| program(1.0) | |
| [buildInfo = dict<tensor<string, []>, tensor<string, []>>({{"coremlc-component-MIL", "3510.2.1"}, {"coremlc-version", "3500.32.1"}, {"coremltools-component-torch", "2.9.1"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0"}})] | |
| { | |
| func main<ios15>(tensor<fp32, [1, 3, 224, 224]> image) { | |
| tensor<fp32, []> image__scaled___y_0 = const()[name = tensor<string, []>("image__scaled___y_0"), val = tensor<fp32, []>(0x1.de5c7ap-7)]; | |
| tensor<fp32, [1, 3, 224, 224]> image__scaled__ = mul(x = image, y = image__scaled___y_0)[name = tensor<string, []>("image__scaled__")]; | |
| tensor<fp32, [1, 3, 1, 1]> image__biased___y_0 = const()[name = tensor<string, []>("image__biased___y_0"), val = tensor<fp32, [1, 3, 1, 1]>([[[[-0x1.cad1b8p+0]], [[-0x1.c0897p+0]], [[-0x1.7aefaep+0]]]])]; | |
| tensor<fp32, [1, 3, 224, 224]> image__biased__ = add(x = image__scaled__, y = image__biased___y_0)[name = tensor<string, []>("image__biased__")]; | |
| tensor<int32, []> var_20 = const()[name = tensor<string, []>("op_20"), val = tensor<int32, []>(1)]; | |
| tensor<string, []> x_1_pad_type_0 = const()[name = tensor<string, []>("x_1_pad_type_0"), val = tensor<string, []>("valid")]; | |
| tensor<int32, [2]> x_1_strides_0 = const()[name = tensor<string, []>("x_1_strides_0"), val = tensor<int32, [2]>([32, 32])]; | |
| tensor<int32, [4]> x_1_pad_0 = const()[name = tensor<string, []>("x_1_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; | |
| tensor<int32, [2]> x_1_dilations_0 = const()[name = tensor<string, []>("x_1_dilations_0"), val = tensor<int32, [2]>([1, 1])]; | |
| tensor<int32, []> x_1_groups_0 = const()[name = tensor<string, []>("x_1_groups_0"), val = tensor<int32, []>(1)]; | |
| tensor<string, []> image_to_fp16_dtype_0 = const()[name = tensor<string, []>("image_to_fp16_dtype_0"), val = tensor<string, []>("fp16")]; | |
| tensor<fp16, [768, 3, 32, 32]> model_visual_conv1_weight_to_fp16 = const()[name = tensor<string, []>("model_visual_conv1_weight_to_fp16"), val = tensor<fp16, [768, 3, 32, 32]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(64)))]; | |
| tensor<fp16, [1, 3, 224, 224]> image_to_fp16 = cast(dtype = image_to_fp16_dtype_0, x = image__biased__)[name = tensor<string, []>("cast_135")]; | |
| tensor<fp16, [1, 768, 7, 7]> x_1_cast_fp16 = conv(dilations = x_1_dilations_0, groups = x_1_groups_0, pad = x_1_pad_0, pad_type = x_1_pad_type_0, strides = x_1_strides_0, weight = model_visual_conv1_weight_to_fp16, x = image_to_fp16)[name = tensor<string, []>("x_1_cast_fp16")]; | |
| tensor<int32, [3]> var_38 = const()[name = tensor<string, []>("op_38"), val = tensor<int32, [3]>([1, 768, -1])]; | |
| tensor<fp16, [1, 768, 49]> x_3_cast_fp16 = reshape(shape = var_38, x = x_1_cast_fp16)[name = tensor<string, []>("x_3_cast_fp16")]; | |
| tensor<int32, [3]> var_40 = const()[name = tensor<string, []>("op_40"), val = tensor<int32, [3]>([0, 2, 1])]; | |
| tensor<bool, []> x_7_interleave_0 = const()[name = tensor<string, []>("x_7_interleave_0"), val = tensor<bool, []>(false)]; | |
| tensor<fp16, [1, 1, 768]> const_3_to_fp16 = const()[name = tensor<string, []>("const_3_to_fp16"), val = tensor<fp16, [1, 1, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4718720)))]; | |
| tensor<fp16, [1, 49, 768]> x_5_cast_fp16 = transpose(perm = var_40, x = x_3_cast_fp16)[name = tensor<string, []>("transpose_85")]; | |
| tensor<fp16, [1, 50, 768]> x_7_cast_fp16 = concat(axis = var_20, interleave = x_7_interleave_0, values = (const_3_to_fp16, x_5_cast_fp16))[name = tensor<string, []>("x_7_cast_fp16")]; | |
| tensor<fp16, [50, 768]> const_4_to_fp16 = const()[name = tensor<string, []>("const_4_to_fp16"), val = tensor<fp16, [50, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4720320)))]; | |
| tensor<fp16, [1, 50, 768]> input_1_cast_fp16 = add(x = x_7_cast_fp16, y = const_4_to_fp16)[name = tensor<string, []>("input_1_cast_fp16")]; | |
| tensor<int32, [1]> x_9_axes_0 = const()[name = tensor<string, []>("x_9_axes_0"), val = tensor<int32, [1]>([-1])]; | |
| tensor<fp16, [768]> model_visual_ln_pre_weight_to_fp16 = const()[name = tensor<string, []>("model_visual_ln_pre_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4797184)))]; | |
| tensor<fp16, [768]> model_visual_ln_pre_bias_to_fp16 = const()[name = tensor<string, []>("model_visual_ln_pre_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4798784)))]; | |
| tensor<fp16, []> var_13_to_fp16 = const()[name = tensor<string, []>("op_13_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)]; | |
| tensor<fp16, [1, 50, 768]> x_9_cast_fp16 = layer_norm(axes = x_9_axes_0, beta = model_visual_ln_pre_bias_to_fp16, epsilon = var_13_to_fp16, gamma = model_visual_ln_pre_weight_to_fp16, x = input_1_cast_fp16)[name = tensor<string, []>("x_9_cast_fp16")]; | |
| tensor<int32, [1]> x_11_axes_0 = const()[name = tensor<string, []>("x_11_axes_0"), val = tensor<int32, [1]>([-1])]; | |
| tensor<fp16, [768]> model_visual_transformer_resblocks_0_ln_1_weight_to_fp16 = const()[name = tensor<string, []>("model_visual_transformer_resblocks_0_ln_1_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4800384)))]; | |
| tensor<fp16, [768]> model_visual_transformer_resblocks_0_ln_1_bias_to_fp16 = const()[name = tensor<string, []>("model_visual_transformer_resblocks_0_ln_1_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4801984)))]; | |
| tensor<fp16, [1, 50, 768]> x_11_cast_fp16 = layer_norm(axes = x_11_axes_0, beta = model_visual_transformer_resblocks_0_ln_1_bias_to_fp16, epsilon = var_13_to_fp16, gamma = model_visual_transformer_resblocks_0_ln_1_weight_to_fp16, x = x_9_cast_fp16)[name = tensor<string, []>("x_11_cast_fp16")]; | |
| tensor<int32, [3]> query_3_perm_0 = const()[name = tensor<string, []>("query_3_perm_0"), val = tensor<int32, [3]>([1, 0, 2])]; | |
| tensor<fp16, [2304, 768]> model_visual_transformer_resblocks_0_attn_in_proj_weight_to_fp16 = const()[name = tensor<string, []>("model_visual_transformer_resblocks_0_attn_in_proj_weight_to_fp16"), val = tensor<fp16, [2304, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4803584)))]; | |
| tensor<fp16, [2304]> model_visual_transformer_resblocks_0_attn_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_visual_transformer_resblocks_0_attn_in_proj_bias_to_fp16"), val = tensor<fp16, [2304]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8342592)))]; | |
| tensor<fp16, [50, 1, 768]> query_3_cast_fp16 = transpose(perm = query_3_perm_0, x = x_11_cast_fp16)[name = tensor<string, []>("transpose_84")]; | |
| tensor<fp16, [50, 1, 2304]> linear_0_cast_fp16 = linear(bias = model_visual_transformer_resblocks_0_attn_in_proj_bias_to_fp16, weight = model_visual_transformer_resblocks_0_attn_in_proj_weight_to_fp16, x = query_3_cast_fp16)[name = tensor<string, []>("linear_0_cast_fp16")]; | |
| tensor<int32, [4]> concat_1 = const()[name = tensor<string, []>("concat_1"), val = tensor<int32, [4]>([50, 1, 3, 768])]; | |
| tensor<fp16, [50, 1, 3, 768]> var_113_cast_fp16 = reshape(shape = concat_1, x = linear_0_cast_fp16)[name = tensor<string, []>("op_113_cast_fp16")]; | |
| tensor<int32, [1]> var_114_axes_0 = const()[name = tensor<string, []>("op_114_axes_0"), val = tensor<int32, [1]>([0])]; | |
| tensor<fp16, [1, 50, 1, 3, 768]> var_114_cast_fp16 = expand_dims(axes = var_114_axes_0, x = var_113_cast_fp16)[name = tensor<string, []>("op_114_cast_fp16")]; | |
| tensor<int32, [5]> var_115_perm_0 = const()[name = tensor<string, []>("op_115_perm_0"), val = tensor<int32, [5]>([-2, 1, 2, 0, 4])]; | |
| tensor<int32, [1]> var_116_axes_0 = const()[name = tensor<string, []>("op_116_axes_0"), val = tensor<int32, [1]>([-2])]; | |
| tensor<fp16, [3, 50, 1, 1, 768]> var_115_cast_fp16 = transpose(perm = var_115_perm_0, x = var_114_cast_fp16)[name = tensor<string, []>("transpose_83")]; | |
| tensor<fp16, [3, 50, 1, 768]> var_116_cast_fp16 = squeeze(axes = var_116_axes_0, x = var_115_cast_fp16)[name = tensor<string, []>("op_116_cast_fp16")]; | |
| tensor<int32, [4]> q_1_begin_0 = const()[name = tensor<string, []>("q_1_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; | |
| tensor<int32, [4]> q_1_end_0 = const()[name = tensor<string, []>("q_1_end_0"), val = tensor<int32, [4]>([1, 50, 1, 768])]; | |
| tensor<bool, [4]> q_1_end_mask_0 = const()[name = tensor<string, []>("q_1_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])]; | |
| tensor<bool, [4]> q_1_squeeze_mask_0 = const()[name = tensor<string, []>("q_1_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])]; | |
| tensor<fp16, [50, 1, 768]> q_1_cast_fp16 = slice_by_index(begin = q_1_begin_0, end = q_1_end_0, end_mask = q_1_end_mask_0, squeeze_mask = q_1_squeeze_mask_0, x = var_116_cast_fp16)[name = tensor<string, []>("q_1_cast_fp16")]; | |
| tensor<int32, [4]> k_1_begin_0 = const()[name = tensor<string, []>("k_1_begin_0"), val = tensor<int32, [4]>([1, 0, 0, 0])]; | |
| tensor<int32, [4]> k_1_end_0 = const()[name = tensor<string, []>("k_1_end_0"), val = tensor<int32, [4]>([2, 50, 1, 768])]; | |
| tensor<bool, [4]> k_1_end_mask_0 = const()[name = tensor<string, []>("k_1_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])]; | |
| tensor<bool, [4]> k_1_squeeze_mask_0 = const()[name = tensor<string, []>("k_1_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])]; | |
| tensor<fp16, [50, 1, 768]> k_1_cast_fp16 = slice_by_index(begin = k_1_begin_0, end = k_1_end_0, end_mask = k_1_end_mask_0, squeeze_mask = k_1_squeeze_mask_0, x = var_116_cast_fp16)[name = tensor<string, []>("k_1_cast_fp16")]; | |
| tensor<int32, [4]> v_1_begin_0 = const()[name = tensor<string, []>("v_1_begin_0"), val = tensor<int32, [4]>([2, 0, 0, 0])]; | |
| tensor<int32, [4]> v_1_end_0 = const()[name = tensor<string, []>("v_1_end_0"), val = tensor<int32, [4]>([3, 50, 1, 768])]; | |
| tensor<bool, [4]> v_1_end_mask_0 = const()[name = tensor<string, []>("v_1_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])]; | |
| tensor<bool, [4]> v_1_squeeze_mask_0 = const()[name = tensor<string, []>("v_1_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])]; | |
| tensor<fp16, [50, 1, 768]> v_1_cast_fp16 = slice_by_index(begin = v_1_begin_0, end = v_1_end_0, end_mask = v_1_end_mask_0, squeeze_mask = v_1_squeeze_mask_0, x = var_116_cast_fp16)[name = tensor<string, []>("v_1_cast_fp16")]; | |
| tensor<int32, [3]> var_124 = const()[name = tensor<string, []>("op_124"), val = tensor<int32, [3]>([50, 12, 64])]; | |
| tensor<fp16, [50, 12, 64]> var_125_cast_fp16 = reshape(shape = var_124, x = q_1_cast_fp16)[name = tensor<string, []>("op_125_cast_fp16")]; | |
| tensor<int32, [3]> q_3_perm_0 = const()[name = tensor<string, []>("q_3_perm_0"), val = tensor<int32, [3]>([1, 0, 2])]; | |
| tensor<int32, [3]> var_131 = const()[name = tensor<string, []>("op_131"), val = tensor<int32, [3]>([50, 12, 64])]; | |
| tensor<fp16, [50, 12, 64]> var_132_cast_fp16 = reshape(shape = var_131, x = k_1_cast_fp16)[name = tensor<string, []>("op_132_cast_fp16")]; | |
| tensor<int32, [3]> k_3_perm_0 = const()[name = tensor<string, []>("k_3_perm_0"), val = tensor<int32, [3]>([1, 0, 2])]; | |
| tensor<int32, [3]> var_138 = const()[name = tensor<string, []>("op_138"), val = tensor<int32, [3]>([50, 12, 64])]; | |
| tensor<fp16, [50, 12, 64]> var_139_cast_fp16 = reshape(shape = var_138, x = v_1_cast_fp16)[name = tensor<string, []>("op_139_cast_fp16")]; | |
| tensor<int32, [3]> v_3_perm_0 = const()[name = tensor<string, []>("v_3_perm_0"), val = tensor<int32, [3]>([1, 0, 2])]; | |
| tensor<int32, [4]> var_142 = const()[name = tensor<string, []>("op_142"), val = tensor<int32, [4]>([1, 12, 50, 64])]; | |
| tensor<fp16, [12, 50, 64]> q_3_cast_fp16 = transpose(perm = q_3_perm_0, x = var_125_cast_fp16)[name = tensor<string, []>("transpose_82")]; | |
| tensor<fp16, [1, 12, 50, 64]> q_5_cast_fp16 = reshape(shape = var_142, x = q_3_cast_fp16)[name = tensor<string, []>("q_5_cast_fp16")]; | |
| tensor<int32, [4]> var_144 = const()[name = tensor<string, []>("op_144"), val = tensor<int32, [4]>([1, 12, 50, 64])]; | |
| tensor<fp16, [12, 50, 64]> k_3_cast_fp16 = transpose(perm = k_3_perm_0, x = var_132_cast_fp16)[name = tensor<string, []>("transpose_81")]; | |
| tensor<fp16, [1, 12, 50, 64]> k_5_cast_fp16 = reshape(shape = var_144, x = k_3_cast_fp16)[name = tensor<string, []>("k_5_cast_fp16")]; | |
| tensor<int32, [4]> var_146 = const()[name = tensor<string, []>("op_146"), val = tensor<int32, [4]>([1, 12, 50, 64])]; | |
| tensor<fp16, [12, 50, 64]> v_3_cast_fp16 = transpose(perm = v_3_perm_0, x = var_139_cast_fp16)[name = tensor<string, []>("transpose_80")]; | |
| tensor<fp16, [1, 12, 50, 64]> v_5_cast_fp16 = reshape(shape = var_146, x = v_3_cast_fp16)[name = tensor<string, []>("v_5_cast_fp16")]; | |
| tensor<fp16, []> mul_1_y_0_to_fp16 = const()[name = tensor<string, []>("mul_1_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)]; | |
| tensor<fp16, [1, 12, 50, 64]> mul_1_cast_fp16 = mul(x = q_5_cast_fp16, y = mul_1_y_0_to_fp16)[name = tensor<string, []>("mul_1_cast_fp16")]; | |
| tensor<bool, []> matmul_0_transpose_y_0 = const()[name = tensor<string, []>("matmul_0_transpose_y_0"), val = tensor<bool, []>(true)]; | |
| tensor<bool, []> matmul_0_transpose_x_0 = const()[name = tensor<string, []>("matmul_0_transpose_x_0"), val = tensor<bool, []>(false)]; | |
| tensor<fp16, [1, 12, 50, 50]> matmul_0_cast_fp16 = matmul(transpose_x = matmul_0_transpose_x_0, transpose_y = matmul_0_transpose_y_0, x = mul_1_cast_fp16, y = k_5_cast_fp16)[name = tensor<string, []>("matmul_0_cast_fp16")]; | |
| tensor<int32, []> softmax_0_axis_0 = const()[name = tensor<string, []>("softmax_0_axis_0"), val = tensor<int32, []>(-1)]; | |
| tensor<fp16, [1, 12, 50, 50]> softmax_0_cast_fp16 = softmax(axis = softmax_0_axis_0, x = matmul_0_cast_fp16)[name = tensor<string, []>("softmax_0_cast_fp16")]; | |
| tensor<bool, []> attn_output_1_transpose_x_0 = const()[name = tensor<string, []>("attn_output_1_transpose_x_0"), val = tensor<bool, []>(false)]; | |
| tensor<bool, []> attn_output_1_transpose_y_0 = const()[name = tensor<string, []>("attn_output_1_transpose_y_0"), val = tensor<bool, []>(false)]; | |
| tensor<fp16, [1, 12, 50, 64]> attn_output_1_cast_fp16 = matmul(transpose_x = attn_output_1_transpose_x_0, transpose_y = attn_output_1_transpose_y_0, x = softmax_0_cast_fp16, y = v_5_cast_fp16)[name = tensor<string, []>("attn_output_1_cast_fp16")]; | |
| tensor<int32, [4]> var_149 = const()[name = tensor<string, []>("op_149"), val = tensor<int32, [4]>([2, 0, 1, 3])]; | |
| tensor<int32, [2]> var_154 = const()[name = tensor<string, []>("op_154"), val = tensor<int32, [2]>([50, 768])]; | |
| tensor<fp16, [50, 1, 12, 64]> var_150_cast_fp16 = transpose(perm = var_149, x = attn_output_1_cast_fp16)[name = tensor<string, []>("transpose_79")]; | |
| tensor<fp16, [50, 768]> attn_output_3_cast_fp16 = reshape(shape = var_154, x = var_150_cast_fp16)[name = tensor<string, []>("attn_output_3_cast_fp16")]; | |
| tensor<fp16, [768, 768]> model_visual_transformer_resblocks_0_attn_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("model_visual_transformer_resblocks_0_attn_out_proj_weight_to_fp16"), val = tensor<fp16, [768, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8347264)))]; | |
| tensor<fp16, [768]> model_visual_transformer_resblocks_0_attn_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_visual_transformer_resblocks_0_attn_out_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(9526976)))]; | |
| tensor<fp16, [50, 768]> linear_1_cast_fp16 = linear(bias = model_visual_transformer_resblocks_0_attn_out_proj_bias_to_fp16, weight = model_visual_transformer_resblocks_0_attn_out_proj_weight_to_fp16, x = attn_output_3_cast_fp16)[name = tensor<string, []>("linear_1_cast_fp16")]; | |
| tensor<int32, [3]> var_158 = const()[name = tensor<string, []>("op_158"), val = tensor<int32, [3]>([50, 1, 768])]; | |
| tensor<fp16, [50, 1, 768]> attn_output_7_cast_fp16 = reshape(shape = var_158, x = linear_1_cast_fp16)[name = tensor<string, []>("attn_output_7_cast_fp16")]; | |
| tensor<int32, [3]> var_160_perm_0 = const()[name = tensor<string, []>("op_160_perm_0"), val = tensor<int32, [3]>([1, 0, 2])]; | |
| tensor<fp16, [1, 50, 768]> var_160_cast_fp16 = transpose(perm = var_160_perm_0, x = attn_output_7_cast_fp16)[name = tensor<string, []>("transpose_78")]; | |
| tensor<fp16, [1, 50, 768]> input_5_cast_fp16 = add(x = x_9_cast_fp16, y = var_160_cast_fp16)[name = tensor<string, []>("input_5_cast_fp16")]; | |
| tensor<int32, [1]> x_13_axes_0 = const()[name = tensor<string, []>("x_13_axes_0"), val = tensor<int32, [1]>([-1])]; | |
| tensor<fp16, [768]> model_visual_transformer_resblocks_0_ln_2_weight_to_fp16 = const()[name = tensor<string, []>("model_visual_transformer_resblocks_0_ln_2_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(9528576)))]; | |
| tensor<fp16, [768]> model_visual_transformer_resblocks_0_ln_2_bias_to_fp16 = const()[name = tensor<string, []>("model_visual_transformer_resblocks_0_ln_2_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(9530176)))]; | |
| tensor<fp16, [1, 50, 768]> x_13_cast_fp16 = layer_norm(axes = x_13_axes_0, beta = model_visual_transformer_resblocks_0_ln_2_bias_to_fp16, epsilon = var_13_to_fp16, gamma = model_visual_transformer_resblocks_0_ln_2_weight_to_fp16, x = input_5_cast_fp16)[name = tensor<string, []>("x_13_cast_fp16")]; | |
| tensor<fp16, [3072, 768]> model_visual_transformer_resblocks_0_mlp_c_fc_weight_to_fp16 = const()[name = tensor<string, []>("model_visual_transformer_resblocks_0_mlp_c_fc_weight_to_fp16"), val = tensor<fp16, [3072, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(9531776)))]; | |
| tensor<fp16, [3072]> model_visual_transformer_resblocks_0_mlp_c_fc_bias_to_fp16 = const()[name = tensor<string, []>("model_visual_transformer_resblocks_0_mlp_c_fc_bias_to_fp16"), val = tensor<fp16, [3072]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(14250432)))]; | |
| tensor<fp16, [1, 50, 3072]> linear_2_cast_fp16 = linear(bias = model_visual_transformer_resblocks_0_mlp_c_fc_bias_to_fp16, weight = model_visual_transformer_resblocks_0_mlp_c_fc_weight_to_fp16, x = x_13_cast_fp16)[name = tensor<string, []>("linear_2_cast_fp16")]; | |
| tensor<string, []> input_11_mode_0 = const()[name = tensor<string, []>("input_11_mode_0"), val = tensor<string, []>("EXACT")]; | |
| tensor<fp16, [1, 50, 3072]> input_11_cast_fp16 = gelu(mode = input_11_mode_0, x = linear_2_cast_fp16)[name = tensor<string, []>("input_11_cast_fp16")]; | |
| tensor<fp16, [768, 3072]> model_visual_transformer_resblocks_0_mlp_c_proj_weight_to_fp16 = const()[name = tensor<string, []>("model_visual_transformer_resblocks_0_mlp_c_proj_weight_to_fp16"), val = tensor<fp16, [768, 3072]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(14256640)))]; | |
| tensor<fp16, [768]> model_visual_transformer_resblocks_0_mlp_c_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_visual_transformer_resblocks_0_mlp_c_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(18975296)))]; | |
| tensor<fp16, [1, 50, 768]> linear_3_cast_fp16 = linear(bias = model_visual_transformer_resblocks_0_mlp_c_proj_bias_to_fp16, weight = model_visual_transformer_resblocks_0_mlp_c_proj_weight_to_fp16, x = input_11_cast_fp16)[name = tensor<string, []>("linear_3_cast_fp16")]; | |
| tensor<fp16, [1, 50, 768]> input_13_cast_fp16 = add(x = input_5_cast_fp16, y = linear_3_cast_fp16)[name = tensor<string, []>("input_13_cast_fp16")]; | |
| tensor<int32, [1]> x_15_axes_0 = const()[name = tensor<string, []>("x_15_axes_0"), val = tensor<int32, [1]>([-1])]; | |
| tensor<fp16, [768]> model_visual_transformer_resblocks_1_ln_1_weight_to_fp16 = const()[name = tensor<string, []>("model_visual_transformer_resblocks_1_ln_1_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(18976896)))]; | |
| tensor<fp16, [768]> model_visual_transformer_resblocks_1_ln_1_bias_to_fp16 = const()[name = tensor<string, []>("model_visual_transformer_resblocks_1_ln_1_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(18978496)))]; | |
| tensor<fp16, [1, 50, 768]> x_15_cast_fp16 = layer_norm(axes = x_15_axes_0, beta = model_visual_transformer_resblocks_1_ln_1_bias_to_fp16, epsilon = var_13_to_fp16, gamma = model_visual_transformer_resblocks_1_ln_1_weight_to_fp16, x = input_13_cast_fp16)[name = tensor<string, []>("x_15_cast_fp16")]; | |
| tensor<int32, [3]> query_7_perm_0 = const()[name = tensor<string, []>("query_7_perm_0"), val = tensor<int32, [3]>([1, 0, 2])]; | |
| tensor<fp16, [2304, 768]> model_visual_transformer_resblocks_1_attn_in_proj_weight_to_fp16 = const()[name = tensor<string, []>("model_visual_transformer_resblocks_1_attn_in_proj_weight_to_fp16"), val = tensor<fp16, [2304, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(18980096)))]; | |
| tensor<fp16, [2304]> model_visual_transformer_resblocks_1_attn_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_visual_transformer_resblocks_1_attn_in_proj_bias_to_fp16"), val = tensor<fp16, [2304]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(22519104)))]; | |
