum(equation = var_4094_equation_0, values = (var_3878_cast_fp16, var_4037_cast_fp16))[name = tensor("op_4094_cast_fp16")]; tensor var_4096_equation_0 = const()[name = tensor("op_4096_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_4096_cast_fp16 = einsum(equation = var_4096_equation_0, values = (var_3878_cast_fp16, var_4038_cast_fp16))[name = tensor("op_4096_cast_fp16")]; tensor var_4098_equation_0 = const()[name = tensor("op_4098_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_4098_cast_fp16 = einsum(equation = var_4098_equation_0, values = (var_3878_cast_fp16, var_4039_cast_fp16))[name = tensor("op_4098_cast_fp16")]; tensor var_4100_equation_0 = const()[name = tensor("op_4100_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_4100_cast_fp16 = einsum(equation = var_4100_equation_0, values = (var_3878_cast_fp16, var_4040_cast_fp16))[name = tensor("op_4100_cast_fp16")]; tensor var_4102_equation_0 = const()[name = tensor("op_4102_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_4102_cast_fp16 = einsum(equation = var_4102_equation_0, values = (var_3882_cast_fp16, var_4041_cast_fp16))[name = tensor("op_4102_cast_fp16")]; tensor var_4104_equation_0 = const()[name = tensor("op_4104_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_4104_cast_fp16 = einsum(equation = var_4104_equation_0, values = (var_3882_cast_fp16, var_4042_cast_fp16))[name = tensor("op_4104_cast_fp16")]; tensor var_4106_equation_0 = const()[name = tensor("op_4106_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_4106_cast_fp16 = einsum(equation = var_4106_equation_0, values = (var_3882_cast_fp16, var_4043_cast_fp16))[name = tensor("op_4106_cast_fp16")]; tensor var_4108_equation_0 = const()[name = tensor("op_4108_equation_0"), val = tensor("bchk,bkhq->bchq")]; tensor var_4108_cast_fp16 = einsum(equation = var_4108_equation_0, values = (var_3882_cast_fp16, var_4044_cast_fp16))[name = tensor("op_4108_cast_fp16")]; tensor var_4110_interleave_0 = const()[name = tensor("op_4110_interleave_0"), val = tensor(false)]; tensor var_4110_cast_fp16 = concat(axis = var_3497, interleave = var_4110_interleave_0, values = (var_4046_cast_fp16, var_4048_cast_fp16, var_4050_cast_fp16, var_4052_cast_fp16))[name = tensor("op_4110_cast_fp16")]; tensor var_4112_interleave_0 = const()[name = tensor("op_4112_interleave_0"), val = tensor(false)]; tensor var_4112_cast_fp16 = concat(axis = var_3497, interleave = var_4112_interleave_0, values = (var_4054_cast_fp16, var_4056_cast_fp16, var_4058_cast_fp16, var_4060_cast_fp16))[name = tensor("op_4112_cast_fp16")]; tensor var_4114_interleave_0 = const()[name = tensor("op_4114_interleave_0"), val = tensor(false)]; tensor var_4114_cast_fp16 = concat(axis = var_3497, interleave = var_4114_interleave_0, values = (var_4062_cast_fp16, var_4064_cast_fp16, var_4066_cast_fp16, var_4068_cast_fp16))[name = tensor("op_4114_cast_fp16")]; tensor var_4116_interleave_0 = const()[name = tensor("op_4116_interleave_0"), val = tensor(false)]; tensor var_4116_cast_fp16 = concat(axis = var_3497, interleave = var_4116_interleave_0, values = (var_4070_cast_fp16, var_4072_cast_fp16, var_4074_cast_fp16, var_4076_cast_fp16))[name = tensor("op_4116_cast_fp16")]; tensor var_4118_interleave_0 = const()[name = tensor("op_4118_interleave_0"), val = tensor(false)]; tensor var_4118_cast_fp16 = concat(axis = var_3497, interleave = var_4118_interleave_0, values = (var_4078_cast_fp16, var_4080_cast_fp16, var_4082_cast_fp16, var_4084_cast_fp16))[name = tensor("op_4118_cast_fp16")]; tensor var_4120_interleave_0 = const()[name = tensor("op_4120_interleave_0"), val = tensor(false)]; tensor var_4120_cast_fp16 = concat(axis = var_3497, interleave = var_4120_interleave_0, values = (var_4086_cast_fp16, var_4088_cast_fp16, var_4090_cast_fp16, var_4092_cast_fp16))[name = tensor("op_4120_cast_fp16")]; tensor var_4122_interleave_0 = const()[name = tensor("op_4122_interleave_0"), val = tensor(false)]; tensor var_4122_cast_fp16 = concat(axis = var_3497, interleave = var_4122_interleave_0, values = (var_4094_cast_fp16, var_4096_cast_fp16, var_4098_cast_fp16, var_4100_cast_fp16))[name = tensor("op_4122_cast_fp16")]; tensor var_4124_interleave_0 = const()[name = tensor("op_4124_interleave_0"), val = tensor(false)]; tensor var_4124_cast_fp16 = concat(axis = var_3497, interleave = var_4124_interleave_0, values = (var_4102_cast_fp16, var_4104_cast_fp16, var_4106_cast_fp16, var_4108_cast_fp16))[name = tensor("op_4124_cast_fp16")]; tensor input_41_interleave_0 = const()[name = tensor("input_41_interleave_0"), val = tensor(false)]; tensor input_41_cast_fp16 = concat(axis = var_3510, interleave = input_41_interleave_0, values = (var_4110_cast_fp16, var_4112_cast_fp16, var_4114_cast_fp16, var_4116_cast_fp16, var_4118_cast_fp16, var_4120_cast_fp16, var_4122_cast_fp16, var_4124_cast_fp16))[name = tensor("input_41_cast_fp16")]; tensor var_4129 = const()[name = tensor("op_4129"), val = tensor([1, 1])]; tensor var_4131 = const()[name = tensor("op_4131"), 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_5_self_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_5_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(36460160)))]; tensor layers_5_self_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_5_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(36984512)))]; tensor obj_cast_fp16 = conv(bias = layers_5_self_attn_o_proj_bias_to_fp16, dilations = var_4131, groups = var_3510, pad = obj_pad_0, pad_type = obj_pad_type_0, strides = var_4129, weight = layers_5_self_attn_o_proj_weight_to_fp16, x = input_41_cast_fp16)[name = tensor("obj_cast_fp16")]; tensor inputs_23_cast_fp16 = add(x = inputs_21_cast_fp16, y = obj_cast_fp16)[name = tensor("inputs_23_cast_fp16")]; tensor var_4137 = const()[name = tensor("op_4137"), val = tensor([1])]; tensor channels_mean_23_cast_fp16 = reduce_mean(axes = var_4137, keep_dims = var_3511, x = inputs_23_cast_fp16)[name = tensor("channels_mean_23_cast_fp16")]; tensor zero_mean_23_cast_fp16 = sub(x = inputs_23_cast_fp16, y = channels_mean_23_cast_fp16)[name = tensor("zero_mean_23_cast_fp16")]; tensor zero_mean_sq_23_cast_fp16 = mul(x = zero_mean_23_cast_fp16, y = zero_mean_23_cast_fp16)[name = tensor("zero_mean_sq_23_cast_fp16")]; tensor var_4141 = const()[name = tensor("op_4141"), val = tensor([1])]; tensor var_4142_cast_fp16 = reduce_mean(axes = var_4141, keep_dims = var_3511, x = zero_mean_sq_23_cast_fp16)[name = tensor("op_4142_cast_fp16")]; tensor var_4143_to_fp16 = const()[name = tensor("op_4143_to_fp16"), val = tensor(0x1.5p-17)]; tensor var_4144_cast_fp16 = add(x = var_4142_cast_fp16, y = var_4143_to_fp16)[name = tensor("op_4144_cast_fp16")]; tensor denom_23_epsilon_0_to_fp16 = const()[name = tensor("denom_23_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; tensor denom_23_cast_fp16 = rsqrt(epsilon = denom_23_epsilon_0_to_fp16, x = var_4144_cast_fp16)[name = tensor("denom_23_cast_fp16")]; tensor out_23_cast_fp16 = mul(x = zero_mean_23_cast_fp16, y = denom_23_cast_fp16)[name = tensor("out_23_cast_fp16")]; tensor input_43_gamma_0_to_fp16 = const()[name = tensor("input_43_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(36985600)))]; tensor input_43_beta_0_to_fp16 = const()[name = tensor("input_43_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(36986688)))]; tensor input_43_epsilon_0_to_fp16 = const()[name = tensor("input_43_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; tensor input_43_cast_fp16 = batch_norm(beta = input_43_beta_0_to_fp16, epsilon = input_43_epsilon_0_to_fp16, gamma = input_43_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_23_cast_fp16)[name = tensor("input_43_cast_fp16")]; tensor var_4155 = const()[name = tensor("op_4155"), val = tensor([1, 1])]; tensor var_4157 = const()[name = tensor("op_4157"), val = tensor([1, 1])]; tensor input_45_pad_type_0 = const()[name = tensor("input_45_pad_type_0"), val = tensor("custom")]; tensor input_45_pad_0 = const()[name = tensor("input_45_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_5_fc1_weight_to_fp16 = const()[name = tensor("layers_5_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(36987776)))]; tensor layers_5_fc1_bias_to_fp16 = const()[name = tensor("layers_5_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39084992)))]; tensor input_45_cast_fp16 = conv(bias = layers_5_fc1_bias_to_fp16, dilations = var_4157, groups = var_3510, pad = input_45_pad_0, pad_type = input_45_pad_type_0, strides = var_4155, weight = layers_5_fc1_weight_to_fp16, x = input_43_cast_fp16)[name = tensor("input_45_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_45_cast_fp16)[name = tensor("input_cast_fp16")]; tensor var_4163 = const()[name = tensor("op_4163"), val = tensor([1, 1])]; tensor var_4165 = const()[name = tensor("op_4165"), 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_5_fc2_weight_to_fp16 = const()[name = tensor("layers_5_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39089152)))]; tensor layers_5_fc2_bias_to_fp16 = const()[name = tensor("layers_5_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41186368)))]; tensor hidden_states_cast_fp16 = conv(bias = layers_5_fc2_bias_to_fp16, dilations = var_4165, groups = var_3510, pad = hidden_states_pad_0, pad_type = hidden_states_pad_type_0, strides = var_4163, weight = layers_5_fc2_weight_to_fp16, x = input_cast_fp16)[name = tensor("hidden_states_cast_fp16")]; tensor inputs_cast_fp16 = add(x = inputs_23_cast_fp16, y = hidden_states_cast_fp16)[name = tensor("inputs_cast_fp16")]; tensor var_4171 = const()[name = tensor("op_4171"), val = tensor(true)]; tensor var_4175 = const()[name = tensor("op_4175"), val = tensor([1])]; tensor channels_mean_cast_fp16 = reduce_mean(axes = var_4175, keep_dims = var_4171, 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_4179 = const()[name = tensor("op_4179"), val = tensor([1])]; tensor var_4180_cast_fp16 = reduce_mean(axes = var_4179, keep_dims = var_4171, x = zero_mean_sq_cast_fp16)[name = tensor("op_4180_cast_fp16")]; tensor var_4181_to_fp16 = const()[name = tensor("op_4181_to_fp16"), val = tensor(0x1.5p-17)]; tensor var_4182_cast_fp16 = add(x = var_4180_cast_fp16, y = var_4181_to_fp16)[name = tensor("op_4182_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_4182_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(41187456)))]; 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(41188544)))]; 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); }