| tensor<fp16, [50, 1, 768]> query_7_cast_fp16 = transpose(perm = query_7_perm_0, x = x_15_cast_fp16)[name = tensor<string, []>("transpose_77")]; | |
| tensor<fp16, [50, 1, 2304]> linear_4_cast_fp16 = linear(bias = model_visual_transformer_resblocks_1_attn_in_proj_bias_to_fp16, weight = model_visual_transformer_resblocks_1_attn_in_proj_weight_to_fp16, x = query_7_cast_fp16)[name = tensor<string, []>("linear_4_cast_fp16")]; | |
| tensor<int32, [4]> concat_2 = const()[name = tensor<string, []>("concat_2"), val = tensor<int32, [4]>([50, 1, 3, 768])]; | |
| tensor<fp16, [50, 1, 3, 768]> var_209_cast_fp16 = reshape(shape = concat_2, x = linear_4_cast_fp16)[name = tensor<string, []>("op_209_cast_fp16")]; | |
| tensor<int32, [1]> var_210_axes_0 = const()[name = tensor<string, []>("op_210_axes_0"), val = tensor<int32, [1]>([0])]; | |
| tensor<fp16, [1, 50, 1, 3, 768]> var_210_cast_fp16 = expand_dims(axes = var_210_axes_0, x = var_209_cast_fp16)[name = tensor<string, []>("op_210_cast_fp16")]; | |
| tensor<int32, [5]> var_211_perm_0 = const()[name = tensor<string, []>("op_211_perm_0"), val = tensor<int32, [5]>([-2, 1, 2, 0, 4])]; | |
| tensor<int32, [1]> var_212_axes_0 = const()[name = tensor<string, []>("op_212_axes_0"), val = tensor<int32, [1]>([-2])]; | |
| tensor<fp16, [3, 50, 1, 1, 768]> var_211_cast_fp16 = transpose(perm = var_211_perm_0, x = var_210_cast_fp16)[name = tensor<string, []>("transpose_76")]; | |
| tensor<fp16, [3, 50, 1, 768]> var_212_cast_fp16 = squeeze(axes = var_212_axes_0, x = var_211_cast_fp16)[name = tensor<string, []>("op_212_cast_fp16")]; | |
| tensor<int32, [4]> q_7_begin_0 = const()[name = tensor<string, []>("q_7_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; | |
| tensor<int32, [4]> q_7_end_0 = const()[name = tensor<string, []>("q_7_end_0"), val = tensor<int32, [4]>([1, 50, 1, 768])]; | |
| tensor<bool, [4]> q_7_end_mask_0 = const()[name = tensor<string, []>("q_7_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])]; | |
| tensor<bool, [4]> q_7_squeeze_mask_0 = const()[name = tensor<string, []>("q_7_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])]; | |
| tensor<fp16, [50, 1, 768]> q_7_cast_fp16 = slice_by_index(begin = q_7_begin_0, end = q_7_end_0, end_mask = q_7_end_mask_0, squeeze_mask = q_7_squeeze_mask_0, x = var_212_cast_fp16)[name = tensor<string, []>("q_7_cast_fp16")]; | |
| tensor<int32, [4]> k_7_begin_0 = const()[name = tensor<string, []>("k_7_begin_0"), val = tensor<int32, [4]>([1, 0, 0, 0])]; | |
| tensor<int32, [4]> k_7_end_0 = const()[name = tensor<string, []>("k_7_end_0"), val = tensor<int32, [4]>([2, 50, 1, 768])]; | |
| tensor<bool, [4]> k_7_end_mask_0 = const()[name = tensor<string, []>("k_7_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])]; | |
| tensor<bool, [4]> k_7_squeeze_mask_0 = const()[name = tensor<string, []>("k_7_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])]; | |
| tensor<fp16, [50, 1, 768]> k_7_cast_fp16 = slice_by_index(begin = k_7_begin_0, end = k_7_end_0, end_mask = k_7_end_mask_0, squeeze_mask = k_7_squeeze_mask_0, x = var_212_cast_fp16)[name = tensor<string, []>("k_7_cast_fp16")]; | |
| tensor<int32, [4]> v_7_begin_0 = const()[name = tensor<string, []>("v_7_begin_0"), val = tensor<int32, [4]>([2, 0, 0, 0])]; | |
| tensor<int32, [4]> v_7_end_0 = const()[name = tensor<string, []>("v_7_end_0"), val = tensor<int32, [4]>([3, 50, 1, 768])]; | |
| tensor<bool, [4]> v_7_end_mask_0 = const()[name = tensor<string, []>("v_7_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])]; | |
| tensor<bool, [4]> v_7_squeeze_mask_0 = const()[name = tensor<string, []>("v_7_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])]; | |
| tensor<fp16, [50, 1, 768]> v_7_cast_fp16 = slice_by_index(begin = v_7_begin_0, end = v_7_end_0, end_mask = v_7_end_mask_0, squeeze_mask = v_7_squeeze_mask_0, x = var_212_cast_fp16)[name = tensor<string, []>("v_7_cast_fp16")]; | |
| tensor<int32, [3]> var_220 = const()[name = tensor<string, []>("op_220"), val = tensor<int32, [3]>([50, 12, 64])]; | |
| tensor<fp16, [50, 12, 64]> var_221_cast_fp16 = reshape(shape = var_220, x = q_7_cast_fp16)[name = tensor<string, []>("op_221_cast_fp16")]; | |
| tensor<int32, [3]> q_9_perm_0 = const()[name = tensor<string, []>("q_9_perm_0"), val = tensor<int32, [3]>([1, 0, 2])]; | |
| tensor<int32, [3]> var_227 = const()[name = tensor<string, []>("op_227"), val = tensor<int32, [3]>([50, 12, 64])]; | |
| tensor<fp16, [50, 12, 64]> var_228_cast_fp16 = reshape(shape = var_227, x = k_7_cast_fp16)[name = tensor<string, []>("op_228_cast_fp16")]; | |
| tensor<int32, [3]> k_9_perm_0 = const()[name = tensor<string, []>("k_9_perm_0"), val = tensor<int32, [3]>([1, 0, 2])]; | |
| tensor<int32, [3]> var_234 = const()[name = tensor<string, []>("op_234"), val = tensor<int32, [3]>([50, 12, 64])]; | |
| tensor<fp16, [50, 12, 64]> var_235_cast_fp16 = reshape(shape = var_234, x = v_7_cast_fp16)[name = tensor<string, []>("op_235_cast_fp16")]; | |
| tensor<int32, [3]> v_9_perm_0 = const()[name = tensor<string, []>("v_9_perm_0"), val = tensor<int32, [3]>([1, 0, 2])]; | |
| tensor<int32, [4]> var_238 = const()[name = tensor<string, []>("op_238"), val = tensor<int32, [4]>([1, 12, 50, 64])]; | |
| tensor<fp16, [12, 50, 64]> q_9_cast_fp16 = transpose(perm = q_9_perm_0, x = var_221_cast_fp16)[name = tensor<string, []>("transpose_75")]; | |
| tensor<fp16, [1, 12, 50, 64]> q_11_cast_fp16 = reshape(shape = var_238, x = q_9_cast_fp16)[name = tensor<string, []>("q_11_cast_fp16")]; | |
| tensor<int32, [4]> var_240 = const()[name = tensor<string, []>("op_240"), val = tensor<int32, [4]>([1, 12, 50, 64])]; | |
| tensor<fp16, [12, 50, 64]> k_9_cast_fp16 = transpose(perm = k_9_perm_0, x = var_228_cast_fp16)[name = tensor<string, []>("transpose_74")]; | |
| tensor<fp16, [1, 12, 50, 64]> k_11_cast_fp16 = reshape(shape = var_240, x = k_9_cast_fp16)[name = tensor<string, []>("k_11_cast_fp16")]; | |
| tensor<int32, [4]> var_242 = const()[name = tensor<string, []>("op_242"), val = tensor<int32, [4]>([1, 12, 50, 64])]; | |
| tensor<fp16, [12, 50, 64]> v_9_cast_fp16 = transpose(perm = v_9_perm_0, x = var_235_cast_fp16)[name = tensor<string, []>("transpose_73")]; | |
| tensor<fp16, [1, 12, 50, 64]> v_11_cast_fp16 = reshape(shape = var_242, x = v_9_cast_fp16)[name = tensor<string, []>("v_11_cast_fp16")]; | |
| tensor<fp16, []> mul_3_y_0_to_fp16 = const()[name = tensor<string, []>("mul_3_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)]; | |
| tensor<fp16, [1, 12, 50, 64]> mul_3_cast_fp16 = mul(x = q_11_cast_fp16, y = mul_3_y_0_to_fp16)[name = tensor<string, []>("mul_3_cast_fp16")]; | |
| tensor<bool, []> matmul_1_transpose_y_0 = const()[name = tensor<string, []>("matmul_1_transpose_y_0"), val = tensor<bool, []>(true)]; | |
| tensor<bool, []> matmul_1_transpose_x_0 = const()[name = tensor<string, []>("matmul_1_transpose_x_0"), val = tensor<bool, []>(false)]; | |
| tensor<fp16, [1, 12, 50, 50]> matmul_1_cast_fp16 = matmul(transpose_x = matmul_1_transpose_x_0, transpose_y = matmul_1_transpose_y_0, x = mul_3_cast_fp16, y = k_11_cast_fp16)[name = tensor<string, []>("matmul_1_cast_fp16")]; | |
| tensor<int32, []> softmax_1_axis_0 = const()[name = tensor<string, []>("softmax_1_axis_0"), val = tensor<int32, []>(-1)]; | |
| tensor<fp16, [1, 12, 50, 50]> softmax_1_cast_fp16 = softmax(axis = softmax_1_axis_0, x = matmul_1_cast_fp16)[name = tensor<string, []>("softmax_1_cast_fp16")]; | |
| tensor<bool, []> attn_output_9_transpose_x_0 = const()[name = tensor<string, []>("attn_output_9_transpose_x_0"), val = tensor<bool, []>(false)]; | |
| tensor<bool, []> attn_output_9_transpose_y_0 = const()[name = tensor<string, []>("attn_output_9_transpose_y_0"), val = tensor<bool, []>(false)]; | |
| tensor<fp16, [1, 12, 50, 64]> attn_output_9_cast_fp16 = matmul(transpose_x = attn_output_9_transpose_x_0, transpose_y = attn_output_9_transpose_y_0, x = softmax_1_cast_fp16, y = v_11_cast_fp16)[name = tensor<string, []>("attn_output_9_cast_fp16")]; | |
| tensor<int32, [4]> var_245 = const()[name = tensor<string, []>("op_245"), val = tensor<int32, [4]>([2, 0, 1, 3])]; | |
| tensor<int32, [2]> var_250 = const()[name = tensor<string, []>("op_250"), val = tensor<int32, [2]>([50, 768])]; | |
| tensor<fp16, [50, 1, 12, 64]> var_246_cast_fp16 = transpose(perm = var_245, x = attn_output_9_cast_fp16)[name = tensor<string, []>("transpose_72")]; | |
| tensor<fp16, [50, 768]> attn_output_11_cast_fp16 = reshape(shape = var_250, x = var_246_cast_fp16)[name = tensor<string, []>("attn_output_11_cast_fp16")]; | |
| tensor<fp16, [768, 768]> model_visual_transformer_resblocks_1_attn_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("model_visual_transformer_resblocks_1_attn_out_proj_weight_to_fp16"), val = tensor<fp16, [768, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(22523776)))]; | |
| tensor<fp16, [768]> model_visual_transformer_resblocks_1_attn_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_visual_transformer_resblocks_1_attn_out_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(23703488)))]; | |
| tensor<fp16, [50, 768]> linear_5_cast_fp16 = linear(bias = model_visual_transformer_resblocks_1_attn_out_proj_bias_to_fp16, weight = model_visual_transformer_resblocks_1_attn_out_proj_weight_to_fp16, x = attn_output_11_cast_fp16)[name = tensor<string, []>("linear_5_cast_fp16")]; | |
| tensor<int32, [3]> var_254 = const()[name = tensor<string, []>("op_254"), val = tensor<int32, [3]>([50, 1, 768])]; | |
| tensor<fp16, [50, 1, 768]> attn_output_15_cast_fp16 = reshape(shape = var_254, x = linear_5_cast_fp16)[name = tensor<string, []>("attn_output_15_cast_fp16")]; | |
| tensor<int32, [3]> var_256_perm_0 = const()[name = tensor<string, []>("op_256_perm_0"), val = tensor<int32, [3]>([1, 0, 2])]; | |
| tensor<fp16, [1, 50, 768]> var_256_cast_fp16 = transpose(perm = var_256_perm_0, x = attn_output_15_cast_fp16)[name = tensor<string, []>("transpose_71")]; | |
| tensor<fp16, [1, 50, 768]> input_15_cast_fp16 = add(x = input_13_cast_fp16, y = var_256_cast_fp16)[name = tensor<string, []>("input_15_cast_fp16")]; | |
| tensor<int32, [1]> x_17_axes_0 = const()[name = tensor<string, []>("x_17_axes_0"), val = tensor<int32, [1]>([-1])]; | |
| tensor<fp16, [768]> model_visual_transformer_resblocks_1_ln_2_weight_to_fp16 = const()[name = tensor<string, []>("model_visual_transformer_resblocks_1_ln_2_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(23705088)))]; | |
| tensor<fp16, [768]> model_visual_transformer_resblocks_1_ln_2_bias_to_fp16 = const()[name = tensor<string, []>("model_visual_transformer_resblocks_1_ln_2_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(23706688)))]; | |
| tensor<fp16, [1, 50, 768]> x_17_cast_fp16 = layer_norm(axes = x_17_axes_0, beta = model_visual_transformer_resblocks_1_ln_2_bias_to_fp16, epsilon = var_13_to_fp16, gamma = model_visual_transformer_resblocks_1_ln_2_weight_to_fp16, x = input_15_cast_fp16)[name = tensor<string, []>("x_17_cast_fp16")]; | |
| tensor<fp16, [3072, 768]> model_visual_transformer_resblocks_1_mlp_c_fc_weight_to_fp16 = const()[name = tensor<string, []>("model_visual_transformer_resblocks_1_mlp_c_fc_weight_to_fp16"), val = tensor<fp16, [3072, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(23708288)))]; | |
| tensor<fp16, [3072]> model_visual_transformer_resblocks_1_mlp_c_fc_bias_to_fp16 = const()[name = tensor<string, []>("model_visual_transformer_resblocks_1_mlp_c_fc_bias_to_fp16"), val = tensor<fp16, [3072]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(28426944)))]; | |
| tensor<fp16, [1, 50, 3072]> linear_6_cast_fp16 = linear(bias = model_visual_transformer_resblocks_1_mlp_c_fc_bias_to_fp16, weight = model_visual_transformer_resblocks_1_mlp_c_fc_weight_to_fp16, x = x_17_cast_fp16)[name = tensor<string, []>("linear_6_cast_fp16")]; | |
| tensor<string, []> input_21_mode_0 = const()[name = tensor<string, []>("input_21_mode_0"), val = tensor<string, []>("EXACT")]; | |
| tensor<fp16, [1, 50, 3072]> input_21_cast_fp16 = gelu(mode = input_21_mode_0, x = linear_6_cast_fp16)[name = tensor<string, []>("input_21_cast_fp16")]; | |
| tensor<fp16, [768, 3072]> model_visual_transformer_resblocks_1_mlp_c_proj_weight_to_fp16 = const()[name = tensor<string, []>("model_visual_transformer_resblocks_1_mlp_c_proj_weight_to_fp16"), val = tensor<fp16, [768, 3072]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(28433152)))]; | |
| tensor<fp16, [768]> model_visual_transformer_resblocks_1_mlp_c_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_visual_transformer_resblocks_1_mlp_c_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(33151808)))]; | |
| tensor<fp16, [1, 50, 768]> linear_7_cast_fp16 = linear(bias = model_visual_transformer_resblocks_1_mlp_c_proj_bias_to_fp16, weight = model_visual_transformer_resblocks_1_mlp_c_proj_weight_to_fp16, x = input_21_cast_fp16)[name = tensor<string, []>("linear_7_cast_fp16")]; | |
| tensor<fp16, [1, 50, 768]> input_23_cast_fp16 = add(x = input_15_cast_fp16, y = linear_7_cast_fp16)[name = tensor<string, []>("input_23_cast_fp16")]; | |
| tensor<int32, [1]> x_19_axes_0 = const()[name = tensor<string, []>("x_19_axes_0"), val = tensor<int32, [1]>([-1])]; | |
| tensor<fp16, [768]> model_visual_transformer_resblocks_2_ln_1_weight_to_fp16 = const()[name = tensor<string, []>("model_visual_transformer_resblocks_2_ln_1_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(33153408)))]; | |
| tensor<fp16, [768]> model_visual_transformer_resblocks_2_ln_1_bias_to_fp16 = const()[name = tensor<string, []>("model_visual_transformer_resblocks_2_ln_1_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(33155008)))]; | |
| tensor<fp16, [1, 50, 768]> x_19_cast_fp16 = layer_norm(axes = x_19_axes_0, beta = model_visual_transformer_resblocks_2_ln_1_bias_to_fp16, epsilon = var_13_to_fp16, gamma = model_visual_transformer_resblocks_2_ln_1_weight_to_fp16, x = input_23_cast_fp16)[name = tensor<string, []>("x_19_cast_fp16")]; | |
| tensor<int32, [3]> query_11_perm_0 = const()[name = tensor<string, []>("query_11_perm_0"), val = tensor<int32, [3]>([1, 0, 2])]; | |
| tensor<fp16, [2304, 768]> model_visual_transformer_resblocks_2_attn_in_proj_weight_to_fp16 = const()[name = tensor<string, []>("model_visual_transformer_resblocks_2_attn_in_proj_weight_to_fp16"), val = tensor<fp16, [2304, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(33156608)))]; | |
| tensor<fp16, [2304]> model_visual_transformer_resblocks_2_attn_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_visual_transformer_resblocks_2_attn_in_proj_bias_to_fp16"), val = tensor<fp16, [2304]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(36695616)))]; | |
| tensor<fp16, [50, 1, 768]> query_11_cast_fp16 = transpose(perm = query_11_perm_0, x = x_19_cast_fp16)[name = tensor<string, []>("transpose_70")]; | |
| tensor<fp16, [50, 1, 2304]> linear_8_cast_fp16 = linear(bias = model_visual_transformer_resblocks_2_attn_in_proj_bias_to_fp16, weight = model_visual_transformer_resblocks_2_attn_in_proj_weight_to_fp16, x = query_11_cast_fp16)[name = tensor<string, []>("linear_8_cast_fp16")]; | |
| tensor<int32, [4]> concat_3 = const()[name = tensor<string, []>("concat_3"), val = tensor<int32, [4]>([50, 1, 3, 768])]; | |
| tensor<fp16, [50, 1, 3, 768]> var_305_cast_fp16 = reshape(shape = concat_3, x = linear_8_cast_fp16)[name = tensor<string, []>("op_305_cast_fp16")]; | |
| tensor<int32, [1]> var_306_axes_0 = const()[name = tensor<string, []>("op_306_axes_0"), val = tensor<int32, [1]>([0])]; | |
| tensor<fp16, [1, 50, 1, 3, 768]> var_306_cast_fp16 = expand_dims(axes = var_306_axes_0, x = var_305_cast_fp16)[name = tensor<string, []>("op_306_cast_fp16")]; | |
| tensor<int32, [5]> var_307_perm_0 = const()[name = tensor<string, []>("op_307_perm_0"), val = tensor<int32, [5]>([-2, 1, 2, 0, 4])]; | |
| tensor<int32, [1]> var_308_axes_0 = const()[name = tensor<string, []>("op_308_axes_0"), val = tensor<int32, [1]>([-2])]; | |
| tensor<fp16, [3, 50, 1, 1, 768]> var_307_cast_fp16 = transpose(perm = var_307_perm_0, x = var_306_cast_fp16)[name = tensor<string, []>("transpose_69")]; | |
| tensor<fp16, [3, 50, 1, 768]> var_308_cast_fp16 = squeeze(axes = var_308_axes_0, x = var_307_cast_fp16)[name = tensor<string, []>("op_308_cast_fp16")]; | |
| tensor<int32, [4]> q_13_begin_0 = const()[name = tensor<string, []>("q_13_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; | |
| tensor<int32, [4]> q_13_end_0 = const()[name = tensor<string, []>("q_13_end_0"), val = tensor<int32, [4]>([1, 50, 1, 768])]; | |
| tensor<bool, [4]> q_13_end_mask_0 = const()[name = tensor<string, []>("q_13_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])]; | |
| tensor<bool, [4]> q_13_squeeze_mask_0 = const()[name = tensor<string, []>("q_13_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])]; | |
| tensor<fp16, [50, 1, 768]> q_13_cast_fp16 = slice_by_index(begin = q_13_begin_0, end = q_13_end_0, end_mask = q_13_end_mask_0, squeeze_mask = q_13_squeeze_mask_0, x = var_308_cast_fp16)[name = tensor<string, []>("q_13_cast_fp16")]; | |
| tensor<int32, [4]> k_13_begin_0 = const()[name = tensor<string, []>("k_13_begin_0"), val = tensor<int32, [4]>([1, 0, 0, 0])]; | |
| tensor<int32, [4]> k_13_end_0 = const()[name = tensor<string, []>("k_13_end_0"), val = tensor<int32, [4]>([2, 50, 1, 768])]; | |
| tensor<bool, [4]> k_13_end_mask_0 = const()[name = tensor<string, []>("k_13_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])]; | |
| tensor<bool, [4]> k_13_squeeze_mask_0 = const()[name = tensor<string, []>("k_13_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])]; | |
| tensor<fp16, [50, 1, 768]> k_13_cast_fp16 = slice_by_index(begin = k_13_begin_0, end = k_13_end_0, end_mask = k_13_end_mask_0, squeeze_mask = k_13_squeeze_mask_0, x = var_308_cast_fp16)[name = tensor<string, []>("k_13_cast_fp16")]; | |
| tensor<int32, [4]> v_13_begin_0 = const()[name = tensor<string, []>("v_13_begin_0"), val = tensor<int32, [4]>([2, 0, 0, 0])]; | |
| tensor<int32, [4]> v_13_end_0 = const()[name = tensor<string, []>("v_13_end_0"), val = tensor<int32, [4]>([3, 50, 1, 768])]; | |
| tensor<bool, [4]> v_13_end_mask_0 = const()[name = tensor<string, []>("v_13_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])]; | |
| tensor<bool, [4]> v_13_squeeze_mask_0 = const()[name = tensor<string, []>("v_13_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])]; | |
| tensor<fp16, [50, 1, 768]> v_13_cast_fp16 = slice_by_index(begin = v_13_begin_0, end = v_13_end_0, end_mask = v_13_end_mask_0, squeeze_mask = v_13_squeeze_mask_0, x = var_308_cast_fp16)[name = tensor<string, []>("v_13_cast_fp16")]; | |
| tensor<int32, [3]> var_316 = const()[name = tensor<string, []>("op_316"), val = tensor<int32, [3]>([50, 12, 64])]; | |
| tensor<fp16, [50, 12, 64]> var_317_cast_fp16 = reshape(shape = var_316, x = q_13_cast_fp16)[name = tensor<string, []>("op_317_cast_fp16")]; | |
| tensor<int32, [3]> q_15_perm_0 = const()[name = tensor<string, []>("q_15_perm_0"), val = tensor<int32, [3]>([1, 0, 2])]; | |
| tensor<int32, [3]> var_323 = const()[name = tensor<string, []>("op_323"), val = tensor<int32, [3]>([50, 12, 64])]; | |
| tensor<fp16, [50, 12, 64]> var_324_cast_fp16 = reshape(shape = var_323, x = k_13_cast_fp16)[name = tensor<string, []>("op_324_cast_fp16")]; | |
| tensor<int32, [3]> k_15_perm_0 = const()[name = tensor<string, []>("k_15_perm_0"), val = tensor<int32, [3]>([1, 0, 2])]; | |
| tensor<int32, [3]> var_330 = const()[name = tensor<string, []>("op_330"), val = tensor<int32, [3]>([50, 12, 64])]; | |
| tensor<fp16, [50, 12, 64]> var_331_cast_fp16 = reshape(shape = var_330, x = v_13_cast_fp16)[name = tensor<string, []>("op_331_cast_fp16")]; | |
| tensor<int32, [3]> v_15_perm_0 = const()[name = tensor<string, []>("v_15_perm_0"), val = tensor<int32, [3]>([1, 0, 2])]; | |
| tensor<int32, [4]> var_334 = const()[name = tensor<string, []>("op_334"), val = tensor<int32, [4]>([1, 12, 50, 64])]; | |
| tensor<fp16, [12, 50, 64]> q_15_cast_fp16 = transpose(perm = q_15_perm_0, x = var_317_cast_fp16)[name = tensor<string, []>("transpose_68")]; | |
| tensor<fp16, [1, 12, 50, 64]> q_17_cast_fp16 = reshape(shape = var_334, x = q_15_cast_fp16)[name = tensor<string, []>("q_17_cast_fp16")]; | |
| tensor<int32, [4]> var_336 = const()[name = tensor<string, []>("op_336"), val = tensor<int32, [4]>([1, 12, 50, 64])]; | |
| tensor<fp16, [12, 50, 64]> k_15_cast_fp16 = transpose(perm = k_15_perm_0, x = var_324_cast_fp16)[name = tensor<string, []>("transpose_67")]; | |
| tensor<fp16, [1, 12, 50, 64]> k_17_cast_fp16 = reshape(shape = var_336, x = k_15_cast_fp16)[name = tensor<string, []>("k_17_cast_fp16")]; | |
| tensor<int32, [4]> var_338 = const()[name = tensor<string, []>("op_338"), val = tensor<int32, [4]>([1, 12, 50, 64])]; | |
| tensor<fp16, [12, 50, 64]> v_15_cast_fp16 = transpose(perm = v_15_perm_0, x = var_331_cast_fp16)[name = tensor<string, []>("transpose_66")]; | |
| tensor<fp16, [1, 12, 50, 64]> v_17_cast_fp16 = reshape(shape = var_338, x = v_15_cast_fp16)[name = tensor<string, []>("v_17_cast_fp16")]; | |
| tensor<fp16, []> mul_5_y_0_to_fp16 = const()[name = tensor<string, []>("mul_5_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)]; | |
| tensor<fp16, [1, 12, 50, 64]> mul_5_cast_fp16 = mul(x = q_17_cast_fp16, y = mul_5_y_0_to_fp16)[name = tensor<string, []>("mul_5_cast_fp16")]; | |
| tensor<bool, []> matmul_2_transpose_y_0 = const()[name = tensor<string, []>("matmul_2_transpose_y_0"), val = tensor<bool, []>(true)]; | |
| tensor<bool, []> matmul_2_transpose_x_0 = const()[name = tensor<string, []>("matmul_2_transpose_x_0"), val = tensor<bool, []>(false)]; | |
| tensor<fp16, [1, 12, 50, 50]> matmul_2_cast_fp16 = matmul(transpose_x = matmul_2_transpose_x_0, transpose_y = matmul_2_transpose_y_0, x = mul_5_cast_fp16, y = k_17_cast_fp16)[name = tensor<string, []>("matmul_2_cast_fp16")]; | |
| tensor<int32, []> softmax_2_axis_0 = const()[name = tensor<string, []>("softmax_2_axis_0"), val = tensor<int32, []>(-1)]; | |
| tensor<fp16, [1, 12, 50, 50]> softmax_2_cast_fp16 = softmax(axis = softmax_2_axis_0, x = matmul_2_cast_fp16)[name = tensor<string, []>("softmax_2_cast_fp16")]; | |
| tensor<bool, []> attn_output_17_transpose_x_0 = const()[name = tensor<string, []>("attn_output_17_transpose_x_0"), val = tensor<bool, []>(false)]; | |
| tensor<bool, []> attn_output_17_transpose_y_0 = const()[name = tensor<string, []>("attn_output_17_transpose_y_0"), val = tensor<bool, []>(false)]; | |
| tensor<fp16, [1, 12, 50, 64]> attn_output_17_cast_fp16 = matmul(transpose_x = attn_output_17_transpose_x_0, transpose_y = attn_output_17_transpose_y_0, x = softmax_2_cast_fp16, y = v_17_cast_fp16)[name = tensor<string, []>("attn_output_17_cast_fp16")]; | |
| tensor<int32, [4]> var_341 = const()[name = tensor<string, []>("op_341"), val = tensor<int32, [4]>([2, 0, 1, 3])]; | |
| tensor<int32, [2]> var_346 = const()[name = tensor<string, []>("op_346"), val = tensor<int32, [2]>([50, 768])]; | |
| tensor<fp16, [50, 1, 12, 64]> var_342_cast_fp16 = transpose(perm = var_341, x = attn_output_17_cast_fp16)[name = tensor<string, []>("transpose_65")]; | |
| tensor<fp16, [50, 768]> attn_output_19_cast_fp16 = reshape(shape = var_346, x = var_342_cast_fp16)[name = tensor<string, []>("attn_output_19_cast_fp16")]; | |
| tensor<fp16, [768, 768]> model_visual_transformer_resblocks_2_attn_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("model_visual_transformer_resblocks_2_attn_out_proj_weight_to_fp16"), val = tensor<fp16, [768, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(36700288)))]; | |
| tensor<fp16, [768]> model_visual_transformer_resblocks_2_attn_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_visual_transformer_resblocks_2_attn_out_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(37880000)))]; | |
| tensor<fp16, [50, 768]> linear_9_cast_fp16 = linear(bias = model_visual_transformer_resblocks_2_attn_out_proj_bias_to_fp16, weight = model_visual_transformer_resblocks_2_attn_out_proj_weight_to_fp16, x = attn_output_19_cast_fp16)[name = tensor<string, []>("linear_9_cast_fp16")]; | |
| tensor<int32, [3]> var_350 = const()[name = tensor<string, []>("op_350"), val = tensor<int32, [3]>([50, 1, 768])]; | |
| tensor<fp16, [50, 1, 768]> attn_output_23_cast_fp16 = reshape(shape = var_350, x = linear_9_cast_fp16)[name = tensor<string, []>("attn_output_23_cast_fp16")]; | |
| tensor<int32, [3]> var_352_perm_0 = const()[name = tensor<string, []>("op_352_perm_0"), val = tensor<int32, [3]>([1, 0, 2])]; | |
| tensor<fp16, [1, 50, 768]> var_352_cast_fp16 = transpose(perm = var_352_perm_0, x = attn_output_23_cast_fp16)[name = tensor<string, []>("transpose_64")]; | |
| tensor<fp16, [1, 50, 768]> input_25_cast_fp16 = add(x = input_23_cast_fp16, y = var_352_cast_fp16)[name = tensor<string, []>("input_25_cast_fp16")]; | |
| tensor<int32, [1]> x_21_axes_0 = const()[name = tensor<string, []>("x_21_axes_0"), val = tensor<int32, [1]>([-1])]; | |
| tensor<fp16, [768]> model_visual_transformer_resblocks_2_ln_2_weight_to_fp16 = const()[name = tensor<string, []>("model_visual_transformer_resblocks_2_ln_2_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(37881600)))]; | |
| tensor<fp16, [768]> model_visual_transformer_resblocks_2_ln_2_bias_to_fp16 = const()[name = tensor<string, []>("model_visual_transformer_resblocks_2_ln_2_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(37883200)))]; | |
| tensor<fp16, [1, 50, 768]> x_21_cast_fp16 = layer_norm(axes = x_21_axes_0, beta = model_visual_transformer_resblocks_2_ln_2_bias_to_fp16, epsilon = var_13_to_fp16, gamma = model_visual_transformer_resblocks_2_ln_2_weight_to_fp16, x = input_25_cast_fp16)[name = tensor<string, []>("x_21_cast_fp16")]; | |
| tensor<fp16, [3072, 768]> model_visual_transformer_resblocks_2_mlp_c_fc_weight_to_fp16 = const()[name = tensor<string, []>("model_visual_transformer_resblocks_2_mlp_c_fc_weight_to_fp16"), val = tensor<fp16, [3072, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(37884800)))]; | |
| tensor<fp16, [3072]> model_visual_transformer_resblocks_2_mlp_c_fc_bias_to_fp16 = const()[name = tensor<string, []>("model_visual_transformer_resblocks_2_mlp_c_fc_bias_to_fp16"), val = tensor<fp16, [3072]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(42603456)))]; | |
| tensor<fp16, [1, 50, 3072]> linear_10_cast_fp16 = linear(bias = model_visual_transformer_resblocks_2_mlp_c_fc_bias_to_fp16, weight = model_visual_transformer_resblocks_2_mlp_c_fc_weight_to_fp16, x = x_21_cast_fp16)[name = tensor<string, []>("linear_10_cast_fp16")]; | |
| tensor<string, []> input_31_mode_0 = const()[name = tensor<string, []>("input_31_mode_0"), val = tensor<string, []>("EXACT")]; | |
| tensor<fp16, [1, 50, 3072]> input_31_cast_fp16 = gelu(mode = input_31_mode_0, x = linear_10_cast_fp16)[name = tensor<string, []>("input_31_cast_fp16")]; | |
| tensor<fp16, [768, 3072]> model_visual_transformer_resblocks_2_mlp_c_proj_weight_to_fp16 = const()[name = tensor<string, []>("model_visual_transformer_resblocks_2_mlp_c_proj_weight_to_fp16"), val = tensor<fp16, [768, 3072]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(42609664)))]; | |
| tensor<fp16, [768]> model_visual_transformer_resblocks_2_mlp_c_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_visual_transformer_resblocks_2_mlp_c_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(47328320)))]; | |
| tensor<fp16, [1, 50, 768]> linear_11_cast_fp16 = linear(bias = model_visual_transformer_resblocks_2_mlp_c_proj_bias_to_fp16, weight = model_visual_transformer_resblocks_2_mlp_c_proj_weight_to_fp16, x = input_31_cast_fp16)[name = tensor<string, []>("linear_11_cast_fp16")]; | |
| tensor<fp16, [1, 50, 768]> input_33_cast_fp16 = add(x = input_25_cast_fp16, y = linear_11_cast_fp16)[name = tensor<string, []>("input_33_cast_fp16")]; | |
| tensor<int32, [1]> x_23_axes_0 = const()[name = tensor<string, []>("x_23_axes_0"), val = tensor<int32, [1]>([-1])]; | |
| tensor<fp16, [768]> model_visual_transformer_resblocks_3_ln_1_weight_to_fp16 = const()[name = tensor<string, []>("model_visual_transformer_resblocks_3_ln_1_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(47329920)))]; | |
| tensor<fp16, [768]> model_visual_transformer_resblocks_3_ln_1_bias_to_fp16 = const()[name = tensor<string, []>("model_visual_transformer_resblocks_3_ln_1_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(47331520)))]; | |
| tensor<fp16, [1, 50, 768]> x_23_cast_fp16 = layer_norm(axes = x_23_axes_0, beta = model_visual_transformer_resblocks_3_ln_1_bias_to_fp16, epsilon = var_13_to_fp16, gamma = model_visual_transformer_resblocks_3_ln_1_weight_to_fp16, x = input_33_cast_fp16)[name = tensor<string, []>("x_23_cast_fp16")]; | |
| tensor<int32, [3]> query_15_perm_0 = const()[name = tensor<string, []>("query_15_perm_0"), val = tensor<int32, [3]>([1, 0, 2])]; | |
| tensor<fp16, [2304, 768]> model_visual_transformer_resblocks_3_attn_in_proj_weight_to_fp16 = const()[name = tensor<string, []>("model_visual_transformer_resblocks_3_attn_in_proj_weight_to_fp16"), val = tensor<fp16, [2304, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(47333120)))]; | |
| tensor<fp16, [2304]> model_visual_transformer_resblocks_3_attn_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_visual_transformer_resblocks_3_attn_in_proj_bias_to_fp16"), val = tensor<fp16, [2304]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(50872128)))]; | |
| tensor<fp16, [50, 1, 768]> query_15_cast_fp16 = transpose(perm = query_15_perm_0, x = x_23_cast_fp16)[name = tensor<string, []>("transpose_63")]; | |
| tensor<fp16, [50, 1, 2304]> linear_12_cast_fp16 = linear(bias = model_visual_transformer_resblocks_3_attn_in_proj_bias_to_fp16, weight = model_visual_transformer_resblocks_3_attn_in_proj_weight_to_fp16, x = query_15_cast_fp16)[name = tensor<string, []>("linear_12_cast_fp16")]; | |
| tensor<int32, [4]> concat_4 = const()[name = tensor<string, []>("concat_4"), val = tensor<int32, [4]>([50, 1, 3, 768])]; | |
| tensor<fp16, [50, 1, 3, 768]> var_401_cast_fp16 = reshape(shape = concat_4, x = linear_12_cast_fp16)[name = tensor<string, []>("op_401_cast_fp16")]; | |
| tensor<int32, [1]> var_402_axes_0 = const()[name = tensor<string, []>("op_402_axes_0"), val = tensor<int32, [1]>([0])]; | |
| tensor<fp16, [1, 50, 1, 3, 768]> var_402_cast_fp16 = expand_dims(axes = var_402_axes_0, x = var_401_cast_fp16)[name = tensor<string, []>("op_402_cast_fp16")]; | |
| tensor<int32, [5]> var_403_perm_0 = const()[name = tensor<string, []>("op_403_perm_0"), val = tensor<int32, [5]>([-2, 1, 2, 0, 4])]; | |
| tensor<int32, [1]> var_404_axes_0 = const()[name = tensor<string, []>("op_404_axes_0"), val = tensor<int32, [1]>([-2])]; | |
| tensor<fp16, [3, 50, 1, 1, 768]> var_403_cast_fp16 = transpose(perm = var_403_perm_0, x = var_402_cast_fp16)[name = tensor<string, []>("transpose_62")]; | |
| tensor<fp16, [3, 50, 1, 768]> var_404_cast_fp16 = squeeze(axes = var_404_axes_0, x = var_403_cast_fp16)[name = tensor<string, []>("op_404_cast_fp16")]; | |
| tensor<int32, [4]> q_19_begin_0 = const()[name = tensor<string, []>("q_19_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; | |
| tensor<int32, [4]> q_19_end_0 = const()[name = tensor<string, []>("q_19_end_0"), val = tensor<int32, [4]>([1, 50, 1, 768])]; | |
| tensor<bool, [4]> q_19_end_mask_0 = const()[name = tensor<string, []>("q_19_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])]; | |
| tensor<bool, [4]> q_19_squeeze_mask_0 = const()[name = tensor<string, []>("q_19_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])]; | |
| tensor<fp16, [50, 1, 768]> q_19_cast_fp16 = slice_by_index(begin = q_19_begin_0, end = q_19_end_0, end_mask = q_19_end_mask_0, squeeze_mask = q_19_squeeze_mask_0, x = var_404_cast_fp16)[name = tensor<string, []>("q_19_cast_fp16")]; | |
| tensor<int32, [4]> k_19_begin_0 = const()[name = tensor<string, []>("k_19_begin_0"), val = tensor<int32, [4]>([1, 0, 0, 0])]; | |
| tensor<int32, [4]> k_19_end_0 = const()[name = tensor<string, []>("k_19_end_0"), val = tensor<int32, [4]>([2, 50, 1, 768])]; | |
| tensor<bool, [4]> k_19_end_mask_0 = const()[name = tensor<string, []>("k_19_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])]; | |
| tensor<bool, [4]> k_19_squeeze_mask_0 = const()[name = tensor<string, []>("k_19_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])]; | |
| tensor<fp16, [50, 1, 768]> k_19_cast_fp16 = slice_by_index(begin = k_19_begin_0, end = k_19_end_0, end_mask = k_19_end_mask_0, squeeze_mask = k_19_squeeze_mask_0, x = var_404_cast_fp16)[name = tensor<string, []>("k_19_cast_fp16")]; | |
| tensor<int32, [4]> v_19_begin_0 = const()[name = tensor<string, []>("v_19_begin_0"), val = tensor<int32, [4]>([2, 0, 0, 0])]; | |
| tensor<int32, [4]> v_19_end_0 = const()[name = tensor<string, []>("v_19_end_0"), val = tensor<int32, [4]>([3, 50, 1, 768])]; | |
| tensor<bool, [4]> v_19_end_mask_0 = const()[name = tensor<string, []>("v_19_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])]; | |
| tensor<bool, [4]> v_19_squeeze_mask_0 = const()[name = tensor<string, []>("v_19_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])]; | |
| tensor<fp16, [50, 1, 768]> v_19_cast_fp16 = slice_by_index(begin = v_19_begin_0, end = v_19_end_0, end_mask = v_19_end_mask_0, squeeze_mask = v_19_squeeze_mask_0, x = var_404_cast_fp16)[name = tensor<string, []>("v_19_cast_fp16")]; | |
| tensor<int32, [3]> var_412 = const()[name = tensor<string, []>("op_412"), val = tensor<int32, [3]>([50, 12, 64])]; | |
| tensor<fp16, [50, 12, 64]> var_413_cast_fp16 = reshape(shape = var_412, x = q_19_cast_fp16)[name = tensor<string, []>("op_413_cast_fp16")]; | |
| tensor<int32, [3]> q_21_perm_0 = const()[name = tensor<string, []>("q_21_perm_0"), val = tensor<int32, [3]>([1, 0, 2])]; | |
| tensor<int32, [3]> var_419 = const()[name = tensor<string, []>("op_419"), val = tensor<int32, [3]>([50, 12, 64])]; | |
| tensor<fp16, [50, 12, 64]> var_420_cast_fp16 = reshape(shape = var_419, x = k_19_cast_fp16)[name = tensor<string, []>("op_420_cast_fp16")]; | |
| tensor<int32, [3]> k_21_perm_0 = const()[name = tensor<string, []>("k_21_perm_0"), val = tensor<int32, [3]>([1, 0, 2])]; | |
| tensor<int32, [3]> var_426 = const()[name = tensor<string, []>("op_426"), val = tensor<int32, [3]>([50, 12, 64])]; | |
| tensor<fp16, [50, 12, 64]> var_427_cast_fp16 = reshape(shape = var_426, x = v_19_cast_fp16)[name = tensor<string, []>("op_427_cast_fp16")]; | |
| tensor<int32, [3]> v_21_perm_0 = const()[name = tensor<string, []>("v_21_perm_0"), val = tensor<int32, [3]>([1, 0, 2])]; | |
| tensor<int32, [4]> var_430 = const()[name = tensor<string, []>("op_430"), val = tensor<int32, [4]>([1, 12, 50, 64])]; | |
| tensor<fp16, [12, 50, 64]> q_21_cast_fp16 = transpose(perm = q_21_perm_0, x = var_413_cast_fp16)[name = tensor<string, []>("transpose_61")]; | |
| tensor<fp16, [1, 12, 50, 64]> q_23_cast_fp16 = reshape(shape = var_430, x = q_21_cast_fp16)[name = tensor<string, []>("q_23_cast_fp16")]; | |
| tensor<int32, [4]> var_432 = const()[name = tensor<string, []>("op_432"), val = tensor<int32, [4]>([1, 12, 50, 64])]; | |
| tensor<fp16, [12, 50, 64]> k_21_cast_fp16 = transpose(perm = k_21_perm_0, x = var_420_cast_fp16)[name = tensor<string, []>("transpose_60")]; | |
| tensor<fp16, [1, 12, 50, 64]> k_23_cast_fp16 = reshape(shape = var_432, x = k_21_cast_fp16)[name = tensor<string, []>("k_23_cast_fp16")]; | |
| tensor<int32, [4]> var_434 = const()[name = tensor<string, []>("op_434"), val = tensor<int32, [4]>([1, 12, 50, 64])]; | |
| tensor<fp16, [12, 50, 64]> v_21_cast_fp16 = transpose(perm = v_21_perm_0, x = var_427_cast_fp16)[name = tensor<string, []>("transpose_59")]; | |
| tensor<fp16, [1, 12, 50, 64]> v_23_cast_fp16 = reshape(shape = var_434, x = v_21_cast_fp16)[name = tensor<string, []>("v_23_cast_fp16")]; | |
| tensor<fp16, []> mul_7_y_0_to_fp16 = const()[name = tensor<string, []>("mul_7_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)]; | |
| tensor<fp16, [1, 12, 50, 64]> mul_7_cast_fp16 = mul(x = q_23_cast_fp16, y = mul_7_y_0_to_fp16)[name = tensor<string, []>("mul_7_cast_fp16")]; | |
| tensor<bool, []> matmul_3_transpose_y_0 = const()[name = tensor<string, []>("matmul_3_transpose_y_0"), val = tensor<bool, []>(true)]; | |
| tensor<bool, []> matmul_3_transpose_x_0 = const()[name = tensor<string, []>("matmul_3_transpose_x_0"), val = tensor<bool, []>(false)]; | |
| tensor<fp16, [1, 12, 50, 50]> matmul_3_cast_fp16 = matmul(transpose_x = matmul_3_transpose_x_0, transpose_y = matmul_3_transpose_y_0, x = mul_7_cast_fp16, y = k_23_cast_fp16)[name = tensor<string, []>("matmul_3_cast_fp16")]; | |
| tensor<int32, []> softmax_3_axis_0 = const()[name = tensor<string, []>("softmax_3_axis_0"), val = tensor<int32, []>(-1)]; | |
| tensor<fp16, [1, 12, 50, 50]> softmax_3_cast_fp16 = softmax(axis = softmax_3_axis_0, x = matmul_3_cast_fp16)[name = tensor<string, []>("softmax_3_cast_fp16")]; | |
| tensor<bool, []> attn_output_25_transpose_x_0 = const()[name = tensor<string, []>("attn_output_25_transpose_x_0"), val = tensor<bool, []>(false)]; | |
| tensor<bool, []> attn_output_25_transpose_y_0 = const()[name = tensor<string, []>("attn_output_25_transpose_y_0"), val = tensor<bool, []>(false)]; | |
| tensor<fp16, [1, 12, 50, 64]> attn_output_25_cast_fp16 = matmul(transpose_x = attn_output_25_transpose_x_0, transpose_y = attn_output_25_transpose_y_0, x = softmax_3_cast_fp16, y = v_23_cast_fp16)[name = tensor<string, []>("attn_output_25_cast_fp16")]; | |
| tensor<int32, [4]> var_437 = const()[name = tensor<string, []>("op_437"), val = tensor<int32, [4]>([2, 0, 1, 3])]; | |
| tensor<int32, [2]> var_442 = const()[name = tensor<string, []>("op_442"), val = tensor<int32, [2]>([50, 768])]; | |
| tensor<fp16, [50, 1, 12, 64]> var_438_cast_fp16 = transpose(perm = var_437, x = attn_output_25_cast_fp16)[name = tensor<string, []>("transpose_58")]; | |
| tensor<fp16, [50, 768]> attn_output_27_cast_fp16 = reshape(shape = var_442, x = var_438_cast_fp16)[name = tensor<string, []>("attn_output_27_cast_fp16")]; | |
| tensor<fp16, [768, 768]> model_visual_transformer_resblocks_3_attn_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("model_visual_transformer_resblocks_3_attn_out_proj_weight_to_fp16"), val = tensor<fp16, [768, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(50876800)))]; | |
| tensor<fp16, [768]> model_visual_transformer_resblocks_3_attn_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_visual_transformer_resblocks_3_attn_out_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(52056512)))]; | |
| tensor<fp16, [50, 768]> linear_13_cast_fp16 = linear(bias = model_visual_transformer_resblocks_3_attn_out_proj_bias_to_fp16, weight = model_visual_transformer_resblocks_3_attn_out_proj_weight_to_fp16, x = attn_output_27_cast_fp16)[name = tensor<string, []>("linear_13_cast_fp16")]; | |
| tensor<int32, [3]> var_446 = const()[name = tensor<string, []>("op_446"), val = tensor<int32, [3]>([50, 1, 768])]; | |
| tensor<fp16, [50, 1, 768]> attn_output_31_cast_fp16 = reshape(shape = var_446, x = linear_13_cast_fp16)[name = tensor<string, []>("attn_output_31_cast_fp16")]; | |
| tensor<int32, [3]> var_448_perm_0 = const()[name = tensor<string, []>("op_448_perm_0"), val = tensor<int32, [3]>([1, 0, 2])]; | |
| tensor<fp16, [1, 50, 768]> var_448_cast_fp16 = transpose(perm = var_448_perm_0, x = attn_output_31_cast_fp16)[name = tensor<string, []>("transpose_57")]; | |
| tensor<fp16, [1, 50, 768]> input_35_cast_fp16 = add(x = input_33_cast_fp16, y = var_448_cast_fp16)[name = tensor<string, []>("input_35_cast_fp16")]; | |
| tensor<int32, [1]> x_25_axes_0 = const()[name = tensor<string, []>("x_25_axes_0"), val = tensor<int32, [1]>([-1])]; | |
| tensor<fp16, [768]> model_visual_transformer_resblocks_3_ln_2_weight_to_fp16 = const()[name = tensor<string, []>("model_visual_transformer_resblocks_3_ln_2_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(52058112)))]; | |
| tensor<fp16, [768]> model_visual_transformer_resblocks_3_ln_2_bias_to_fp16 = const()[name = tensor<string, []>("model_visual_transformer_resblocks_3_ln_2_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(52059712)))]; | |
| tensor<fp16, [1, 50, 768]> x_25_cast_fp16 = layer_norm(axes = x_25_axes_0, beta = model_visual_transformer_resblocks_3_ln_2_bias_to_fp16, epsilon = var_13_to_fp16, gamma = model_visual_transformer_resblocks_3_ln_2_weight_to_fp16, x = input_35_cast_fp16)[name = tensor<string, []>("x_25_cast_fp16")]; | |
| tensor<fp16, [3072, 768]> model_visual_transformer_resblocks_3_mlp_c_fc_weight_to_fp16 = const()[name = tensor<string, []>("model_visual_transformer_resblocks_3_mlp_c_fc_weight_to_fp16"), val = tensor<fp16, [3072, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(52061312)))]; | |
| tensor<fp16, [3072]> model_visual_transformer_resblocks_3_mlp_c_fc_bias_to_fp16 = const()[name = tensor<string, []>("model_visual_transformer_resblocks_3_mlp_c_fc_bias_to_fp16"), val = tensor<fp16, [3072]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(56779968)))]; | |
| tensor<fp16, [1, 50, 3072]> linear_14_cast_fp16 = linear(bias = model_visual_transformer_resblocks_3_mlp_c_fc_bias_to_fp16, weight = model_visual_transformer_resblocks_3_mlp_c_fc_weight_to_fp16, x = x_25_cast_fp16)[name = tensor<string, []>("linear_14_cast_fp16")]; | |
| tensor<string, []> input_41_mode_0 = const()[name = tensor<string, []>("input_41_mode_0"), val = tensor<string, []>("EXACT")]; | |
| tensor<fp16, [1, 50, 3072]> input_41_cast_fp16 = gelu(mode = input_41_mode_0, x = linear_14_cast_fp16)[name = tensor<string, []>("input_41_cast_fp16")]; | |
| tensor<fp16, [768, 3072]> model_visual_transformer_resblocks_3_mlp_c_proj_weight_to_fp16 = const()[name = tensor<string, []>("model_visual_transformer_resblocks_3_mlp_c_proj_weight_to_fp16"), val = tensor<fp16, [768, 3072]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(56786176)))]; | |
| tensor<fp16, [768]> model_visual_transformer_resblocks_3_mlp_c_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_visual_transformer_resblocks_3_mlp_c_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(61504832)))]; | |
| tensor<fp16, [1, 50, 768]> linear_15_cast_fp16 = linear(bias = model_visual_transformer_resblocks_3_mlp_c_proj_bias_to_fp16, weight = model_visual_transformer_resblocks_3_mlp_c_proj_weight_to_fp16, x = input_41_cast_fp16)[name = tensor<string, []>("linear_15_cast_fp16")]; | |
| tensor<fp16, [1, 50, 768]> input_43_cast_fp16 = add(x = input_35_cast_fp16, y = linear_15_cast_fp16)[name = tensor<string, []>("input_43_cast_fp16")]; | |
| tensor<int32, [1]> x_27_axes_0 = const()[name = tensor<string, []>("x_27_axes_0"), val = tensor<int32, [1]>([-1])]; | |
| tensor<fp16, [768]> model_visual_transformer_resblocks_4_ln_1_weight_to_fp16 = const()[name = tensor<string, []>("model_visual_transformer_resblocks_4_ln_1_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(61506432)))]; | |
| tensor<fp16, [768]> model_visual_transformer_resblocks_4_ln_1_bias_to_fp16 = const()[name = tensor<string, []>("model_visual_transformer_resblocks_4_ln_1_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(61508032)))]; | |
| tensor<fp16, [1, 50, 768]> x_27_cast_fp16 = layer_norm(axes = x_27_axes_0, beta = model_visual_transformer_resblocks_4_ln_1_bias_to_fp16, epsilon = var_13_to_fp16, gamma = model_visual_transformer_resblocks_4_ln_1_weight_to_fp16, x = input_43_cast_fp16)[name = tensor<string, []>("x_27_cast_fp16")]; | |
| tensor<int32, [3]> query_19_perm_0 = const()[name = tensor<string, []>("query_19_perm_0"), val = tensor<int32, [3]>([1, 0, 2])]; | |
| tensor<fp16, [2304, 768]> model_visual_transformer_resblocks_4_attn_in_proj_weight_to_fp16 = const()[name = tensor<string, []>("model_visual_transformer_resblocks_4_attn_in_proj_weight_to_fp16"), val = tensor<fp16, [2304, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(61509632)))]; | |
| tensor<fp16, [2304]> model_visual_transformer_resblocks_4_attn_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_visual_transformer_resblocks_4_attn_in_proj_bias_to_fp16"), val = tensor<fp16, [2304]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(65048640)))]; | |
| tensor<fp16, [50, 1, 768]> query_19_cast_fp16 = transpose(perm = query_19_perm_0, x = x_27_cast_fp16)[name = tensor<string, []>("transpose_56")]; | |
| tensor<fp16, [50, 1, 2304]> linear_16_cast_fp16 = linear(bias = model_visual_transformer_resblocks_4_attn_in_proj_bias_to_fp16, weight = model_visual_transformer_resblocks_4_attn_in_proj_weight_to_fp16, x = query_19_cast_fp16)[name = tensor<string, []>("linear_16_cast_fp16")]; | |
| tensor<int32, [4]> concat_5 = const()[name = tensor<string, []>("concat_5"), val = tensor<int32, [4]>([50, 1, 3, 768])]; | |
| tensor<fp16, [50, 1, 3, 768]> var_497_cast_fp16 = reshape(shape = concat_5, x = linear_16_cast_fp16)[name = tensor<string, []>("op_497_cast_fp16")]; | |
| tensor<int32, [1]> var_498_axes_0 = const()[name = tensor<string, []>("op_498_axes_0"), val = tensor<int32, [1]>([0])]; | |
| tensor<fp16, [1, 50, 1, 3, 768]> var_498_cast_fp16 = expand_dims(axes = var_498_axes_0, x = var_497_cast_fp16)[name = tensor<string, []>("op_498_cast_fp16")]; | |
| tensor<int32, [5]> var_499_perm_0 = const()[name = tensor<string, []>("op_499_perm_0"), val = tensor<int32, [5]>([-2, 1, 2, 0, 4])]; | |
| tensor<int32, [1]> var_500_axes_0 = const()[name = tensor<string, []>("op_500_axes_0"), val = tensor<int32, [1]>([-2])]; | |
| tensor<fp16, [3, 50, 1, 1, 768]> var_499_cast_fp16 = transpose(perm = var_499_perm_0, x = var_498_cast_fp16)[name = tensor<string, []>("transpose_55")]; | |
| tensor<fp16, [3, 50, 1, 768]> var_500_cast_fp16 = squeeze(axes = var_500_axes_0, x = var_499_cast_fp16)[name = tensor<string, []>("op_500_cast_fp16")]; | |
| tensor<int32, [4]> q_25_begin_0 = const()[name = tensor<string, []>("q_25_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; | |
| tensor<int32, [4]> q_25_end_0 = const()[name = tensor<string, []>("q_25_end_0"), val = tensor<int32, [4]>([1, 50, 1, 768])]; | |
| tensor<bool, [4]> q_25_end_mask_0 = const()[name = tensor<string, []>("q_25_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])]; | |
| tensor<bool, [4]> q_25_squeeze_mask_0 = const()[name = tensor<string, []>("q_25_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])]; | |
| tensor<fp16, [50, 1, 768]> q_25_cast_fp16 = slice_by_index(begin = q_25_begin_0, end = q_25_end_0, end_mask = q_25_end_mask_0, squeeze_mask = q_25_squeeze_mask_0, x = var_500_cast_fp16)[name = tensor<string, []>("q_25_cast_fp16")]; | |
| tensor<int32, [4]> k_25_begin_0 = const()[name = tensor<string, []>("k_25_begin_0"), val = tensor<int32, [4]>([1, 0, 0, 0])]; | |
| tensor<int32, [4]> k_25_end_0 = const()[name = tensor<string, []>("k_25_end_0"), val = tensor<int32, [4]>([2, 50, 1, 768])]; | |
| tensor<bool, [4]> k_25_end_mask_0 = const()[name = tensor<string, []>("k_25_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])]; | |
| tensor<bool, [4]> k_25_squeeze_mask_0 = const()[name = tensor<string, []>("k_25_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])]; | |
| tensor<fp16, [50, 1, 768]> k_25_cast_fp16 = slice_by_index(begin = k_25_begin_0, end = k_25_end_0, end_mask = k_25_end_mask_0, squeeze_mask = k_25_squeeze_mask_0, x = var_500_cast_fp16)[name = tensor<string, []>("k_25_cast_fp16")]; | |
| tensor<int32, [4]> v_25_begin_0 = const()[name = tensor<string, []>("v_25_begin_0"), val = tensor<int32, [4]>([2, 0, 0, 0])]; | |
| tensor<int32, [4]> v_25_end_0 = const()[name = tensor<string, []>("v_25_end_0"), val = tensor<int32, [4]>([3, 50, 1, 768])]; | |
| tensor<bool, [4]> v_25_end_mask_0 = const()[name = tensor<string, []>("v_25_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])]; | |
| tensor<bool, [4]> v_25_squeeze_mask_0 = const()[name = tensor<string, []>("v_25_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])]; | |
| tensor<fp16, [50, 1, 768]> v_25_cast_fp16 = slice_by_index(begin = v_25_begin_0, end = v_25_end_0, end_mask = v_25_end_mask_0, squeeze_mask = v_25_squeeze_mask_0, x = var_500_cast_fp16)[name = tensor<string, []>("v_25_cast_fp16")]; | |
| tensor<int32, [3]> var_508 = const()[name = tensor<string, []>("op_508"), val = tensor<int32, [3]>([50, 12, 64])]; | |
| tensor<fp16, [50, 12, 64]> var_509_cast_fp16 = reshape(shape = var_508, x = q_25_cast_fp16)[name = tensor<string, []>("op_509_cast_fp16")]; | |
| tensor<int32, [3]> q_27_perm_0 = const()[name = tensor<string, []>("q_27_perm_0"), val = tensor<int32, [3]>([1, 0, 2])]; | |
| tensor<int32, [3]> var_515 = const()[name = tensor<string, []>("op_515"), val = tensor<int32, [3]>([50, 12, 64])]; | |
| tensor<fp16, [50, 12, 64]> var_516_cast_fp16 = reshape(shape = var_515, x = k_25_cast_fp16)[name = tensor<string, []>("op_516_cast_fp16")]; | |
| tensor<int32, [3]> k_27_perm_0 = const()[name = tensor<string, []>("k_27_perm_0"), val = tensor<int32, [3]>([1, 0, 2])]; | |
| tensor<int32, [3]> var_522 = const()[name = tensor<string, []>("op_522"), val = tensor<int32, [3]>([50, 12, 64])]; | |
| tensor<fp16, [50, 12, 64]> var_523_cast_fp16 = reshape(shape = var_522, x = v_25_cast_fp16)[name = tensor<string, []>("op_523_cast_fp16")]; | |
| tensor<int32, [3]> v_27_perm_0 = const()[name = tensor<string, []>("v_27_perm_0"), val = tensor<int32, [3]>([1, 0, 2])]; | |
| tensor<int32, [4]> var_526 = const()[name = tensor<string, []>("op_526"), val = tensor<int32, [4]>([1, 12, 50, 64])]; | |
| tensor<fp16, [12, 50, 64]> q_27_cast_fp16 = transpose(perm = q_27_perm_0, x = var_509_cast_fp16)[name = tensor<string, []>("transpose_54")]; | |
| tensor<fp16, [1, 12, 50, 64]> q_29_cast_fp16 = reshape(shape = var_526, x = q_27_cast_fp16)[name = tensor<string, []>("q_29_cast_fp16")]; | |
| tensor<int32, [4]> var_528 = const()[name = tensor<string, []>("op_528"), val = tensor<int32, [4]>([1, 12, 50, 64])]; | |
| tensor<fp16, [12, 50, 64]> k_27_cast_fp16 = transpose(perm = k_27_perm_0, x = var_516_cast_fp16)[name = tensor<string, []>("transpose_53")]; | |
| tensor<fp16, [1, 12, 50, 64]> k_29_cast_fp16 = reshape(shape = var_528, x = k_27_cast_fp16)[name = tensor<string, []>("k_29_cast_fp16")]; | |
| tensor<int32, [4]> var_530 = const()[name = tensor<string, []>("op_530"), val = tensor<int32, [4]>([1, 12, 50, 64])]; | |
| tensor<fp16, [12, 50, 64]> v_27_cast_fp16 = transpose(perm = v_27_perm_0, x = var_523_cast_fp16)[name = tensor<string, []>("transpose_52")]; | |
| tensor<fp16, [1, 12, 50, 64]> v_29_cast_fp16 = reshape(shape = var_530, x = v_27_cast_fp16)[name = tensor<string, []>("v_29_cast_fp16")]; | |
| tensor<fp16, []> mul_9_y_0_to_fp16 = const()[name = tensor<string, []>("mul_9_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)]; | |
| tensor<fp16, [1, 12, 50, 64]> mul_9_cast_fp16 = mul(x = q_29_cast_fp16, y = mul_9_y_0_to_fp16)[name = tensor<string, []>("mul_9_cast_fp16")]; | |
| tensor<bool, []> matmul_4_transpose_y_0 = const()[name = tensor<string, []>("matmul_4_transpose_y_0"), val = tensor<bool, []>(true)]; | |
| tensor<bool, []> matmul_4_transpose_x_0 = const()[name = tensor<string, []>("matmul_4_transpose_x_0"), val = tensor<bool, []>(false)]; | |
| tensor<fp16, [1, 12, 50, 50]> matmul_4_cast_fp16 = matmul(transpose_x = matmul_4_transpose_x_0, transpose_y = matmul_4_transpose_y_0, x = mul_9_cast_fp16, y = k_29_cast_fp16)[name = tensor<string, []>("matmul_4_cast_fp16")]; | |
| tensor<int32, []> softmax_4_axis_0 = const()[name = tensor<string, []>("softmax_4_axis_0"), val = tensor<int32, []>(-1)]; | |
| tensor<fp16, [1, 12, 50, 50]> softmax_4_cast_fp16 = softmax(axis = softmax_4_axis_0, x = matmul_4_cast_fp16)[name = tensor<string, []>("softmax_4_cast_fp16")]; | |
| tensor<bool, []> attn_output_33_transpose_x_0 = const()[name = tensor<string, []>("attn_output_33_transpose_x_0"), val = tensor<bool, []>(false)]; | |
| tensor<bool, []> attn_output_33_transpose_y_0 = const()[name = tensor<string, []>("attn_output_33_transpose_y_0"), val = tensor<bool, []>(false)]; | |
| tensor<fp16, [1, 12, 50, 64]> attn_output_33_cast_fp16 = matmul(transpose_x = attn_output_33_transpose_x_0, transpose_y = attn_output_33_transpose_y_0, x = softmax_4_cast_fp16, y = v_29_cast_fp16)[name = tensor<string, []>("attn_output_33_cast_fp16")]; | |
| tensor<int32, [4]> var_533 = const()[name = tensor<string, []>("op_533"), val = tensor<int32, [4]>([2, 0, 1, 3])]; | |
| tensor<int32, [2]> var_538 = const()[name = tensor<string, []>("op_538"), val = tensor<int32, [2]>([50, 768])]; | |
| tensor<fp16, [50, 1, 12, 64]> var_534_cast_fp16 = transpose(perm = var_533, x = attn_output_33_cast_fp16)[name = tensor<string, []>("transpose_51")]; | |
| tensor<fp16, [50, 768]> attn_output_35_cast_fp16 = reshape(shape = var_538, x = var_534_cast_fp16)[name = tensor<string, []>("attn_output_35_cast_fp16")]; | |
| tensor<fp16, [768, 768]> model_visual_transformer_resblocks_4_attn_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("model_visual_transformer_resblocks_4_attn_out_proj_weight_to_fp16"), val = tensor<fp16, [768, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(65053312)))]; | |
| tensor<fp16, [768]> model_visual_transformer_resblocks_4_attn_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_visual_transformer_resblocks_4_attn_out_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(66233024)))]; | |
| tensor<fp16, [50, 768]> linear_17_cast_fp16 = linear(bias = model_visual_transformer_resblocks_4_attn_out_proj_bias_to_fp16, weight = model_visual_transformer_resblocks_4_attn_out_proj_weight_to_fp16, x = attn_output_35_cast_fp16)[name = tensor<string, []>("linear_17_cast_fp16")]; | |
| tensor<int32, [3]> var_542 = const()[name = tensor<string, []>("op_542"), val = tensor<int32, [3]>([50, 1, 768])]; | |
| tensor<fp16, [50, 1, 768]> attn_output_39_cast_fp16 = reshape(shape = var_542, x = linear_17_cast_fp16)[name = tensor<string, []>("attn_output_39_cast_fp16")]; | |
| tensor<int32, [3]> var_544_perm_0 = const()[name = tensor<string, []>("op_544_perm_0"), val = tensor<int32, [3]>([1, 0, 2])]; | |
| tensor<fp16, [1, 50, 768]> var_544_cast_fp16 = transpose(perm = var_544_perm_0, x = attn_output_39_cast_fp16)[name = tensor<string, []>("transpose_50")]; | |
| tensor<fp16, [1, 50, 768]> input_45_cast_fp16 = add(x = input_43_cast_fp16, y = var_544_cast_fp16)[name = tensor<string, []>("input_45_cast_fp16")]; | |
| tensor<int32, [1]> x_29_axes_0 = const()[name = tensor<string, []>("x_29_axes_0"), val = tensor<int32, [1]>([-1])]; | |
| tensor<fp16, [768]> model_visual_transformer_resblocks_4_ln_2_weight_to_fp16 = const()[name = tensor<string, []>("model_visual_transformer_resblocks_4_ln_2_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(66234624)))]; | |
| tensor<fp16, [768]> model_visual_transformer_resblocks_4_ln_2_bias_to_fp16 = const()[name = tensor<string, []>("model_visual_transformer_resblocks_4_ln_2_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(66236224)))]; | |
| tensor<fp16, [1, 50, 768]> x_29_cast_fp16 = layer_norm(axes = x_29_axes_0, beta = model_visual_transformer_resblocks_4_ln_2_bias_to_fp16, epsilon = var_13_to_fp16, gamma = model_visual_transformer_resblocks_4_ln_2_weight_to_fp16, x = input_45_cast_fp16)[name = tensor<string, []>("x_29_cast_fp16")]; | |
| tensor<fp16, [3072, 768]> model_visual_transformer_resblocks_4_mlp_c_fc_weight_to_fp16 = const()[name = tensor<string, []>("model_visual_transformer_resblocks_4_mlp_c_fc_weight_to_fp16"), val = tensor<fp16, [3072, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(66237824)))]; | |
| tensor<fp16, [3072]> model_visual_transformer_resblocks_4_mlp_c_fc_bias_to_fp16 = const()[name = tensor<string, []>("model_visual_transformer_resblocks_4_mlp_c_fc_bias_to_fp16"), val = tensor<fp16, [3072]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(70956480)))]; | |
| tensor<fp16, [1, 50, 3072]> linear_18_cast_fp16 = linear(bias = model_visual_transformer_resblocks_4_mlp_c_fc_bias_to_fp16, weight = model_visual_transformer_resblocks_4_mlp_c_fc_weight_to_fp16, x = x_29_cast_fp16)[name = tensor<string, []>("linear_18_cast_fp16")]; | |
| tensor<string, []> input_51_mode_0 = const()[name = tensor<string, []>("input_51_mode_0"), val = tensor<string, []>("EXACT")]; | |
| tensor<fp16, [1, 50, 3072]> input_51_cast_fp16 = gelu(mode = input_51_mode_0, x = linear_18_cast_fp16)[name = tensor<string, []>("input_51_cast_fp16")]; | |
| tensor<fp16, [768, 3072]> model_visual_transformer_resblocks_4_mlp_c_proj_weight_to_fp16 = const()[name = tensor<string, []>("model_visual_transformer_resblocks_4_mlp_c_proj_weight_to_fp16"), val = tensor<fp16, [768, 3072]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(70962688)))]; | |
| tensor<fp16, [768]> model_visual_transformer_resblocks_4_mlp_c_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_visual_transformer_resblocks_4_mlp_c_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(75681344)))]; | |
| tensor<fp16, [1, 50, 768]> linear_19_cast_fp16 = linear(bias = model_visual_transformer_resblocks_4_mlp_c_proj_bias_to_fp16, weight = model_visual_transformer_resblocks_4_mlp_c_proj_weight_to_fp16, x = input_51_cast_fp16)[name = tensor<string, []>("linear_19_cast_fp16")]; | |
| tensor<fp16, [1, 50, 768]> input_53_cast_fp16 = add(x = input_45_cast_fp16, y = linear_19_cast_fp16)[name = tensor<string, []>("input_53_cast_fp16")]; | |
| tensor<int32, [1]> x_31_axes_0 = const()[name = tensor<string, []>("x_31_axes_0"), val = tensor<int32, [1]>([-1])]; | |
| tensor<fp16, [768]> model_visual_transformer_resblocks_5_ln_1_weight_to_fp16 = const()[name = tensor<string, []>("model_visual_transformer_resblocks_5_ln_1_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(75682944)))]; | |
| tensor<fp16, [768]> model_visual_transformer_resblocks_5_ln_1_bias_to_fp16 = const()[name = tensor<string, []>("model_visual_transformer_resblocks_5_ln_1_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(75684544)))]; | |
| tensor<fp16, [1, 50, 768]> x_31_cast_fp16 = layer_norm(axes = x_31_axes_0, beta = model_visual_transformer_resblocks_5_ln_1_bias_to_fp16, epsilon = var_13_to_fp16, gamma = model_visual_transformer_resblocks_5_ln_1_weight_to_fp16, x = input_53_cast_fp16)[name = tensor<string, []>("x_31_cast_fp16")]; | |
| tensor<int32, [3]> query_23_perm_0 = const()[name = tensor<string, []>("query_23_perm_0"), val = tensor<int32, [3]>([1, 0, 2])]; | |
| tensor<fp16, [2304, 768]> model_visual_transformer_resblocks_5_attn_in_proj_weight_to_fp16 = const()[name = tensor<string, []>("model_visual_transformer_resblocks_5_attn_in_proj_weight_to_fp16"), val = tensor<fp16, [2304, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(75686144)))]; | |
| tensor<fp16, [2304]> model_visual_transformer_resblocks_5_attn_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_visual_transformer_resblocks_5_attn_in_proj_bias_to_fp16"), val = tensor<fp16, [2304]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(79225152)))]; | |
| tensor<fp16, [50, 1, 768]> query_23_cast_fp16 = transpose(perm = query_23_perm_0, x = x_31_cast_fp16)[name = tensor<string, []>("transpose_49")]; | |
| tensor<fp16, [50, 1, 2304]> linear_20_cast_fp16 = linear(bias = model_visual_transformer_resblocks_5_attn_in_proj_bias_to_fp16, weight = model_visual_transformer_resblocks_5_attn_in_proj_weight_to_fp16, x = query_23_cast_fp16)[name = tensor<string, []>("linear_20_cast_fp16")]; | |
| tensor<int32, [4]> concat_6 = const()[name = tensor<string, []>("concat_6"), val = tensor<int32, [4]>([50, 1, 3, 768])]; | |
| tensor<fp16, [50, 1, 3, 768]> var_593_cast_fp16 = reshape(shape = concat_6, x = linear_20_cast_fp16)[name = tensor<string, []>("op_593_cast_fp16")]; | |
| tensor<int32, [1]> var_594_axes_0 = const()[name = tensor<string, []>("op_594_axes_0"), val = tensor<int32, [1]>([0])]; | |
| tensor<fp16, [1, 50, 1, 3, 768]> var_594_cast_fp16 = expand_dims(axes = var_594_axes_0, x = var_593_cast_fp16)[name = tensor<string, []>("op_594_cast_fp16")]; | |
| tensor<int32, [5]> var_595_perm_0 = const()[name = tensor<string, []>("op_595_perm_0"), val = tensor<int32, [5]>([-2, 1, 2, 0, 4])]; | |
| tensor<int32, [1]> var_596_axes_0 = const()[name = tensor<string, []>("op_596_axes_0"), val = tensor<int32, [1]>([-2])]; | |
| tensor<fp16, [3, 50, 1, 1, 768]> var_595_cast_fp16 = transpose(perm = var_595_perm_0, x = var_594_cast_fp16)[name = tensor<string, []>("transpose_48")]; | |
| tensor<fp16, [3, 50, 1, 768]> var_596_cast_fp16 = squeeze(axes = var_596_axes_0, x = var_595_cast_fp16)[name = tensor<string, []>("op_596_cast_fp16")]; | |
| tensor<int32, [4]> q_31_begin_0 = const()[name = tensor<string, []>("q_31_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; | |
| tensor<int32, [4]> q_31_end_0 = const()[name = tensor<string, []>("q_31_end_0"), val = tensor<int32, [4]>([1, 50, 1, 768])]; | |
| tensor<bool, [4]> q_31_end_mask_0 = const()[name = tensor<string, []>("q_31_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])]; | |
| tensor<bool, [4]> q_31_squeeze_mask_0 = const()[name = tensor<string, []>("q_31_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])]; | |
| tensor<fp16, [50, 1, 768]> q_31_cast_fp16 = slice_by_index(begin = q_31_begin_0, end = q_31_end_0, end_mask = q_31_end_mask_0, squeeze_mask = q_31_squeeze_mask_0, x = var_596_cast_fp16)[name = tensor<string, []>("q_31_cast_fp16")]; | |
| tensor<int32, [4]> k_31_begin_0 = const()[name = tensor<string, []>("k_31_begin_0"), val = tensor<int32, [4]>([1, 0, 0, 0])]; | |
| tensor<int32, [4]> k_31_end_0 = const()[name = tensor<string, []>("k_31_end_0"), val = tensor<int32, [4]>([2, 50, 1, 768])]; | |
| tensor<bool, [4]> k_31_end_mask_0 = const()[name = tensor<string, []>("k_31_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])]; | |
| tensor<bool, [4]> k_31_squeeze_mask_0 = const()[name = tensor<string, []>("k_31_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])]; | |
| tensor<fp16, [50, 1, 768]> k_31_cast_fp16 = slice_by_index(begin = k_31_begin_0, end = k_31_end_0, end_mask = k_31_end_mask_0, squeeze_mask = k_31_squeeze_mask_0, x = var_596_cast_fp16)[name = tensor<string, []>("k_31_cast_fp16")]; | |
| tensor<int32, [4]> v_31_begin_0 = const()[name = tensor<string, []>("v_31_begin_0"), val = tensor<int32, [4]>([2, 0, 0, 0])]; | |
| tensor<int32, [4]> v_31_end_0 = const()[name = tensor<string, []>("v_31_end_0"), val = tensor<int32, [4]>([3, 50, 1, 768])]; | |
| tensor<bool, [4]> v_31_end_mask_0 = const()[name = tensor<string, []>("v_31_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])]; | |
| tensor<bool, [4]> v_31_squeeze_mask_0 = const()[name = tensor<string, []>("v_31_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])]; | |
| tensor<fp16, [50, 1, 768]> v_31_cast_fp16 = slice_by_index(begin = v_31_begin_0, end = v_31_end_0, end_mask = v_31_end_mask_0, squeeze_mask = v_31_squeeze_mask_0, x = var_596_cast_fp16)[name = tensor<string, []>("v_31_cast_fp16")]; | |
| tensor<int32, [3]> var_604 = const()[name = tensor<string, []>("op_604"), val = tensor<int32, [3]>([50, 12, 64])]; | |
| tensor<fp16, [50, 12, 64]> var_605_cast_fp16 = reshape(shape = var_604, x = q_31_cast_fp16)[name = tensor<string, []>("op_605_cast_fp16")]; | |
| tensor<int32, [3]> q_33_perm_0 = const()[name = tensor<string, []>("q_33_perm_0"), val = tensor<int32, [3]>([1, 0, 2])]; | |
| tensor<int32, [3]> var_611 = const()[name = tensor<string, []>("op_611"), val = tensor<int32, [3]>([50, 12, 64])]; | |
| tensor<fp16, [50, 12, 64]> var_612_cast_fp16 = reshape(shape = var_611, x = k_31_cast_fp16)[name = tensor<string, []>("op_612_cast_fp16")]; | |
| tensor<int32, [3]> k_33_perm_0 = const()[name = tensor<string, []>("k_33_perm_0"), val = tensor<int32, [3]>([1, 0, 2])]; | |
| tensor<int32, [3]> var_618 = const()[name = tensor<string, []>("op_618"), val = tensor<int32, [3]>([50, 12, 64])]; | |
| tensor<fp16, [50, 12, 64]> var_619_cast_fp16 = reshape(shape = var_618, x = v_31_cast_fp16)[name = tensor<string, []>("op_619_cast_fp16")]; | |
| tensor<int32, [3]> v_33_perm_0 = const()[name = tensor<string, []>("v_33_perm_0"), val = tensor<int32, [3]>([1, 0, 2])]; | |
| tensor<int32, [4]> var_622 = const()[name = tensor<string, []>("op_622"), val = tensor<int32, [4]>([1, 12, 50, 64])]; | |
| tensor<fp16, [12, 50, 64]> q_33_cast_fp16 = transpose(perm = q_33_perm_0, x = var_605_cast_fp16)[name = tensor<string, []>("transpose_47")]; | |
| tensor<fp16, [1, 12, 50, 64]> q_35_cast_fp16 = reshape(shape = var_622, x = q_33_cast_fp16)[name = tensor<string, []>("q_35_cast_fp16")]; | |
| tensor<int32, [4]> var_624 = const()[name = tensor<string, []>("op_624"), val = tensor<int32, [4]>([1, 12, 50, 64])]; | |
| tensor<fp16, [12, 50, 64]> k_33_cast_fp16 = transpose(perm = k_33_perm_0, x = var_612_cast_fp16)[name = tensor<string, []>("transpose_46")]; | |
| tensor<fp16, [1, 12, 50, 64]> k_35_cast_fp16 = reshape(shape = var_624, x = k_33_cast_fp16)[name = tensor<string, []>("k_35_cast_fp16")]; | |
| tensor<int32, [4]> var_626 = const()[name = tensor<string, []>("op_626"), val = tensor<int32, [4]>([1, 12, 50, 64])]; | |
| tensor<fp16, [12, 50, 64]> v_33_cast_fp16 = transpose(perm = v_33_perm_0, x = var_619_cast_fp16)[name = tensor<string, []>("transpose_45")]; | |
| tensor<fp16, [1, 12, 50, 64]> v_35_cast_fp16 = reshape(shape = var_626, x = v_33_cast_fp16)[name = tensor<string, []>("v_35_cast_fp16")]; | |
| tensor<fp16, []> mul_11_y_0_to_fp16 = const()[name = tensor<string, []>("mul_11_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)]; | |
| tensor<fp16, [1, 12, 50, 64]> mul_11_cast_fp16 = mul(x = q_35_cast_fp16, y = mul_11_y_0_to_fp16)[name = tensor<string, []>("mul_11_cast_fp16")]; | |
| tensor<bool, []> matmul_5_transpose_y_0 = const()[name = tensor<string, []>("matmul_5_transpose_y_0"), val = tensor<bool, []>(true)]; | |
| tensor<bool, []> matmul_5_transpose_x_0 = const()[name = tensor<string, []>("matmul_5_transpose_x_0"), val = tensor<bool, []>(false)]; | |
| tensor<fp16, [1, 12, 50, 50]> matmul_5_cast_fp16 = matmul(transpose_x = matmul_5_transpose_x_0, transpose_y = matmul_5_transpose_y_0, x = mul_11_cast_fp16, y = k_35_cast_fp16)[name = tensor<string, []>("matmul_5_cast_fp16")]; | |
| tensor<int32, []> softmax_5_axis_0 = const()[name = tensor<string, []>("softmax_5_axis_0"), val = tensor<int32, []>(-1)]; | |
| tensor<fp16, [1, 12, 50, 50]> softmax_5_cast_fp16 = softmax(axis = softmax_5_axis_0, x = matmul_5_cast_fp16)[name = tensor<string, []>("softmax_5_cast_fp16")]; | |
| tensor<bool, []> attn_output_41_transpose_x_0 = const()[name = tensor<string, []>("attn_output_41_transpose_x_0"), val = tensor<bool, []>(false)]; | |
| tensor<bool, []> attn_output_41_transpose_y_0 = const()[name = tensor<string, []>("attn_output_41_transpose_y_0"), val = tensor<bool, []>(false)]; | |
| tensor<fp16, [1, 12, 50, 64]> attn_output_41_cast_fp16 = matmul(transpose_x = attn_output_41_transpose_x_0, transpose_y = attn_output_41_transpose_y_0, x = softmax_5_cast_fp16, y = v_35_cast_fp16)[name = tensor<string, []>("attn_output_41_cast_fp16")]; | |
| tensor<int32, [4]> var_629 = const()[name = tensor<string, []>("op_629"), val = tensor<int32, [4]>([2, 0, 1, 3])]; | |
| tensor<int32, [2]> var_634 = const()[name = tensor<string, []>("op_634"), val = tensor<int32, [2]>([50, 768])]; | |
| tensor<fp16, [50, 1, 12, 64]> var_630_cast_fp16 = transpose(perm = var_629, x = attn_output_41_cast_fp16)[name = tensor<string, []>("transpose_44")]; | |
| tensor<fp16, [50, 768]> attn_output_43_cast_fp16 = reshape(shape = var_634, x = var_630_cast_fp16)[name = tensor<string, []>("attn_output_43_cast_fp16")]; | |
| tensor<fp16, [768, 768]> model_visual_transformer_resblocks_5_attn_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("model_visual_transformer_resblocks_5_attn_out_proj_weight_to_fp16"), val = tensor<fp16, [768, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(79229824)))]; | |
| tensor<fp16, [768]> model_visual_transformer_resblocks_5_attn_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_visual_transformer_resblocks_5_attn_out_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(80409536)))]; | |
| tensor<fp16, [50, 768]> linear_21_cast_fp16 = linear(bias = model_visual_transformer_resblocks_5_attn_out_proj_bias_to_fp16, weight = model_visual_transformer_resblocks_5_attn_out_proj_weight_to_fp16, x = attn_output_43_cast_fp16)[name = tensor<string, []>("linear_21_cast_fp16")]; | |
| tensor<int32, [3]> var_638 = const()[name = tensor<string, []>("op_638"), val = tensor<int32, [3]>([50, 1, 768])]; | |
| tensor<fp16, [50, 1, 768]> attn_output_47_cast_fp16 = reshape(shape = var_638, x = linear_21_cast_fp16)[name = tensor<string, []>("attn_output_47_cast_fp16")]; | |
| tensor<int32, [3]> var_640_perm_0 = const()[name = tensor<string, []>("op_640_perm_0"), val = tensor<int32, [3]>([1, 0, 2])]; | |
| tensor<fp16, [1, 50, 768]> var_640_cast_fp16 = transpose(perm = var_640_perm_0, x = attn_output_47_cast_fp16)[name = tensor<string, []>("transpose_43")]; | |
| tensor<fp16, [1, 50, 768]> input_55_cast_fp16 = add(x = input_53_cast_fp16, y = var_640_cast_fp16)[name = tensor<string, []>("input_55_cast_fp16")]; | |
| tensor<int32, [1]> x_33_axes_0 = const()[name = tensor<string, []>("x_33_axes_0"), val = tensor<int32, [1]>([-1])]; | |
| tensor<fp16, [768]> model_visual_transformer_resblocks_5_ln_2_weight_to_fp16 = const()[name = tensor<string, []>("model_visual_transformer_resblocks_5_ln_2_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(80411136)))]; | |
| tensor<fp16, [768]> model_visual_transformer_resblocks_5_ln_2_bias_to_fp16 = const()[name = tensor<string, []>("model_visual_transformer_resblocks_5_ln_2_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(80412736)))]; | |
| tensor<fp16, [1, 50, 768]> x_33_cast_fp16 = layer_norm(axes = x_33_axes_0, beta = model_visual_transformer_resblocks_5_ln_2_bias_to_fp16, epsilon = var_13_to_fp16, gamma = model_visual_transformer_resblocks_5_ln_2_weight_to_fp16, x = input_55_cast_fp16)[name = tensor<string, []>("x_33_cast_fp16")]; | |
| tensor<fp16, [3072, 768]> model_visual_transformer_resblocks_5_mlp_c_fc_weight_to_fp16 = const()[name = tensor<string, []>("model_visual_transformer_resblocks_5_mlp_c_fc_weight_to_fp16"), val = tensor<fp16, [3072, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(80414336)))]; | |
| tensor<fp16, [3072]> model_visual_transformer_resblocks_5_mlp_c_fc_bias_to_fp16 = const()[name = tensor<string, []>("model_visual_transformer_resblocks_5_mlp_c_fc_bias_to_fp16"), val = tensor<fp16, [3072]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(85132992)))]; | |
| tensor<fp16, [1, 50, 3072]> linear_22_cast_fp16 = linear(bias = model_visual_transformer_resblocks_5_mlp_c_fc_bias_to_fp16, weight = model_visual_transformer_resblocks_5_mlp_c_fc_weight_to_fp16, x = x_33_cast_fp16)[name = tensor<string, []>("linear_22_cast_fp16")]; | |
| tensor<string, []> input_61_mode_0 = const()[name = tensor<string, []>("input_61_mode_0"), val = tensor<string, []>("EXACT")]; | |
| tensor<fp16, [1, 50, 3072]> input_61_cast_fp16 = gelu(mode = input_61_mode_0, x = linear_22_cast_fp16)[name = tensor<string, []>("input_61_cast_fp16")]; | |
| tensor<fp16, [768, 3072]> model_visual_transformer_resblocks_5_mlp_c_proj_weight_to_fp16 = const()[name = tensor<string, []>("model_visual_transformer_resblocks_5_mlp_c_proj_weight_to_fp16"), val = tensor<fp16, [768, 3072]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(85139200)))]; | |
| tensor<fp16, [768]> model_visual_transformer_resblocks_5_mlp_c_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_visual_transformer_resblocks_5_mlp_c_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(89857856)))]; | |
| tensor<fp16, [1, 50, 768]> linear_23_cast_fp16 = linear(bias = model_visual_transformer_resblocks_5_mlp_c_proj_bias_to_fp16, weight = model_visual_transformer_resblocks_5_mlp_c_proj_weight_to_fp16, x = input_61_cast_fp16)[name = tensor<string, []>("linear_23_cast_fp16")]; | |
| tensor<fp16, [1, 50, 768]> input_63_cast_fp16 = add(x = input_55_cast_fp16, y = linear_23_cast_fp16)[name = tensor<string, []>("input_63_cast_fp16")]; | |
| tensor<int32, [1]> x_35_axes_0 = const()[name = tensor<string, []>("x_35_axes_0"), val = tensor<int32, [1]>([-1])]; | |
| tensor<fp16, [768]> model_visual_transformer_resblocks_6_ln_1_weight_to_fp16 = const()[name = tensor<string, []>("model_visual_transformer_resblocks_6_ln_1_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(89859456)))]; | |
| tensor<fp16, [768]> model_visual_transformer_resblocks_6_ln_1_bias_to_fp16 = const()[name = tensor<string, []>("model_visual_transformer_resblocks_6_ln_1_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(89861056)))]; | |
| tensor<fp16, [1, 50, 768]> x_35_cast_fp16 = layer_norm(axes = x_35_axes_0, beta = model_visual_transformer_resblocks_6_ln_1_bias_to_fp16, epsilon = var_13_to_fp16, gamma = model_visual_transformer_resblocks_6_ln_1_weight_to_fp16, x = input_63_cast_fp16)[name = tensor<string, []>("x_35_cast_fp16")]; | |
| tensor<int32, [3]> query_27_perm_0 = const()[name = tensor<string, []>("query_27_perm_0"), val = tensor<int32, [3]>([1, 0, 2])]; | |
| tensor<fp16, [2304, 768]> model_visual_transformer_resblocks_6_attn_in_proj_weight_to_fp16 = const()[name = tensor<string, []>("model_visual_transformer_resblocks_6_attn_in_proj_weight_to_fp16"), val = tensor<fp16, [2304, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(89862656)))]; | |
| tensor<fp16, [2304]> model_visual_transformer_resblocks_6_attn_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_visual_transformer_resblocks_6_attn_in_proj_bias_to_fp16"), val = tensor<fp16, [2304]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(93401664)))]; | |
| tensor<fp16, [50, 1, 768]> query_27_cast_fp16 = transpose(perm = query_27_perm_0, x = x_35_cast_fp16)[name = tensor<string, []>("transpose_42")]; | |
| tensor<fp16, [50, 1, 2304]> linear_24_cast_fp16 = linear(bias = model_visual_transformer_resblocks_6_attn_in_proj_bias_to_fp16, weight = model_visual_transformer_resblocks_6_attn_in_proj_weight_to_fp16, x = query_27_cast_fp16)[name = tensor<string, []>("linear_24_cast_fp16")]; | |
| tensor<int32, [4]> concat_7 = const()[name = tensor<string, []>("concat_7"), val = tensor<int32, [4]>([50, 1, 3, 768])]; | |
| tensor<fp16, [50, 1, 3, 768]> var_689_cast_fp16 = reshape(shape = concat_7, x = linear_24_cast_fp16)[name = tensor<string, []>("op_689_cast_fp16")]; | |
| tensor<int32, [1]> var_690_axes_0 = const()[name = tensor<string, []>("op_690_axes_0"), val = tensor<int32, [1]>([0])]; | |
| tensor<fp16, [1, 50, 1, 3, 768]> var_690_cast_fp16 = expand_dims(axes = var_690_axes_0, x = var_689_cast_fp16)[name = tensor<string, []>("op_690_cast_fp16")]; | |
| tensor<int32, [5]> var_691_perm_0 = const()[name = tensor<string, []>("op_691_perm_0"), val = tensor<int32, [5]>([-2, 1, 2, 0, 4])]; | |
| tensor<int32, [1]> var_692_axes_0 = const()[name = tensor<string, []>("op_692_axes_0"), val = tensor<int32, [1]>([-2])]; | |
| tensor<fp16, [3, 50, 1, 1, 768]> var_691_cast_fp16 = transpose(perm = var_691_perm_0, x = var_690_cast_fp16)[name = tensor<string, []>("transpose_41")]; | |
| tensor<fp16, [3, 50, 1, 768]> var_692_cast_fp16 = squeeze(axes = var_692_axes_0, x = var_691_cast_fp16)[name = tensor<string, []>("op_692_cast_fp16")]; | |
| tensor<int32, [4]> q_37_begin_0 = const()[name = tensor<string, []>("q_37_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; | |
| tensor<int32, [4]> q_37_end_0 = const()[name = tensor<string, []>("q_37_end_0"), val = tensor<int32, [4]>([1, 50, 1, 768])]; | |
| tensor<bool, [4]> q_37_end_mask_0 = const()[name = tensor<string, []>("q_37_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])]; | |
| tensor<bool, [4]> q_37_squeeze_mask_0 = const()[name = tensor<string, []>("q_37_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])]; | |
| tensor<fp16, [50, 1, 768]> q_37_cast_fp16 = slice_by_index(begin = q_37_begin_0, end = q_37_end_0, end_mask = q_37_end_mask_0, squeeze_mask = q_37_squeeze_mask_0, x = var_692_cast_fp16)[name = tensor<string, []>("q_37_cast_fp16")]; | |
| tensor<int32, [4]> k_37_begin_0 = const()[name = tensor<string, []>("k_37_begin_0"), val = tensor<int32, [4]>([1, 0, 0, 0])]; | |
| tensor<int32, [4]> k_37_end_0 = const()[name = tensor<string, []>("k_37_end_0"), val = tensor<int32, [4]>([2, 50, 1, 768])]; | |
| tensor<bool, [4]> k_37_end_mask_0 = const()[name = tensor<string, []>("k_37_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])]; | |
| tensor<bool, [4]> k_37_squeeze_mask_0 = const()[name = tensor<string, []>("k_37_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])]; | |
| tensor<fp16, [50, 1, 768]> k_37_cast_fp16 = slice_by_index(begin = k_37_begin_0, end = k_37_end_0, end_mask = k_37_end_mask_0, squeeze_mask = k_37_squeeze_mask_0, x = var_692_cast_fp16)[name = tensor<string, []>("k_37_cast_fp16")]; | |
| tensor<int32, [4]> v_37_begin_0 = const()[name = tensor<string, []>("v_37_begin_0"), val = tensor<int32, [4]>([2, 0, 0, 0])]; | |
| tensor<int32, [4]> v_37_end_0 = const()[name = tensor<string, []>("v_37_end_0"), val = tensor<int32, [4]>([3, 50, 1, 768])]; | |
| tensor<bool, [4]> v_37_end_mask_0 = const()[name = tensor<string, []>("v_37_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])]; | |
| tensor<bool, [4]> v_37_squeeze_mask_0 = const()[name = tensor<string, []>("v_37_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])]; | |
| tensor<fp16, [50, 1, 768]> v_37_cast_fp16 = slice_by_index(begin = v_37_begin_0, end = v_37_end_0, end_mask = v_37_end_mask_0, squeeze_mask = v_37_squeeze_mask_0, x = var_692_cast_fp16)[name = tensor<string, []>("v_37_cast_fp16")]; | |
| tensor<int32, [3]> var_700 = const()[name = tensor<string, []>("op_700"), val = tensor<int32, [3]>([50, 12, 64])]; | |
| tensor<fp16, [50, 12, 64]> var_701_cast_fp16 = reshape(shape = var_700, x = q_37_cast_fp16)[name = tensor<string, []>("op_701_cast_fp16")]; | |
| tensor<int32, [3]> q_39_perm_0 = const()[name = tensor<string, []>("q_39_perm_0"), val = tensor<int32, [3]>([1, 0, 2])]; | |
| tensor<int32, [3]> var_707 = const()[name = tensor<string, []>("op_707"), val = tensor<int32, [3]>([50, 12, 64])]; | |
| tensor<fp16, [50, 12, 64]> var_708_cast_fp16 = reshape(shape = var_707, x = k_37_cast_fp16)[name = tensor<string, []>("op_708_cast_fp16")]; | |
| tensor<int32, [3]> k_39_perm_0 = const()[name = tensor<string, []>("k_39_perm_0"), val = tensor<int32, [3]>([1, 0, 2])]; | |
| tensor<int32, [3]> var_714 = const()[name = tensor<string, []>("op_714"), val = tensor<int32, [3]>([50, 12, 64])]; | |
| tensor<fp16, [50, 12, 64]> var_715_cast_fp16 = reshape(shape = var_714, x = v_37_cast_fp16)[name = tensor<string, []>("op_715_cast_fp16")]; | |
| tensor<int32, [3]> v_39_perm_0 = const()[name = tensor<string, []>("v_39_perm_0"), val = tensor<int32, [3]>([1, 0, 2])]; | |
| tensor<int32, [4]> var_718 = const()[name = tensor<string, []>("op_718"), val = tensor<int32, [4]>([1, 12, 50, 64])]; | |
| tensor<fp16, [12, 50, 64]> q_39_cast_fp16 = transpose(perm = q_39_perm_0, x = var_701_cast_fp16)[name = tensor<string, []>("transpose_40")]; | |
| tensor<fp16, [1, 12, 50, 64]> q_41_cast_fp16 = reshape(shape = var_718, x = q_39_cast_fp16)[name = tensor<string, []>("q_41_cast_fp16")]; | |
| tensor<int32, [4]> var_720 = const()[name = tensor<string, []>("op_720"), val = tensor<int32, [4]>([1, 12, 50, 64])]; | |
| tensor<fp16, [12, 50, 64]> k_39_cast_fp16 = transpose(perm = k_39_perm_0, x = var_708_cast_fp16)[name = tensor<string, []>("transpose_39")]; | |
| tensor<fp16, [1, 12, 50, 64]> k_41_cast_fp16 = reshape(shape = var_720, x = k_39_cast_fp16)[name = tensor<string, []>("k_41_cast_fp16")]; | |
| tensor<int32, [4]> var_722 = const()[name = tensor<string, []>("op_722"), val = tensor<int32, [4]>([1, 12, 50, 64])]; | |
| tensor<fp16, [12, 50, 64]> v_39_cast_fp16 = transpose(perm = v_39_perm_0, x = var_715_cast_fp16)[name = tensor<string, []>("transpose_38")]; | |
| tensor<fp16, [1, 12, 50, 64]> v_41_cast_fp16 = reshape(shape = var_722, x = v_39_cast_fp16)[name = tensor<string, []>("v_41_cast_fp16")]; | |
| tensor<fp16, []> mul_13_y_0_to_fp16 = const()[name = tensor<string, []>("mul_13_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)]; | |
| tensor<fp16, [1, 12, 50, 64]> mul_13_cast_fp16 = mul(x = q_41_cast_fp16, y = mul_13_y_0_to_fp16)[name = tensor<string, []>("mul_13_cast_fp16")]; | |
| tensor<bool, []> matmul_6_transpose_y_0 = const()[name = tensor<string, []>("matmul_6_transpose_y_0"), val = tensor<bool, []>(true)]; | |
| tensor<bool, []> matmul_6_transpose_x_0 = const()[name = tensor<string, []>("matmul_6_transpose_x_0"), val = tensor<bool, []>(false)]; | |
| tensor<fp16, [1, 12, 50, 50]> matmul_6_cast_fp16 = matmul(transpose_x = matmul_6_transpose_x_0, transpose_y = matmul_6_transpose_y_0, x = mul_13_cast_fp16, y = k_41_cast_fp16)[name = tensor<string, []>("matmul_6_cast_fp16")]; | |
| tensor<int32, []> softmax_6_axis_0 = const()[name = tensor<string, []>("softmax_6_axis_0"), val = tensor<int32, []>(-1)]; | |
| tensor<fp16, [1, 12, 50, 50]> softmax_6_cast_fp16 = softmax(axis = softmax_6_axis_0, x = matmul_6_cast_fp16)[name = tensor<string, []>("softmax_6_cast_fp16")]; | |
| tensor<bool, []> attn_output_49_transpose_x_0 = const()[name = tensor<string, []>("attn_output_49_transpose_x_0"), val = tensor<bool, []>(false)]; | |
| tensor<bool, []> attn_output_49_transpose_y_0 = const()[name = tensor<string, []>("attn_output_49_transpose_y_0"), val = tensor<bool, []>(false)]; | |
| tensor<fp16, [1, 12, 50, 64]> attn_output_49_cast_fp16 = matmul(transpose_x = attn_output_49_transpose_x_0, transpose_y = attn_output_49_transpose_y_0, x = softmax_6_cast_fp16, y = v_41_cast_fp16)[name = tensor<string, []>("attn_output_49_cast_fp16")]; | |
| tensor<int32, [4]> var_725 = const()[name = tensor<string, []>("op_725"), val = tensor<int32, [4]>([2, 0, 1, 3])]; | |
| tensor<int32, [2]> var_730 = const()[name = tensor<string, []>("op_730"), val = tensor<int32, [2]>([50, 768])]; | |
| tensor<fp16, [50, 1, 12, 64]> var_726_cast_fp16 = transpose(perm = var_725, x = attn_output_49_cast_fp16)[name = tensor<string, []>("transpose_37")]; | |
| tensor<fp16, [50, 768]> attn_output_51_cast_fp16 = reshape(shape = var_730, x = var_726_cast_fp16)[name = tensor<string, []>("attn_output_51_cast_fp16")]; | |
| tensor<fp16, [768, 768]> model_visual_transformer_resblocks_6_attn_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("model_visual_transformer_resblocks_6_attn_out_proj_weight_to_fp16"), val = tensor<fp16, [768, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(93406336)))]; | |
| tensor<fp16, [768]> model_visual_transformer_resblocks_6_attn_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_visual_transformer_resblocks_6_attn_out_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(94586048)))]; | |
| tensor<fp16, [50, 768]> linear_25_cast_fp16 = linear(bias = model_visual_transformer_resblocks_6_attn_out_proj_bias_to_fp16, weight = model_visual_transformer_resblocks_6_attn_out_proj_weight_to_fp16, x = attn_output_51_cast_fp16)[name = tensor<string, []>("linear_25_cast_fp16")]; | |
| tensor<int32, [3]> var_734 = const()[name = tensor<string, []>("op_734"), val = tensor<int32, [3]>([50, 1, 768])]; | |
| tensor<fp16, [50, 1, 768]> attn_output_55_cast_fp16 = reshape(shape = var_734, x = linear_25_cast_fp16)[name = tensor<string, []>("attn_output_55_cast_fp16")]; | |
| tensor<int32, [3]> var_736_perm_0 = const()[name = tensor<string, []>("op_736_perm_0"), val = tensor<int32, [3]>([1, 0, 2])]; | |
| tensor<fp16, [1, 50, 768]> var_736_cast_fp16 = transpose(perm = var_736_perm_0, x = attn_output_55_cast_fp16)[name = tensor<string, []>("transpose_36")]; | |
| tensor<fp16, [1, 50, 768]> input_65_cast_fp16 = add(x = input_63_cast_fp16, y = var_736_cast_fp16)[name = tensor<string, []>("input_65_cast_fp16")]; | |
| tensor<int32, [1]> x_37_axes_0 = const()[name = tensor<string, []>("x_37_axes_0"), val = tensor<int32, [1]>([-1])]; | |
| tensor<fp16, [768]> model_visual_transformer_resblocks_6_ln_2_weight_to_fp16 = const()[name = tensor<string, []>("model_visual_transformer_resblocks_6_ln_2_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(94587648)))]; | |
| tensor<fp16, [768]> model_visual_transformer_resblocks_6_ln_2_bias_to_fp16 = const()[name = tensor<string, []>("model_visual_transformer_resblocks_6_ln_2_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(94589248)))]; | |
| tensor<fp16, [1, 50, 768]> x_37_cast_fp16 = layer_norm(axes = x_37_axes_0, beta = model_visual_transformer_resblocks_6_ln_2_bias_to_fp16, epsilon = var_13_to_fp16, gamma = model_visual_transformer_resblocks_6_ln_2_weight_to_fp16, x = input_65_cast_fp16)[name = tensor<string, []>("x_37_cast_fp16")]; | |
| tensor<fp16, [3072, 768]> model_visual_transformer_resblocks_6_mlp_c_fc_weight_to_fp16 = const()[name = tensor<string, []>("model_visual_transformer_resblocks_6_mlp_c_fc_weight_to_fp16"), val = tensor<fp16, [3072, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(94590848)))]; | |
| tensor<fp16, [3072]> model_visual_transformer_resblocks_6_mlp_c_fc_bias_to_fp16 = const()[name = tensor<string, []>("model_visual_transformer_resblocks_6_mlp_c_fc_bias_to_fp16"), val = tensor<fp16, [3072]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(99309504)))]; | |
| tensor<fp16, [1, 50, 3072]> linear_26_cast_fp16 = linear(bias = model_visual_transformer_resblocks_6_mlp_c_fc_bias_to_fp16, weight = model_visual_transformer_resblocks_6_mlp_c_fc_weight_to_fp16, x = x_37_cast_fp16)[name = tensor<string, []>("linear_26_cast_fp16")]; | |
| tensor<string, []> input_71_mode_0 = const()[name = tensor<string, []>("input_71_mode_0"), val = tensor<string, []>("EXACT")]; | |
| tensor<fp16, [1, 50, 3072]> input_71_cast_fp16 = gelu(mode = input_71_mode_0, x = linear_26_cast_fp16)[name = tensor<string, []>("input_71_cast_fp16")]; | |
| tensor<fp16, [768, 3072]> model_visual_transformer_resblocks_6_mlp_c_proj_weight_to_fp16 = const()[name = tensor<string, []>("model_visual_transformer_resblocks_6_mlp_c_proj_weight_to_fp16"), val = tensor<fp16, [768, 3072]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(99315712)))]; | |
| tensor<fp16, [768]> model_visual_transformer_resblocks_6_mlp_c_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_visual_transformer_resblocks_6_mlp_c_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(104034368)))]; | |
| tensor<fp16, [1, 50, 768]> linear_27_cast_fp16 = linear(bias = model_visual_transformer_resblocks_6_mlp_c_proj_bias_to_fp16, weight = model_visual_transformer_resblocks_6_mlp_c_proj_weight_to_fp16, x = input_71_cast_fp16)[name = tensor<string, []>("linear_27_cast_fp16")]; | |
| tensor<fp16, [1, 50, 768]> input_73_cast_fp16 = add(x = input_65_cast_fp16, y = linear_27_cast_fp16)[name = tensor<string, []>("input_73_cast_fp16")]; | |
| tensor<int32, [1]> x_39_axes_0 = const()[name = tensor<string, []>("x_39_axes_0"), val = tensor<int32, [1]>([-1])]; | |
| tensor<fp16, [768]> model_visual_transformer_resblocks_7_ln_1_weight_to_fp16 = const()[name = tensor<string, []>("model_visual_transformer_resblocks_7_ln_1_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(104035968)))]; | |
| tensor<fp16, [768]> model_visual_transformer_resblocks_7_ln_1_bias_to_fp16 = const()[name = tensor<string, []>("model_visual_transformer_resblocks_7_ln_1_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(104037568)))]; | |
| tensor<fp16, [1, 50, 768]> x_39_cast_fp16 = layer_norm(axes = x_39_axes_0, beta = model_visual_transformer_resblocks_7_ln_1_bias_to_fp16, epsilon = var_13_to_fp16, gamma = model_visual_transformer_resblocks_7_ln_1_weight_to_fp16, x = input_73_cast_fp16)[name = tensor<string, []>("x_39_cast_fp16")]; | |
| tensor<int32, [3]> query_31_perm_0 = const()[name = tensor<string, []>("query_31_perm_0"), val = tensor<int32, [3]>([1, 0, 2])]; | |
| tensor<fp16, [2304, 768]> model_visual_transformer_resblocks_7_attn_in_proj_weight_to_fp16 = const()[name = tensor<string, []>("model_visual_transformer_resblocks_7_attn_in_proj_weight_to_fp16"), val = tensor<fp16, [2304, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(104039168)))]; | |
| tensor<fp16, [2304]> model_visual_transformer_resblocks_7_attn_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_visual_transformer_resblocks_7_attn_in_proj_bias_to_fp16"), val = tensor<fp16, [2304]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(107578176)))]; | |
| tensor<fp16, [50, 1, 768]> query_31_cast_fp16 = transpose(perm = query_31_perm_0, x = x_39_cast_fp16)[name = tensor<string, []>("transpose_35")]; | |
| tensor<fp16, [50, 1, 2304]> linear_28_cast_fp16 = linear(bias = model_visual_transformer_resblocks_7_attn_in_proj_bias_to_fp16, weight = model_visual_transformer_resblocks_7_attn_in_proj_weight_to_fp16, x = query_31_cast_fp16)[name = tensor<string, []>("linear_28_cast_fp16")]; | |
| tensor<int32, [4]> concat_8 = const()[name = tensor<string, []>("concat_8"), val = tensor<int32, [4]>([50, 1, 3, 768])]; | |
| tensor<fp16, [50, 1, 3, 768]> var_785_cast_fp16 = reshape(shape = concat_8, x = linear_28_cast_fp16)[name = tensor<string, []>("op_785_cast_fp16")]; | |
| tensor<int32, [1]> var_786_axes_0 = const()[name = tensor<string, []>("op_786_axes_0"), val = tensor<int32, [1]>([0])]; | |
| tensor<fp16, [1, 50, 1, 3, 768]> var_786_cast_fp16 = expand_dims(axes = var_786_axes_0, x = var_785_cast_fp16)[name = tensor<string, []>("op_786_cast_fp16")]; | |
| tensor<int32, [5]> var_787_perm_0 = const()[name = tensor<string, []>("op_787_perm_0"), val = tensor<int32, [5]>([-2, 1, 2, 0, 4])]; | |
| tensor<int32, [1]> var_788_axes_0 = const()[name = tensor<string, []>("op_788_axes_0"), val = tensor<int32, [1]>([-2])]; | |
| tensor<fp16, [3, 50, 1, 1, 768]> var_787_cast_fp16 = transpose(perm = var_787_perm_0, x = var_786_cast_fp16)[name = tensor<string, []>("transpose_34")]; | |
| tensor<fp16, [3, 50, 1, 768]> var_788_cast_fp16 = squeeze(axes = var_788_axes_0, x = var_787_cast_fp16)[name = tensor<string, []>("op_788_cast_fp16")]; | |
| tensor<int32, [4]> q_43_begin_0 = const()[name = tensor<string, []>("q_43_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; | |
| tensor<int32, [4]> q_43_end_0 = const()[name = tensor<string, []>("q_43_end_0"), val = tensor<int32, [4]>([1, 50, 1, 768])]; | |
| tensor<bool, [4]> q_43_end_mask_0 = const()[name = tensor<string, []>("q_43_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])]; | |
| tensor<bool, [4]> q_43_squeeze_mask_0 = const()[name = tensor<string, []>("q_43_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])]; | |
| tensor<fp16, [50, 1, 768]> q_43_cast_fp16 = slice_by_index(begin = q_43_begin_0, end = q_43_end_0, end_mask = q_43_end_mask_0, squeeze_mask = q_43_squeeze_mask_0, x = var_788_cast_fp16)[name = tensor<string, []>("q_43_cast_fp16")]; | |
| tensor<int32, [4]> k_43_begin_0 = const()[name = tensor<string, []>("k_43_begin_0"), val = tensor<int32, [4]>([1, 0, 0, 0])]; | |
| tensor<int32, [4]> k_43_end_0 = const()[name = tensor<string, []>("k_43_end_0"), val = tensor<int32, [4]>([2, 50, 1, 768])]; | |
| tensor<bool, [4]> k_43_end_mask_0 = const()[name = tensor<string, []>("k_43_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])]; | |
| tensor<bool, [4]> k_43_squeeze_mask_0 = const()[name = tensor<string, []>("k_43_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])]; | |
| tensor<fp16, [50, 1, 768]> k_43_cast_fp16 = slice_by_index(begin = k_43_begin_0, end = k_43_end_0, end_mask = k_43_end_mask_0, squeeze_mask = k_43_squeeze_mask_0, x = var_788_cast_fp16)[name = tensor<string, []>("k_43_cast_fp16")]; | |
| tensor<int32, [4]> v_43_begin_0 = const()[name = tensor<string, []>("v_43_begin_0"), val = tensor<int32, [4]>([2, 0, 0, 0])]; | |
| tensor<int32, [4]> v_43_end_0 = const()[name = tensor<string, []>("v_43_end_0"), val = tensor<int32, [4]>([3, 50, 1, 768])]; | |
| tensor<bool, [4]> v_43_end_mask_0 = const()[name = tensor<string, []>("v_43_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])]; | |
| tensor<bool, [4]> v_43_squeeze_mask_0 = const()[name = tensor<string, []>("v_43_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])]; | |
| tensor<fp16, [50, 1, 768]> v_43_cast_fp16 = slice_by_index(begin = v_43_begin_0, end = v_43_end_0, end_mask = v_43_end_mask_0, squeeze_mask = v_43_squeeze_mask_0, x = var_788_cast_fp16)[name = tensor<string, []>("v_43_cast_fp16")]; | |
| tensor<int32, [3]> var_796 = const()[name = tensor<string, []>("op_796"), val = tensor<int32, [3]>([50, 12, 64])]; | |
| tensor<fp16, [50, 12, 64]> var_797_cast_fp16 = reshape(shape = var_796, x = q_43_cast_fp16)[name = tensor<string, []>("op_797_cast_fp16")]; | |
| tensor<int32, [3]> q_45_perm_0 = const()[name = tensor<string, []>("q_45_perm_0"), val = tensor<int32, [3]>([1, 0, 2])]; | |
| tensor<int32, [3]> var_803 = const()[name = tensor<string, []>("op_803"), val = tensor<int32, [3]>([50, 12, 64])]; | |
| tensor<fp16, [50, 12, 64]> var_804_cast_fp16 = reshape(shape = var_803, x = k_43_cast_fp16)[name = tensor<string, []>("op_804_cast_fp16")]; | |
| tensor<int32, [3]> k_45_perm_0 = const()[name = tensor<string, []>("k_45_perm_0"), val = tensor<int32, [3]>([1, 0, 2])]; | |
| tensor<int32, [3]> var_810 = const()[name = tensor<string, []>("op_810"), val = tensor<int32, [3]>([50, 12, 64])]; | |
| tensor<fp16, [50, 12, 64]> var_811_cast_fp16 = reshape(shape = var_810, x = v_43_cast_fp16)[name = tensor<string, []>("op_811_cast_fp16")]; | |
| tensor<int32, [3]> v_45_perm_0 = const()[name = tensor<string, []>("v_45_perm_0"), val = tensor<int32, [3]>([1, 0, 2])]; | |
| tensor<int32, [4]> var_814 = const()[name = tensor<string, []>("op_814"), val = tensor<int32, [4]>([1, 12, 50, 64])]; | |
| tensor<fp16, [12, 50, 64]> q_45_cast_fp16 = transpose(perm = q_45_perm_0, x = var_797_cast_fp16)[name = tensor<string, []>("transpose_33")]; | |
| tensor<fp16, [1, 12, 50, 64]> q_47_cast_fp16 = reshape(shape = var_814, x = q_45_cast_fp16)[name = tensor<string, []>("q_47_cast_fp16")]; | |
| tensor<int32, [4]> var_816 = const()[name = tensor<string, []>("op_816"), val = tensor<int32, [4]>([1, 12, 50, 64])]; | |
| tensor<fp16, [12, 50, 64]> k_45_cast_fp16 = transpose(perm = k_45_perm_0, x = var_804_cast_fp16)[name = tensor<string, []>("transpose_32")]; | |
| tensor<fp16, [1, 12, 50, 64]> k_47_cast_fp16 = reshape(shape = var_816, x = k_45_cast_fp16)[name = tensor<string, []>("k_47_cast_fp16")]; | |
| tensor<int32, [4]> var_818 = const()[name = tensor<string, []>("op_818"), val = tensor<int32, [4]>([1, 12, 50, 64])]; | |
| tensor<fp16, [12, 50, 64]> v_45_cast_fp16 = transpose(perm = v_45_perm_0, x = var_811_cast_fp16)[name = tensor<string, []>("transpose_31")]; | |
| tensor<fp16, [1, 12, 50, 64]> v_47_cast_fp16 = reshape(shape = var_818, x = v_45_cast_fp16)[name = tensor<string, []>("v_47_cast_fp16")]; | |
| tensor<fp16, []> mul_15_y_0_to_fp16 = const()[name = tensor<string, []>("mul_15_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)]; | |
| tensor<fp16, [1, 12, 50, 64]> mul_15_cast_fp16 = mul(x = q_47_cast_fp16, y = mul_15_y_0_to_fp16)[name = tensor<string, []>("mul_15_cast_fp16")]; | |
| tensor<bool, []> matmul_7_transpose_y_0 = const()[name = tensor<string, []>("matmul_7_transpose_y_0"), val = tensor<bool, []>(true)]; | |
| tensor<bool, []> matmul_7_transpose_x_0 = const()[name = tensor<string, []>("matmul_7_transpose_x_0"), val = tensor<bool, []>(false)]; | |
| tensor<fp16, [1, 12, 50, 50]> matmul_7_cast_fp16 = matmul(transpose_x = matmul_7_transpose_x_0, transpose_y = matmul_7_transpose_y_0, x = mul_15_cast_fp16, y = k_47_cast_fp16)[name = tensor<string, []>("matmul_7_cast_fp16")]; | |
| tensor<int32, []> softmax_7_axis_0 = const()[name = tensor<string, []>("softmax_7_axis_0"), val = tensor<int32, []>(-1)]; | |
| tensor<fp16, [1, 12, 50, 50]> softmax_7_cast_fp16 = softmax(axis = softmax_7_axis_0, x = matmul_7_cast_fp16)[name = tensor<string, []>("softmax_7_cast_fp16")]; | |
| tensor<bool, []> attn_output_57_transpose_x_0 = const()[name = tensor<string, []>("attn_output_57_transpose_x_0"), val = tensor<bool, []>(false)]; | |
| tensor<bool, []> attn_output_57_transpose_y_0 = const()[name = tensor<string, []>("attn_output_57_transpose_y_0"), val = tensor<bool, []>(false)]; | |
| tensor<fp16, [1, 12, 50, 64]> attn_output_57_cast_fp16 = matmul(transpose_x = attn_output_57_transpose_x_0, transpose_y = attn_output_57_transpose_y_0, x = softmax_7_cast_fp16, y = v_47_cast_fp16)[name = tensor<string, []>("attn_output_57_cast_fp16")]; | |
| tensor<int32, [4]> var_821 = const()[name = tensor<string, []>("op_821"), val = tensor<int32, [4]>([2, 0, 1, 3])]; | |
| tensor<int32, [2]> var_826 = const()[name = tensor<string, []>("op_826"), val = tensor<int32, [2]>([50, 768])]; | |
| tensor<fp16, [50, 1, 12, 64]> var_822_cast_fp16 = transpose(perm = var_821, x = attn_output_57_cast_fp16)[name = tensor<string, []>("transpose_30")]; | |
| tensor<fp16, [50, 768]> attn_output_59_cast_fp16 = reshape(shape = var_826, x = var_822_cast_fp16)[name = tensor<string, []>("attn_output_59_cast_fp16")]; | |
| tensor<fp16, [768, 768]> model_visual_transformer_resblocks_7_attn_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("model_visual_transformer_resblocks_7_attn_out_proj_weight_to_fp16"), val = tensor<fp16, [768, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(107582848)))]; | |
| tensor<fp16, [768]> model_visual_transformer_resblocks_7_attn_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_visual_transformer_resblocks_7_attn_out_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(108762560)))]; | |
| tensor<fp16, [50, 768]> linear_29_cast_fp16 = linear(bias = model_visual_transformer_resblocks_7_attn_out_proj_bias_to_fp16, weight = model_visual_transformer_resblocks_7_attn_out_proj_weight_to_fp16, x = attn_output_59_cast_fp16)[name = tensor<string, []>("linear_29_cast_fp16")]; | |
| tensor<int32, [3]> var_830 = const()[name = tensor<string, []>("op_830"), val = tensor<int32, [3]>([50, 1, 768])]; | |
| tensor<fp16, [50, 1, 768]> attn_output_63_cast_fp16 = reshape(shape = var_830, x = linear_29_cast_fp16)[name = tensor<string, []>("attn_output_63_cast_fp16")]; | |
| tensor<int32, [3]> var_832_perm_0 = const()[name = tensor<string, []>("op_832_perm_0"), val = tensor<int32, [3]>([1, 0, 2])]; | |
| tensor<fp16, [1, 50, 768]> var_832_cast_fp16 = transpose(perm = var_832_perm_0, x = attn_output_63_cast_fp16)[name = tensor<string, []>("transpose_29")]; | |
| tensor<fp16, [1, 50, 768]> input_75_cast_fp16 = add(x = input_73_cast_fp16, y = var_832_cast_fp16)[name = tensor<string, []>("input_75_cast_fp16")]; | |
| tensor<int32, [1]> x_41_axes_0 = const()[name = tensor<string, []>("x_41_axes_0"), val = tensor<int32, [1]>([-1])]; | |
| tensor<fp16, [768]> model_visual_transformer_resblocks_7_ln_2_weight_to_fp16 = const()[name = tensor<string, []>("model_visual_transformer_resblocks_7_ln_2_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(108764160)))]; | |
| tensor<fp16, [768]> model_visual_transformer_resblocks_7_ln_2_bias_to_fp16 = const()[name = tensor<string, []>("model_visual_transformer_resblocks_7_ln_2_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(108765760)))]; | |
| tensor<fp16, [1, 50, 768]> x_41_cast_fp16 = layer_norm(axes = x_41_axes_0, beta = model_visual_transformer_resblocks_7_ln_2_bias_to_fp16, epsilon = var_13_to_fp16, gamma = model_visual_transformer_resblocks_7_ln_2_weight_to_fp16, x = input_75_cast_fp16)[name = tensor<string, []>("x_41_cast_fp16")]; | |
| tensor<fp16, [3072, 768]> model_visual_transformer_resblocks_7_mlp_c_fc_weight_to_fp16 = const()[name = tensor<string, []>("model_visual_transformer_resblocks_7_mlp_c_fc_weight_to_fp16"), val = tensor<fp16, [3072, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(108767360)))]; | |
| tensor<fp16, [3072]> model_visual_transformer_resblocks_7_mlp_c_fc_bias_to_fp16 = const()[name = tensor<string, []>("model_visual_transformer_resblocks_7_mlp_c_fc_bias_to_fp16"), val = tensor<fp16, [3072]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(113486016)))]; | |
| tensor<fp16, [1, 50, 3072]> linear_30_cast_fp16 = linear(bias = model_visual_transformer_resblocks_7_mlp_c_fc_bias_to_fp16, weight = model_visual_transformer_resblocks_7_mlp_c_fc_weight_to_fp16, x = x_41_cast_fp16)[name = tensor<string, []>("linear_30_cast_fp16")]; | |
| tensor<string, []> input_81_mode_0 = const()[name = tensor<string, []>("input_81_mode_0"), val = tensor<string, []>("EXACT")]; | |
| tensor<fp16, [1, 50, 3072]> input_81_cast_fp16 = gelu(mode = input_81_mode_0, x = linear_30_cast_fp16)[name = tensor<string, []>("input_81_cast_fp16")]; | |
| tensor<fp16, [768, 3072]> model_visual_transformer_resblocks_7_mlp_c_proj_weight_to_fp16 = const()[name = tensor<string, []>("model_visual_transformer_resblocks_7_mlp_c_proj_weight_to_fp16"), val = tensor<fp16, [768, 3072]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(113492224)))]; | |
| tensor<fp16, [768]> model_visual_transformer_resblocks_7_mlp_c_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_visual_transformer_resblocks_7_mlp_c_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(118210880)))]; | |
| tensor<fp16, [1, 50, 768]> linear_31_cast_fp16 = linear(bias = model_visual_transformer_resblocks_7_mlp_c_proj_bias_to_fp16, weight = model_visual_transformer_resblocks_7_mlp_c_proj_weight_to_fp16, x = input_81_cast_fp16)[name = tensor<string, []>("linear_31_cast_fp16")]; | |
| tensor<fp16, [1, 50, 768]> input_83_cast_fp16 = add(x = input_75_cast_fp16, y = linear_31_cast_fp16)[name = tensor<string, []>("input_83_cast_fp16")]; | |
| tensor<int32, [1]> x_43_axes_0 = const()[name = tensor<string, []>("x_43_axes_0"), val = tensor<int32, [1]>([-1])]; | |
| tensor<fp16, [768]> model_visual_transformer_resblocks_8_ln_1_weight_to_fp16 = const()[name = tensor<string, []>("model_visual_transformer_resblocks_8_ln_1_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(118212480)))]; | |
| tensor<fp16, [768]> model_visual_transformer_resblocks_8_ln_1_bias_to_fp16 = const()[name = tensor<string, []>("model_visual_transformer_resblocks_8_ln_1_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(118214080)))]; | |
| tensor<fp16, [1, 50, 768]> x_43_cast_fp16 = layer_norm(axes = x_43_axes_0, beta = model_visual_transformer_resblocks_8_ln_1_bias_to_fp16, epsilon = var_13_to_fp16, gamma = model_visual_transformer_resblocks_8_ln_1_weight_to_fp16, x = input_83_cast_fp16)[name = tensor<string, []>("x_43_cast_fp16")]; | |
| tensor<int32, [3]> query_35_perm_0 = const()[name = tensor<string, []>("query_35_perm_0"), val = tensor<int32, [3]>([1, 0, 2])]; | |
| tensor<fp16, [2304, 768]> model_visual_transformer_resblocks_8_attn_in_proj_weight_to_fp16 = const()[name = tensor<string, []>("model_visual_transformer_resblocks_8_attn_in_proj_weight_to_fp16"), val = tensor<fp16, [2304, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(118215680)))]; | |
| tensor<fp16, [2304]> model_visual_transformer_resblocks_8_attn_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_visual_transformer_resblocks_8_attn_in_proj_bias_to_fp16"), val = tensor<fp16, [2304]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(121754688)))]; | |
| tensor<fp16, [50, 1, 768]> query_35_cast_fp16 = transpose(perm = query_35_perm_0, x = x_43_cast_fp16)[name = tensor<string, []>("transpose_28")]; | |
| tensor<fp16, [50, 1, 2304]> linear_32_cast_fp16 = linear(bias = model_visual_transformer_resblocks_8_attn_in_proj_bias_to_fp16, weight = model_visual_transformer_resblocks_8_attn_in_proj_weight_to_fp16, x = query_35_cast_fp16)[name = tensor<string, []>("linear_32_cast_fp16")]; | |
| tensor<int32, [4]> concat_9 = const()[name = tensor<string, []>("concat_9"), val = tensor<int32, [4]>([50, 1, 3, 768])]; | |
| tensor<fp16, [50, 1, 3, 768]> var_881_cast_fp16 = reshape(shape = concat_9, x = linear_32_cast_fp16)[name = tensor<string, []>("op_881_cast_fp16")]; | |
| tensor<int32, [1]> var_882_axes_0 = const()[name = tensor<string, []>("op_882_axes_0"), val = tensor<int32, [1]>([0])]; | |
| tensor<fp16, [1, 50, 1, 3, 768]> var_882_cast_fp16 = expand_dims(axes = var_882_axes_0, x = var_881_cast_fp16)[name = tensor<string, []>("op_882_cast_fp16")]; | |
| tensor<int32, [5]> var_883_perm_0 = const()[name = tensor<string, []>("op_883_perm_0"), val = tensor<int32, [5]>([-2, 1, 2, 0, 4])]; | |
| tensor<int32, [1]> var_884_axes_0 = const()[name = tensor<string, []>("op_884_axes_0"), val = tensor<int32, [1]>([-2])]; | |
| tensor<fp16, [3, 50, 1, 1, 768]> var_883_cast_fp16 = transpose(perm = var_883_perm_0, x = var_882_cast_fp16)[name = tensor<string, []>("transpose_27")]; | |
| tensor<fp16, [3, 50, 1, 768]> var_884_cast_fp16 = squeeze(axes = var_884_axes_0, x = var_883_cast_fp16)[name = tensor<string, []>("op_884_cast_fp16")]; | |
| tensor<int32, [4]> q_49_begin_0 = const()[name = tensor<string, []>("q_49_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; | |
| tensor<int32, [4]> q_49_end_0 = const()[name = tensor<string, []>("q_49_end_0"), val = tensor<int32, [4]>([1, 50, 1, 768])]; | |
| tensor<bool, [4]> q_49_end_mask_0 = const()[name = tensor<string, []>("q_49_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])]; | |
| tensor<bool, [4]> q_49_squeeze_mask_0 = const()[name = tensor<string, []>("q_49_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])]; | |
| tensor<fp16, [50, 1, 768]> q_49_cast_fp16 = slice_by_index(begin = q_49_begin_0, end = q_49_end_0, end_mask = q_49_end_mask_0, squeeze_mask = q_49_squeeze_mask_0, x = var_884_cast_fp16)[name = tensor<string, []>("q_49_cast_fp16")]; | |
| tensor<int32, [4]> k_49_begin_0 = const()[name = tensor<string, []>("k_49_begin_0"), val = tensor<int32, [4]>([1, 0, 0, 0])]; | |
| tensor<int32, [4]> k_49_end_0 = const()[name = tensor<string, []>("k_49_end_0"), val = tensor<int32, [4]>([2, 50, 1, 768])]; | |
| tensor<bool, [4]> k_49_end_mask_0 = const()[name = tensor<string, []>("k_49_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])]; | |
| tensor<bool, [4]> k_49_squeeze_mask_0 = const()[name = tensor<string, []>("k_49_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])]; | |
| tensor<fp16, [50, 1, 768]> k_49_cast_fp16 = slice_by_index(begin = k_49_begin_0, end = k_49_end_0, end_mask = k_49_end_mask_0, squeeze_mask = k_49_squeeze_mask_0, x = var_884_cast_fp16)[name = tensor<string, []>("k_49_cast_fp16")]; | |
| tensor<int32, [4]> v_49_begin_0 = const()[name = tensor<string, []>("v_49_begin_0"), val = tensor<int32, [4]>([2, 0, 0, 0])]; | |
| tensor<int32, [4]> v_49_end_0 = const()[name = tensor<string, []>("v_49_end_0"), val = tensor<int32, [4]>([3, 50, 1, 768])]; | |
| tensor<bool, [4]> v_49_end_mask_0 = const()[name = tensor<string, []>("v_49_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])]; | |
| tensor<bool, [4]> v_49_squeeze_mask_0 = const()[name = tensor<string, []>("v_49_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])]; | |
| tensor<fp16, [50, 1, 768]> v_49_cast_fp16 = slice_by_index(begin = v_49_begin_0, end = v_49_end_0, end_mask = v_49_end_mask_0, squeeze_mask = v_49_squeeze_mask_0, x = var_884_cast_fp16)[name = tensor<string, []>("v_49_cast_fp16")]; | |
| tensor<int32, [3]> var_892 = const()[name = tensor<string, []>("op_892"), val = tensor<int32, [3]>([50, 12, 64])]; | |
| tensor<fp16, [50, 12, 64]> var_893_cast_fp16 = reshape(shape = var_892, x = q_49_cast_fp16)[name = tensor<string, []>("op_893_cast_fp16")]; | |
| tensor<int32, [3]> q_51_perm_0 = const()[name = tensor<string, []>("q_51_perm_0"), val = tensor<int32, [3]>([1, 0, 2])]; | |
| tensor<int32, [3]> var_899 = const()[name = tensor<string, []>("op_899"), val = tensor<int32, [3]>([50, 12, 64])]; | |
| tensor<fp16, [50, 12, 64]> var_900_cast_fp16 = reshape(shape = var_899, x = k_49_cast_fp16)[name = tensor<string, []>("op_900_cast_fp16")]; | |
| tensor<int32, [3]> k_51_perm_0 = const()[name = tensor<string, []>("k_51_perm_0"), val = tensor<int32, [3]>([1, 0, 2])]; | |
| tensor<int32, [3]> var_906 = const()[name = tensor<string, []>("op_906"), val = tensor<int32, [3]>([50, 12, 64])]; | |
| tensor<fp16, [50, 12, 64]> var_907_cast_fp16 = reshape(shape = var_906, x = v_49_cast_fp16)[name = tensor<string, []>("op_907_cast_fp16")]; | |
| tensor<int32, [3]> v_51_perm_0 = const()[name = tensor<string, []>("v_51_perm_0"), val = tensor<int32, [3]>([1, 0, 2])]; | |
| tensor<int32, [4]> var_910 = const()[name = tensor<string, []>("op_910"), val = tensor<int32, [4]>([1, 12, 50, 64])]; | |
| tensor<fp16, [12, 50, 64]> q_51_cast_fp16 = transpose(perm = q_51_perm_0, x = var_893_cast_fp16)[name = tensor<string, []>("transpose_26")]; | |
| tensor<fp16, [1, 12, 50, 64]> q_53_cast_fp16 = reshape(shape = var_910, x = q_51_cast_fp16)[name = tensor<string, []>("q_53_cast_fp16")]; | |
| tensor<int32, [4]> var_912 = const()[name = tensor<string, []>("op_912"), val = tensor<int32, [4]>([1, 12, 50, 64])]; | |
| tensor<fp16, [12, 50, 64]> k_51_cast_fp16 = transpose(perm = k_51_perm_0, x = var_900_cast_fp16)[name = tensor<string, []>("transpose_25")]; | |
| tensor<fp16, [1, 12, 50, 64]> k_53_cast_fp16 = reshape(shape = var_912, x = k_51_cast_fp16)[name = tensor<string, []>("k_53_cast_fp16")]; | |
| tensor<int32, [4]> var_914 = const()[name = tensor<string, []>("op_914"), val = tensor<int32, [4]>([1, 12, 50, 64])]; | |
| tensor<fp16, [12, 50, 64]> v_51_cast_fp16 = transpose(perm = v_51_perm_0, x = var_907_cast_fp16)[name = tensor<string, []>("transpose_24")]; | |
| tensor<fp16, [1, 12, 50, 64]> v_53_cast_fp16 = reshape(shape = var_914, x = v_51_cast_fp16)[name = tensor<string, []>("v_53_cast_fp16")]; | |
| tensor<fp16, []> mul_17_y_0_to_fp16 = const()[name = tensor<string, []>("mul_17_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)]; | |
| tensor<fp16, [1, 12, 50, 64]> mul_17_cast_fp16 = mul(x = q_53_cast_fp16, y = mul_17_y_0_to_fp16)[name = tensor<string, []>("mul_17_cast_fp16")]; | |
| tensor<bool, []> matmul_8_transpose_y_0 = const()[name = tensor<string, []>("matmul_8_transpose_y_0"), val = tensor<bool, []>(true)]; | |
| tensor<bool, []> matmul_8_transpose_x_0 = const()[name = tensor<string, []>("matmul_8_transpose_x_0"), val = tensor<bool, []>(false)]; | |
| tensor<fp16, [1, 12, 50, 50]> matmul_8_cast_fp16 = matmul(transpose_x = matmul_8_transpose_x_0, transpose_y = matmul_8_transpose_y_0, x = mul_17_cast_fp16, y = k_53_cast_fp16)[name = tensor<string, []>("matmul_8_cast_fp16")]; | |
| tensor<int32, []> softmax_8_axis_0 = const()[name = tensor<string, []>("softmax_8_axis_0"), val = tensor<int32, []>(-1)]; | |
| tensor<fp16, [1, 12, 50, 50]> softmax_8_cast_fp16 = softmax(axis = softmax_8_axis_0, x = matmul_8_cast_fp16)[name = tensor<string, []>("softmax_8_cast_fp16")]; | |
| tensor<bool, []> attn_output_65_transpose_x_0 = const()[name = tensor<string, []>("attn_output_65_transpose_x_0"), val = tensor<bool, []>(false)]; | |
| tensor<bool, []> attn_output_65_transpose_y_0 = const()[name = tensor<string, []>("attn_output_65_transpose_y_0"), val = tensor<bool, []>(false)]; | |
| tensor<fp16, [1, 12, 50, 64]> attn_output_65_cast_fp16 = matmul(transpose_x = attn_output_65_transpose_x_0, transpose_y = attn_output_65_transpose_y_0, x = softmax_8_cast_fp16, y = v_53_cast_fp16)[name = tensor<string, []>("attn_output_65_cast_fp16")]; | |
| tensor<int32, [4]> var_917 = const()[name = tensor<string, []>("op_917"), val = tensor<int32, [4]>([2, 0, 1, 3])]; | |
| tensor<int32, [2]> var_922 = const()[name = tensor<string, []>("op_922"), val = tensor<int32, [2]>([50, 768])]; | |
| tensor<fp16, [50, 1, 12, 64]> var_918_cast_fp16 = transpose(perm = var_917, x = attn_output_65_cast_fp16)[name = tensor<string, []>("transpose_23")]; | |
| tensor<fp16, [50, 768]> attn_output_67_cast_fp16 = reshape(shape = var_922, x = var_918_cast_fp16)[name = tensor<string, []>("attn_output_67_cast_fp16")]; | |
| tensor<fp16, [768, 768]> model_visual_transformer_resblocks_8_attn_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("model_visual_transformer_resblocks_8_attn_out_proj_weight_to_fp16"), val = tensor<fp16, [768, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(121759360)))]; | |
| tensor<fp16, [768]> model_visual_transformer_resblocks_8_attn_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_visual_transformer_resblocks_8_attn_out_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(122939072)))]; | |
| tensor<fp16, [50, 768]> linear_33_cast_fp16 = linear(bias = model_visual_transformer_resblocks_8_attn_out_proj_bias_to_fp16, weight = model_visual_transformer_resblocks_8_attn_out_proj_weight_to_fp16, x = attn_output_67_cast_fp16)[name = tensor<string, []>("linear_33_cast_fp16")]; | |
| tensor<int32, [3]> var_926 = const()[name = tensor<string, []>("op_926"), val = tensor<int32, [3]>([50, 1, 768])]; | |
| tensor<fp16, [50, 1, 768]> attn_output_71_cast_fp16 = reshape(shape = var_926, x = linear_33_cast_fp16)[name = tensor<string, []>("attn_output_71_cast_fp16")]; | |
| tensor<int32, [3]> var_928_perm_0 = const()[name = tensor<string, []>("op_928_perm_0"), val = tensor<int32, [3]>([1, 0, 2])]; | |
| tensor<fp16, [1, 50, 768]> var_928_cast_fp16 = transpose(perm = var_928_perm_0, x = attn_output_71_cast_fp16)[name = tensor<string, []>("transpose_22")]; | |
| tensor<fp16, [1, 50, 768]> input_85_cast_fp16 = add(x = input_83_cast_fp16, y = var_928_cast_fp16)[name = tensor<string, []>("input_85_cast_fp16")]; | |
| tensor<int32, [1]> x_45_axes_0 = const()[name = tensor<string, []>("x_45_axes_0"), val = tensor<int32, [1]>([-1])]; | |
| tensor<fp16, [768]> model_visual_transformer_resblocks_8_ln_2_weight_to_fp16 = const()[name = tensor<string, []>("model_visual_transformer_resblocks_8_ln_2_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(122940672)))]; | |
| tensor<fp16, [768]> model_visual_transformer_resblocks_8_ln_2_bias_to_fp16 = const()[name = tensor<string, []>("model_visual_transformer_resblocks_8_ln_2_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(122942272)))]; | |
| tensor<fp16, [1, 50, 768]> x_45_cast_fp16 = layer_norm(axes = x_45_axes_0, beta = model_visual_transformer_resblocks_8_ln_2_bias_to_fp16, epsilon = var_13_to_fp16, gamma = model_visual_transformer_resblocks_8_ln_2_weight_to_fp16, x = input_85_cast_fp16)[name = tensor<string, []>("x_45_cast_fp16")]; | |
| tensor<fp16, [3072, 768]> model_visual_transformer_resblocks_8_mlp_c_fc_weight_to_fp16 = const()[name = tensor<string, []>("model_visual_transformer_resblocks_8_mlp_c_fc_weight_to_fp16"), val = tensor<fp16, [3072, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(122943872)))]; | |
| tensor<fp16, [3072]> model_visual_transformer_resblocks_8_mlp_c_fc_bias_to_fp16 = const()[name = tensor<string, []>("model_visual_transformer_resblocks_8_mlp_c_fc_bias_to_fp16"), val = tensor<fp16, [3072]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(127662528)))]; | |
| tensor<fp16, [1, 50, 3072]> linear_34_cast_fp16 = linear(bias = model_visual_transformer_resblocks_8_mlp_c_fc_bias_to_fp16, weight = model_visual_transformer_resblocks_8_mlp_c_fc_weight_to_fp16, x = x_45_cast_fp16)[name = tensor<string, []>("linear_34_cast_fp16")]; | |
| tensor<string, []> input_91_mode_0 = const()[name = tensor<string, []>("input_91_mode_0"), val = tensor<string, []>("EXACT")]; | |
| tensor<fp16, [1, 50, 3072]> input_91_cast_fp16 = gelu(mode = input_91_mode_0, x = linear_34_cast_fp16)[name = tensor<string, []>("input_91_cast_fp16")]; | |
| tensor<fp16, [768, 3072]> model_visual_transformer_resblocks_8_mlp_c_proj_weight_to_fp16 = const()[name = tensor<string, []>("model_visual_transformer_resblocks_8_mlp_c_proj_weight_to_fp16"), val = tensor<fp16, [768, 3072]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(127668736)))]; | |
| tensor<fp16, [768]> model_visual_transformer_resblocks_8_mlp_c_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_visual_transformer_resblocks_8_mlp_c_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(132387392)))]; | |
| tensor<fp16, [1, 50, 768]> linear_35_cast_fp16 = linear(bias = model_visual_transformer_resblocks_8_mlp_c_proj_bias_to_fp16, weight = model_visual_transformer_resblocks_8_mlp_c_proj_weight_to_fp16, x = input_91_cast_fp16)[name = tensor<string, []>("linear_35_cast_fp16")]; | |
| tensor<fp16, [1, 50, 768]> input_93_cast_fp16 = add(x = input_85_cast_fp16, y = linear_35_cast_fp16)[name = tensor<string, []>("input_93_cast_fp16")]; | |
| tensor<int32, [1]> x_47_axes_0 = const()[name = tensor<string, []>("x_47_axes_0"), val = tensor<int32, [1]>([-1])]; | |
| tensor<fp16, [768]> model_visual_transformer_resblocks_9_ln_1_weight_to_fp16 = const()[name = tensor<string, []>("model_visual_transformer_resblocks_9_ln_1_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(132388992)))]; | |
| tensor<fp16, [768]> model_visual_transformer_resblocks_9_ln_1_bias_to_fp16 = const()[name = tensor<string, []>("model_visual_transformer_resblocks_9_ln_1_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(132390592)))]; | |
| tensor<fp16, [1, 50, 768]> x_47_cast_fp16 = layer_norm(axes = x_47_axes_0, beta = model_visual_transformer_resblocks_9_ln_1_bias_to_fp16, epsilon = var_13_to_fp16, gamma = model_visual_transformer_resblocks_9_ln_1_weight_to_fp16, x = input_93_cast_fp16)[name = tensor<string, []>("x_47_cast_fp16")]; | |
| tensor<int32, [3]> query_39_perm_0 = const()[name = tensor<string, []>("query_39_perm_0"), val = tensor<int32, [3]>([1, 0, 2])]; | |
| tensor<fp16, [2304, 768]> model_visual_transformer_resblocks_9_attn_in_proj_weight_to_fp16 = const()[name = tensor<string, []>("model_visual_transformer_resblocks_9_attn_in_proj_weight_to_fp16"), val = tensor<fp16, [2304, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(132392192)))]; | |
| tensor<fp16, [2304]> model_visual_transformer_resblocks_9_attn_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_visual_transformer_resblocks_9_attn_in_proj_bias_to_fp16"), val = tensor<fp16, [2304]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(135931200)))]; | |
| tensor<fp16, [50, 1, 768]> query_39_cast_fp16 = transpose(perm = query_39_perm_0, x = x_47_cast_fp16)[name = tensor<string, []>("transpose_21")]; | |
| tensor<fp16, [50, 1, 2304]> linear_36_cast_fp16 = linear(bias = model_visual_transformer_resblocks_9_attn_in_proj_bias_to_fp16, weight = model_visual_transformer_resblocks_9_attn_in_proj_weight_to_fp16, x = query_39_cast_fp16)[name = tensor<string, []>("linear_36_cast_fp16")]; | |
| tensor<int32, [4]> concat_10 = const()[name = tensor<string, []>("concat_10"), val = tensor<int32, [4]>([50, 1, 3, 768])]; | |
| tensor<fp16, [50, 1, 3, 768]> var_977_cast_fp16 = reshape(shape = concat_10, x = linear_36_cast_fp16)[name = tensor<string, []>("op_977_cast_fp16")]; | |
| tensor<int32, [1]> var_978_axes_0 = const()[name = tensor<string, []>("op_978_axes_0"), val = tensor<int32, [1]>([0])]; | |
| tensor<fp16, [1, 50, 1, 3, 768]> var_978_cast_fp16 = expand_dims(axes = var_978_axes_0, x = var_977_cast_fp16)[name = tensor<string, []>("op_978_cast_fp16")]; | |
| tensor<int32, [5]> var_979_perm_0 = const()[name = tensor<string, []>("op_979_perm_0"), val = tensor<int32, [5]>([-2, 1, 2, 0, 4])]; | |
| tensor<int32, [1]> var_980_axes_0 = const()[name = tensor<string, []>("op_980_axes_0"), val = tensor<int32, [1]>([-2])]; | |
| tensor<fp16, [3, 50, 1, 1, 768]> var_979_cast_fp16 = transpose(perm = var_979_perm_0, x = var_978_cast_fp16)[name = tensor<string, []>("transpose_20")]; | |
| tensor<fp16, [3, 50, 1, 768]> var_980_cast_fp16 = squeeze(axes = var_980_axes_0, x = var_979_cast_fp16)[name = tensor<string, []>("op_980_cast_fp16")]; | |
| tensor<int32, [4]> q_55_begin_0 = const()[name = tensor<string, []>("q_55_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; | |
| tensor<int32, [4]> q_55_end_0 = const()[name = tensor<string, []>("q_55_end_0"), val = tensor<int32, [4]>([1, 50, 1, 768])]; | |
| tensor<bool, [4]> q_55_end_mask_0 = const()[name = tensor<string, []>("q_55_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])]; | |
| tensor<bool, [4]> q_55_squeeze_mask_0 = const()[name = tensor<string, []>("q_55_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])]; | |
| tensor<fp16, [50, 1, 768]> q_55_cast_fp16 = slice_by_index(begin = q_55_begin_0, end = q_55_end_0, end_mask = q_55_end_mask_0, squeeze_mask = q_55_squeeze_mask_0, x = var_980_cast_fp16)[name = tensor<string, []>("q_55_cast_fp16")]; | |
| tensor<int32, [4]> k_55_begin_0 = const()[name = tensor<string, []>("k_55_begin_0"), val = tensor<int32, [4]>([1, 0, 0, 0])]; | |
| tensor<int32, [4]> k_55_end_0 = const()[name = tensor<string, []>("k_55_end_0"), val = tensor<int32, [4]>([2, 50, 1, 768])]; | |
| tensor<bool, [4]> k_55_end_mask_0 = const()[name = tensor<string, []>("k_55_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])]; | |
| tensor<bool, [4]> k_55_squeeze_mask_0 = const()[name = tensor<string, []>("k_55_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])]; | |
| tensor<fp16, [50, 1, 768]> k_55_cast_fp16 = slice_by_index(begin = k_55_begin_0, end = k_55_end_0, end_mask = k_55_end_mask_0, squeeze_mask = k_55_squeeze_mask_0, x = var_980_cast_fp16)[name = tensor<string, []>("k_55_cast_fp16")]; | |
| tensor<int32, [4]> v_55_begin_0 = const()[name = tensor<string, []>("v_55_begin_0"), val = tensor<int32, [4]>([2, 0, 0, 0])]; | |
| tensor<int32, [4]> v_55_end_0 = const()[name = tensor<string, []>("v_55_end_0"), val = tensor<int32, [4]>([3, 50, 1, 768])]; | |
| tensor<bool, [4]> v_55_end_mask_0 = const()[name = tensor<string, []>("v_55_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])]; | |
| tensor<bool, [4]> v_55_squeeze_mask_0 = const()[name = tensor<string, []>("v_55_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])]; | |
| tensor<fp16, [50, 1, 768]> v_55_cast_fp16 = slice_by_index(begin = v_55_begin_0, end = v_55_end_0, end_mask = v_55_end_mask_0, squeeze_mask = v_55_squeeze_mask_0, x = var_980_cast_fp16)[name = tensor<string, []>("v_55_cast_fp16")]; | |
| tensor<int32, [3]> var_988 = const()[name = tensor<string, []>("op_988"), val = tensor<int32, [3]>([50, 12, 64])]; | |
| tensor<fp16, [50, 12, 64]> var_989_cast_fp16 = reshape(shape = var_988, x = q_55_cast_fp16)[name = tensor<string, []>("op_989_cast_fp16")]; | |
| tensor<int32, [3]> q_57_perm_0 = const()[name = tensor<string, []>("q_57_perm_0"), val = tensor<int32, [3]>([1, 0, 2])]; | |
| tensor<int32, [3]> var_995 = const()[name = tensor<string, []>("op_995"), val = tensor<int32, [3]>([50, 12, 64])]; | |
| tensor<fp16, [50, 12, 64]> var_996_cast_fp16 = reshape(shape = var_995, x = k_55_cast_fp16)[name = tensor<string, []>("op_996_cast_fp16")]; | |
| tensor<int32, [3]> k_57_perm_0 = const()[name = tensor<string, []>("k_57_perm_0"), val = tensor<int32, [3]>([1, 0, 2])]; | |
| tensor<int32, [3]> var_1002 = const()[name = tensor<string, []>("op_1002"), val = tensor<int32, [3]>([50, 12, 64])]; | |
| tensor<fp16, [50, 12, 64]> var_1003_cast_fp16 = reshape(shape = var_1002, x = v_55_cast_fp16)[name = tensor<string, []>("op_1003_cast_fp16")]; | |
| tensor<int32, [3]> v_57_perm_0 = const()[name = tensor<string, []>("v_57_perm_0"), val = tensor<int32, [3]>([1, 0, 2])]; | |
| tensor<int32, [4]> var_1006 = const()[name = tensor<string, []>("op_1006"), val = tensor<int32, [4]>([1, 12, 50, 64])]; | |
| tensor<fp16, [12, 50, 64]> q_57_cast_fp16 = transpose(perm = q_57_perm_0, x = var_989_cast_fp16)[name = tensor<string, []>("transpose_19")]; | |
| tensor<fp16, [1, 12, 50, 64]> q_59_cast_fp16 = reshape(shape = var_1006, x = q_57_cast_fp16)[name = tensor<string, []>("q_59_cast_fp16")]; | |
| tensor<int32, [4]> var_1008 = const()[name = tensor<string, []>("op_1008"), val = tensor<int32, [4]>([1, 12, 50, 64])]; | |
| tensor<fp16, [12, 50, 64]> k_57_cast_fp16 = transpose(perm = k_57_perm_0, x = var_996_cast_fp16)[name = tensor<string, []>("transpose_18")]; | |
| tensor<fp16, [1, 12, 50, 64]> k_59_cast_fp16 = reshape(shape = var_1008, x = k_57_cast_fp16)[name = tensor<string, []>("k_59_cast_fp16")]; | |
| tensor<int32, [4]> var_1010 = const()[name = tensor<string, []>("op_1010"), val = tensor<int32, [4]>([1, 12, 50, 64])]; | |
| tensor<fp16, [12, 50, 64]> v_57_cast_fp16 = transpose(perm = v_57_perm_0, x = var_1003_cast_fp16)[name = tensor<string, []>("transpose_17")]; | |
| tensor<fp16, [1, 12, 50, 64]> v_59_cast_fp16 = reshape(shape = var_1010, x = v_57_cast_fp16)[name = tensor<string, []>("v_59_cast_fp16")]; | |
| tensor<fp16, []> mul_19_y_0_to_fp16 = const()[name = tensor<string, []>("mul_19_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)]; | |
| tensor<fp16, [1, 12, 50, 64]> mul_19_cast_fp16 = mul(x = q_59_cast_fp16, y = mul_19_y_0_to_fp16)[name = tensor<string, []>("mul_19_cast_fp16")]; | |
| tensor<bool, []> matmul_9_transpose_y_0 = const()[name = tensor<string, []>("matmul_9_transpose_y_0"), val = tensor<bool, []>(true)]; | |
| tensor<bool, []> matmul_9_transpose_x_0 = const()[name = tensor<string, []>("matmul_9_transpose_x_0"), val = tensor<bool, []>(false)]; | |
| tensor<fp16, [1, 12, 50, 50]> matmul_9_cast_fp16 = matmul(transpose_x = matmul_9_transpose_x_0, transpose_y = matmul_9_transpose_y_0, x = mul_19_cast_fp16, y = k_59_cast_fp16)[name = tensor<string, []>("matmul_9_cast_fp16")]; | |
| tensor<int32, []> softmax_9_axis_0 = const()[name = tensor<string, []>("softmax_9_axis_0"), val = tensor<int32, []>(-1)]; | |
| tensor<fp16, [1, 12, 50, 50]> softmax_9_cast_fp16 = softmax(axis = softmax_9_axis_0, x = matmul_9_cast_fp16)[name = tensor<string, []>("softmax_9_cast_fp16")]; | |
| tensor<bool, []> attn_output_73_transpose_x_0 = const()[name = tensor<string, []>("attn_output_73_transpose_x_0"), val = tensor<bool, []>(false)]; | |
| tensor<bool, []> attn_output_73_transpose_y_0 = const()[name = tensor<string, []>("attn_output_73_transpose_y_0"), val = tensor<bool, []>(false)]; | |
| tensor<fp16, [1, 12, 50, 64]> attn_output_73_cast_fp16 = matmul(transpose_x = attn_output_73_transpose_x_0, transpose_y = attn_output_73_transpose_y_0, x = softmax_9_cast_fp16, y = v_59_cast_fp16)[name = tensor<string, []>("attn_output_73_cast_fp16")]; | |
| tensor<int32, [4]> var_1013 = const()[name = tensor<string, []>("op_1013"), val = tensor<int32, [4]>([2, 0, 1, 3])]; | |
| tensor<int32, [2]> var_1018 = const()[name = tensor<string, []>("op_1018"), val = tensor<int32, [2]>([50, 768])]; | |
| tensor<fp16, [50, 1, 12, 64]> var_1014_cast_fp16 = transpose(perm = var_1013, x = attn_output_73_cast_fp16)[name = tensor<string, []>("transpose_16")]; | |
| tensor<fp16, [50, 768]> attn_output_75_cast_fp16 = reshape(shape = var_1018, x = var_1014_cast_fp16)[name = tensor<string, []>("attn_output_75_cast_fp16")]; | |
| tensor<fp16, [768, 768]> model_visual_transformer_resblocks_9_attn_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("model_visual_transformer_resblocks_9_attn_out_proj_weight_to_fp16"), val = tensor<fp16, [768, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(135935872)))]; | |
| tensor<fp16, [768]> model_visual_transformer_resblocks_9_attn_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_visual_transformer_resblocks_9_attn_out_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(137115584)))]; | |
| tensor<fp16, [50, 768]> linear_37_cast_fp16 = linear(bias = model_visual_transformer_resblocks_9_attn_out_proj_bias_to_fp16, weight = model_visual_transformer_resblocks_9_attn_out_proj_weight_to_fp16, x = attn_output_75_cast_fp16)[name = tensor<string, []>("linear_37_cast_fp16")]; | |
| tensor<int32, [3]> var_1022 = const()[name = tensor<string, []>("op_1022"), val = tensor<int32, [3]>([50, 1, 768])]; | |
| tensor<fp16, [50, 1, 768]> attn_output_79_cast_fp16 = reshape(shape = var_1022, x = linear_37_cast_fp16)[name = tensor<string, []>("attn_output_79_cast_fp16")]; | |
| tensor<int32, [3]> var_1024_perm_0 = const()[name = tensor<string, []>("op_1024_perm_0"), val = tensor<int32, [3]>([1, 0, 2])]; | |
| tensor<fp16, [1, 50, 768]> var_1024_cast_fp16 = transpose(perm = var_1024_perm_0, x = attn_output_79_cast_fp16)[name = tensor<string, []>("transpose_15")]; | |
| tensor<fp16, [1, 50, 768]> input_95_cast_fp16 = add(x = input_93_cast_fp16, y = var_1024_cast_fp16)[name = tensor<string, []>("input_95_cast_fp16")]; | |
| tensor<int32, [1]> x_49_axes_0 = const()[name = tensor<string, []>("x_49_axes_0"), val = tensor<int32, [1]>([-1])]; | |
| tensor<fp16, [768]> model_visual_transformer_resblocks_9_ln_2_weight_to_fp16 = const()[name = tensor<string, []>("model_visual_transformer_resblocks_9_ln_2_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(137117184)))]; | |
| tensor<fp16, [768]> model_visual_transformer_resblocks_9_ln_2_bias_to_fp16 = const()[name = tensor<string, []>("model_visual_transformer_resblocks_9_ln_2_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(137118784)))]; | |
| tensor<fp16, [1, 50, 768]> x_49_cast_fp16 = layer_norm(axes = x_49_axes_0, beta = model_visual_transformer_resblocks_9_ln_2_bias_to_fp16, epsilon = var_13_to_fp16, gamma = model_visual_transformer_resblocks_9_ln_2_weight_to_fp16, x = input_95_cast_fp16)[name = tensor<string, []>("x_49_cast_fp16")]; | |
| tensor<fp16, [3072, 768]> model_visual_transformer_resblocks_9_mlp_c_fc_weight_to_fp16 = const()[name = tensor<string, []>("model_visual_transformer_resblocks_9_mlp_c_fc_weight_to_fp16"), val = tensor<fp16, [3072, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(137120384)))]; | |
| tensor<fp16, [3072]> model_visual_transformer_resblocks_9_mlp_c_fc_bias_to_fp16 = const()[name = tensor<string, []>("model_visual_transformer_resblocks_9_mlp_c_fc_bias_to_fp16"), val = tensor<fp16, [3072]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(141839040)))]; | |
| tensor<fp16, [1, 50, 3072]> linear_38_cast_fp16 = linear(bias = model_visual_transformer_resblocks_9_mlp_c_fc_bias_to_fp16, weight = model_visual_transformer_resblocks_9_mlp_c_fc_weight_to_fp16, x = x_49_cast_fp16)[name = tensor<string, []>("linear_38_cast_fp16")]; | |
| tensor<string, []> input_101_mode_0 = const()[name = tensor<string, []>("input_101_mode_0"), val = tensor<string, []>("EXACT")]; | |
| tensor<fp16, [1, 50, 3072]> input_101_cast_fp16 = gelu(mode = input_101_mode_0, x = linear_38_cast_fp16)[name = tensor<string, []>("input_101_cast_fp16")]; | |
| tensor<fp16, [768, 3072]> model_visual_transformer_resblocks_9_mlp_c_proj_weight_to_fp16 = const()[name = tensor<string, []>("model_visual_transformer_resblocks_9_mlp_c_proj_weight_to_fp16"), val = tensor<fp16, [768, 3072]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(141845248)))]; | |
| tensor<fp16, [768]> model_visual_transformer_resblocks_9_mlp_c_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_visual_transformer_resblocks_9_mlp_c_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(146563904)))]; | |
| tensor<fp16, [1, 50, 768]> linear_39_cast_fp16 = linear(bias = model_visual_transformer_resblocks_9_mlp_c_proj_bias_to_fp16, weight = model_visual_transformer_resblocks_9_mlp_c_proj_weight_to_fp16, x = input_101_cast_fp16)[name = tensor<string, []>("linear_39_cast_fp16")]; | |
| tensor<fp16, [1, 50, 768]> input_103_cast_fp16 = add(x = input_95_cast_fp16, y = linear_39_cast_fp16)[name = tensor<string, []>("input_103_cast_fp16")]; | |
| tensor<int32, [1]> x_51_axes_0 = const()[name = tensor<string, []>("x_51_axes_0"), val = tensor<int32, [1]>([-1])]; | |
| tensor<fp16, [768]> model_visual_transformer_resblocks_10_ln_1_weight_to_fp16 = const()[name = tensor<string, []>("model_visual_transformer_resblocks_10_ln_1_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(146565504)))]; | |
| tensor<fp16, [768]> model_visual_transformer_resblocks_10_ln_1_bias_to_fp16 = const()[name = tensor<string, []>("model_visual_transformer_resblocks_10_ln_1_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(146567104)))]; | |
| tensor<fp16, [1, 50, 768]> x_51_cast_fp16 = layer_norm(axes = x_51_axes_0, beta = model_visual_transformer_resblocks_10_ln_1_bias_to_fp16, epsilon = var_13_to_fp16, gamma = model_visual_transformer_resblocks_10_ln_1_weight_to_fp16, x = input_103_cast_fp16)[name = tensor<string, []>("x_51_cast_fp16")]; | |
| tensor<int32, [3]> query_43_perm_0 = const()[name = tensor<string, []>("query_43_perm_0"), val = tensor<int32, [3]>([1, 0, 2])]; | |
| tensor<fp16, [2304, 768]> model_visual_transformer_resblocks_10_attn_in_proj_weight_to_fp16 = const()[name = tensor<string, []>("model_visual_transformer_resblocks_10_attn_in_proj_weight_to_fp16"), val = tensor<fp16, [2304, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(146568704)))]; | |
| tensor<fp16, [2304]> model_visual_transformer_resblocks_10_attn_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_visual_transformer_resblocks_10_attn_in_proj_bias_to_fp16"), val = tensor<fp16, [2304]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(150107712)))]; | |
| tensor<fp16, [50, 1, 768]> query_43_cast_fp16 = transpose(perm = query_43_perm_0, x = x_51_cast_fp16)[name = tensor<string, []>("transpose_14")]; | |
| tensor<fp16, [50, 1, 2304]> linear_40_cast_fp16 = linear(bias = model_visual_transformer_resblocks_10_attn_in_proj_bias_to_fp16, weight = model_visual_transformer_resblocks_10_attn_in_proj_weight_to_fp16, x = query_43_cast_fp16)[name = tensor<string, []>("linear_40_cast_fp16")]; | |
| tensor<int32, [4]> concat_11 = const()[name = tensor<string, []>("concat_11"), val = tensor<int32, [4]>([50, 1, 3, 768])]; | |
| tensor<fp16, [50, 1, 3, 768]> var_1073_cast_fp16 = reshape(shape = concat_11, x = linear_40_cast_fp16)[name = tensor<string, []>("op_1073_cast_fp16")]; | |
| tensor<int32, [1]> var_1074_axes_0 = const()[name = tensor<string, []>("op_1074_axes_0"), val = tensor<int32, [1]>([0])]; | |
| tensor<fp16, [1, 50, 1, 3, 768]> var_1074_cast_fp16 = expand_dims(axes = var_1074_axes_0, x = var_1073_cast_fp16)[name = tensor<string, []>("op_1074_cast_fp16")]; | |
| tensor<int32, [5]> var_1075_perm_0 = const()[name = tensor<string, []>("op_1075_perm_0"), val = tensor<int32, [5]>([-2, 1, 2, 0, 4])]; | |
| tensor<int32, [1]> var_1076_axes_0 = const()[name = tensor<string, []>("op_1076_axes_0"), val = tensor<int32, [1]>([-2])]; | |
| tensor<fp16, [3, 50, 1, 1, 768]> var_1075_cast_fp16 = transpose(perm = var_1075_perm_0, x = var_1074_cast_fp16)[name = tensor<string, []>("transpose_13")]; | |
| tensor<fp16, [3, 50, 1, 768]> var_1076_cast_fp16 = squeeze(axes = var_1076_axes_0, x = var_1075_cast_fp16)[name = tensor<string, []>("op_1076_cast_fp16")]; | |
| tensor<int32, [4]> q_61_begin_0 = const()[name = tensor<string, []>("q_61_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; | |
| tensor<int32, [4]> q_61_end_0 = const()[name = tensor<string, []>("q_61_end_0"), val = tensor<int32, [4]>([1, 50, 1, 768])]; | |
| tensor<bool, [4]> q_61_end_mask_0 = const()[name = tensor<string, []>("q_61_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])]; | |
| tensor<bool, [4]> q_61_squeeze_mask_0 = const()[name = tensor<string, []>("q_61_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])]; | |
| tensor<fp16, [50, 1, 768]> q_61_cast_fp16 = slice_by_index(begin = q_61_begin_0, end = q_61_end_0, end_mask = q_61_end_mask_0, squeeze_mask = q_61_squeeze_mask_0, x = var_1076_cast_fp16)[name = tensor<string, []>("q_61_cast_fp16")]; | |
| tensor<int32, [4]> k_61_begin_0 = const()[name = tensor<string, []>("k_61_begin_0"), val = tensor<int32, [4]>([1, 0, 0, 0])]; | |
| tensor<int32, [4]> k_61_end_0 = const()[name = tensor<string, []>("k_61_end_0"), val = tensor<int32, [4]>([2, 50, 1, 768])]; | |
| tensor<bool, [4]> k_61_end_mask_0 = const()[name = tensor<string, []>("k_61_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])]; | |
| tensor<bool, [4]> k_61_squeeze_mask_0 = const()[name = tensor<string, []>("k_61_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])]; | |
| tensor<fp16, [50, 1, 768]> k_61_cast_fp16 = slice_by_index(begin = k_61_begin_0, end = k_61_end_0, end_mask = k_61_end_mask_0, squeeze_mask = k_61_squeeze_mask_0, x = var_1076_cast_fp16)[name = tensor<string, []>("k_61_cast_fp16")]; | |
| tensor<int32, [4]> v_61_begin_0 = const()[name = tensor<string, []>("v_61_begin_0"), val = tensor<int32, [4]>([2, 0, 0, 0])]; | |
| tensor<int32, [4]> v_61_end_0 = const()[name = tensor<string, []>("v_61_end_0"), val = tensor<int32, [4]>([3, 50, 1, 768])]; | |
| tensor<bool, [4]> v_61_end_mask_0 = const()[name = tensor<string, []>("v_61_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])]; | |
| tensor<bool, [4]> v_61_squeeze_mask_0 = const()[name = tensor<string, []>("v_61_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])]; | |
| tensor<fp16, [50, 1, 768]> v_61_cast_fp16 = slice_by_index(begin = v_61_begin_0, end = v_61_end_0, end_mask = v_61_end_mask_0, squeeze_mask = v_61_squeeze_mask_0, x = var_1076_cast_fp16)[name = tensor<string, []>("v_61_cast_fp16")]; | |
| tensor<int32, [3]> var_1084 = const()[name = tensor<string, []>("op_1084"), val = tensor<int32, [3]>([50, 12, 64])]; | |
| tensor<fp16, [50, 12, 64]> var_1085_cast_fp16 = reshape(shape = var_1084, x = q_61_cast_fp16)[name = tensor<string, []>("op_1085_cast_fp16")]; | |
| tensor<int32, [3]> q_63_perm_0 = const()[name = tensor<string, []>("q_63_perm_0"), val = tensor<int32, [3]>([1, 0, 2])]; | |
| tensor<int32, [3]> var_1091 = const()[name = tensor<string, []>("op_1091"), val = tensor<int32, [3]>([50, 12, 64])]; | |
| tensor<fp16, [50, 12, 64]> var_1092_cast_fp16 = reshape(shape = var_1091, x = k_61_cast_fp16)[name = tensor<string, []>("op_1092_cast_fp16")]; | |
| tensor<int32, [3]> k_63_perm_0 = const()[name = tensor<string, []>("k_63_perm_0"), val = tensor<int32, [3]>([1, 0, 2])]; | |
| tensor<int32, [3]> var_1098 = const()[name = tensor<string, []>("op_1098"), val = tensor<int32, [3]>([50, 12, 64])]; | |
| tensor<fp16, [50, 12, 64]> var_1099_cast_fp16 = reshape(shape = var_1098, x = v_61_cast_fp16)[name = tensor<string, []>("op_1099_cast_fp16")]; | |
| tensor<int32, [3]> v_63_perm_0 = const()[name = tensor<string, []>("v_63_perm_0"), val = tensor<int32, [3]>([1, 0, 2])]; | |
| tensor<int32, [4]> var_1102 = const()[name = tensor<string, []>("op_1102"), val = tensor<int32, [4]>([1, 12, 50, 64])]; | |
| tensor<fp16, [12, 50, 64]> q_63_cast_fp16 = transpose(perm = q_63_perm_0, x = var_1085_cast_fp16)[name = tensor<string, []>("transpose_12")]; | |
| tensor<fp16, [1, 12, 50, 64]> q_65_cast_fp16 = reshape(shape = var_1102, x = q_63_cast_fp16)[name = tensor<string, []>("q_65_cast_fp16")]; | |
| tensor<int32, [4]> var_1104 = const()[name = tensor<string, []>("op_1104"), val = tensor<int32, [4]>([1, 12, 50, 64])]; | |
| tensor<fp16, [12, 50, 64]> k_63_cast_fp16 = transpose(perm = k_63_perm_0, x = var_1092_cast_fp16)[name = tensor<string, []>("transpose_11")]; | |
| tensor<fp16, [1, 12, 50, 64]> k_65_cast_fp16 = reshape(shape = var_1104, x = k_63_cast_fp16)[name = tensor<string, []>("k_65_cast_fp16")]; | |
| tensor<int32, [4]> var_1106 = const()[name = tensor<string, []>("op_1106"), val = tensor<int32, [4]>([1, 12, 50, 64])]; | |
| tensor<fp16, [12, 50, 64]> v_63_cast_fp16 = transpose(perm = v_63_perm_0, x = var_1099_cast_fp16)[name = tensor<string, []>("transpose_10")]; | |
| tensor<fp16, [1, 12, 50, 64]> v_65_cast_fp16 = reshape(shape = var_1106, x = v_63_cast_fp16)[name = tensor<string, []>("v_65_cast_fp16")]; | |
| tensor<fp16, []> mul_21_y_0_to_fp16 = const()[name = tensor<string, []>("mul_21_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)]; | |
| tensor<fp16, [1, 12, 50, 64]> mul_21_cast_fp16 = mul(x = q_65_cast_fp16, y = mul_21_y_0_to_fp16)[name = tensor<string, []>("mul_21_cast_fp16")]; | |
| tensor<bool, []> matmul_10_transpose_y_0 = const()[name = tensor<string, []>("matmul_10_transpose_y_0"), val = tensor<bool, []>(true)]; | |
| tensor<bool, []> matmul_10_transpose_x_0 = const()[name = tensor<string, []>("matmul_10_transpose_x_0"), val = tensor<bool, []>(false)]; | |
| tensor<fp16, [1, 12, 50, 50]> matmul_10_cast_fp16 = matmul(transpose_x = matmul_10_transpose_x_0, transpose_y = matmul_10_transpose_y_0, x = mul_21_cast_fp16, y = k_65_cast_fp16)[name = tensor<string, []>("matmul_10_cast_fp16")]; | |
| tensor<int32, []> softmax_10_axis_0 = const()[name = tensor<string, []>("softmax_10_axis_0"), val = tensor<int32, []>(-1)]; | |
| tensor<fp16, [1, 12, 50, 50]> softmax_10_cast_fp16 = softmax(axis = softmax_10_axis_0, x = matmul_10_cast_fp16)[name = tensor<string, []>("softmax_10_cast_fp16")]; | |
| tensor<bool, []> attn_output_81_transpose_x_0 = const()[name = tensor<string, []>("attn_output_81_transpose_x_0"), val = tensor<bool, []>(false)]; | |
| tensor<bool, []> attn_output_81_transpose_y_0 = const()[name = tensor<string, []>("attn_output_81_transpose_y_0"), val = tensor<bool, []>(false)]; | |
| tensor<fp16, [1, 12, 50, 64]> attn_output_81_cast_fp16 = matmul(transpose_x = attn_output_81_transpose_x_0, transpose_y = attn_output_81_transpose_y_0, x = softmax_10_cast_fp16, y = v_65_cast_fp16)[name = tensor<string, []>("attn_output_81_cast_fp16")]; | |
| tensor<int32, [4]> var_1109 = const()[name = tensor<string, []>("op_1109"), val = tensor<int32, [4]>([2, 0, 1, 3])]; | |
| tensor<int32, [2]> var_1114 = const()[name = tensor<string, []>("op_1114"), val = tensor<int32, [2]>([50, 768])]; | |
| tensor<fp16, [50, 1, 12, 64]> var_1110_cast_fp16 = transpose(perm = var_1109, x = attn_output_81_cast_fp16)[name = tensor<string, []>("transpose_9")]; | |
| tensor<fp16, [50, 768]> attn_output_83_cast_fp16 = reshape(shape = var_1114, x = var_1110_cast_fp16)[name = tensor<string, []>("attn_output_83_cast_fp16")]; | |
| tensor<fp16, [768, 768]> model_visual_transformer_resblocks_10_attn_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("model_visual_transformer_resblocks_10_attn_out_proj_weight_to_fp16"), val = tensor<fp16, [768, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(150112384)))]; | |
| tensor<fp16, [768]> model_visual_transformer_resblocks_10_attn_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_visual_transformer_resblocks_10_attn_out_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(151292096)))]; | |
| tensor<fp16, [50, 768]> linear_41_cast_fp16 = linear(bias = model_visual_transformer_resblocks_10_attn_out_proj_bias_to_fp16, weight = model_visual_transformer_resblocks_10_attn_out_proj_weight_to_fp16, x = attn_output_83_cast_fp16)[name = tensor<string, []>("linear_41_cast_fp16")]; | |
| tensor<int32, [3]> var_1118 = const()[name = tensor<string, []>("op_1118"), val = tensor<int32, [3]>([50, 1, 768])]; | |
| tensor<fp16, [50, 1, 768]> attn_output_87_cast_fp16 = reshape(shape = var_1118, x = linear_41_cast_fp16)[name = tensor<string, []>("attn_output_87_cast_fp16")]; | |
| tensor<int32, [3]> var_1120_perm_0 = const()[name = tensor<string, []>("op_1120_perm_0"), val = tensor<int32, [3]>([1, 0, 2])]; | |
| tensor<fp16, [1, 50, 768]> var_1120_cast_fp16 = transpose(perm = var_1120_perm_0, x = attn_output_87_cast_fp16)[name = tensor<string, []>("transpose_8")]; | |
| tensor<fp16, [1, 50, 768]> input_105_cast_fp16 = add(x = input_103_cast_fp16, y = var_1120_cast_fp16)[name = tensor<string, []>("input_105_cast_fp16")]; | |
| tensor<int32, [1]> x_53_axes_0 = const()[name = tensor<string, []>("x_53_axes_0"), val = tensor<int32, [1]>([-1])]; | |
| tensor<fp16, [768]> model_visual_transformer_resblocks_10_ln_2_weight_to_fp16 = const()[name = tensor<string, []>("model_visual_transformer_resblocks_10_ln_2_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(151293696)))]; | |
| tensor<fp16, [768]> model_visual_transformer_resblocks_10_ln_2_bias_to_fp16 = const()[name = tensor<string, []>("model_visual_transformer_resblocks_10_ln_2_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(151295296)))]; | |
| tensor<fp16, [1, 50, 768]> x_53_cast_fp16 = layer_norm(axes = x_53_axes_0, beta = model_visual_transformer_resblocks_10_ln_2_bias_to_fp16, epsilon = var_13_to_fp16, gamma = model_visual_transformer_resblocks_10_ln_2_weight_to_fp16, x = input_105_cast_fp16)[name = tensor<string, []>("x_53_cast_fp16")]; | |
| tensor<fp16, [3072, 768]> model_visual_transformer_resblocks_10_mlp_c_fc_weight_to_fp16 = const()[name = tensor<string, []>("model_visual_transformer_resblocks_10_mlp_c_fc_weight_to_fp16"), val = tensor<fp16, [3072, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(151296896)))]; | |
| tensor<fp16, [3072]> model_visual_transformer_resblocks_10_mlp_c_fc_bias_to_fp16 = const()[name = tensor<string, []>("model_visual_transformer_resblocks_10_mlp_c_fc_bias_to_fp16"), val = tensor<fp16, [3072]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(156015552)))]; | |
| tensor<fp16, [1, 50, 3072]> linear_42_cast_fp16 = linear(bias = model_visual_transformer_resblocks_10_mlp_c_fc_bias_to_fp16, weight = model_visual_transformer_resblocks_10_mlp_c_fc_weight_to_fp16, x = x_53_cast_fp16)[name = tensor<string, []>("linear_42_cast_fp16")]; | |
| tensor<string, []> input_111_mode_0 = const()[name = tensor<string, []>("input_111_mode_0"), val = tensor<string, []>("EXACT")]; | |
| tensor<fp16, [1, 50, 3072]> input_111_cast_fp16 = gelu(mode = input_111_mode_0, x = linear_42_cast_fp16)[name = tensor<string, []>("input_111_cast_fp16")]; | |
| tensor<fp16, [768, 3072]> model_visual_transformer_resblocks_10_mlp_c_proj_weight_to_fp16 = const()[name = tensor<string, []>("model_visual_transformer_resblocks_10_mlp_c_proj_weight_to_fp16"), val = tensor<fp16, [768, 3072]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(156021760)))]; | |
| tensor<fp16, [768]> model_visual_transformer_resblocks_10_mlp_c_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_visual_transformer_resblocks_10_mlp_c_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(160740416)))]; | |
| tensor<fp16, [1, 50, 768]> linear_43_cast_fp16 = linear(bias = model_visual_transformer_resblocks_10_mlp_c_proj_bias_to_fp16, weight = model_visual_transformer_resblocks_10_mlp_c_proj_weight_to_fp16, x = input_111_cast_fp16)[name = tensor<string, []>("linear_43_cast_fp16")]; | |
| tensor<fp16, [1, 50, 768]> input_113_cast_fp16 = add(x = input_105_cast_fp16, y = linear_43_cast_fp16)[name = tensor<string, []>("input_113_cast_fp16")]; | |
| tensor<int32, [1]> x_55_axes_0 = const()[name = tensor<string, []>("x_55_axes_0"), val = tensor<int32, [1]>([-1])]; | |
| tensor<fp16, [768]> model_visual_transformer_resblocks_11_ln_1_weight_to_fp16 = const()[name = tensor<string, []>("model_visual_transformer_resblocks_11_ln_1_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(160742016)))]; | |
| tensor<fp16, [768]> model_visual_transformer_resblocks_11_ln_1_bias_to_fp16 = const()[name = tensor<string, []>("model_visual_transformer_resblocks_11_ln_1_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(160743616)))]; | |
| tensor<fp16, [1, 50, 768]> x_55_cast_fp16 = layer_norm(axes = x_55_axes_0, beta = model_visual_transformer_resblocks_11_ln_1_bias_to_fp16, epsilon = var_13_to_fp16, gamma = model_visual_transformer_resblocks_11_ln_1_weight_to_fp16, x = input_113_cast_fp16)[name = tensor<string, []>("x_55_cast_fp16")]; | |
| tensor<int32, [3]> query_perm_0 = const()[name = tensor<string, []>("query_perm_0"), val = tensor<int32, [3]>([1, 0, 2])]; | |
| tensor<fp16, [2304, 768]> model_visual_transformer_resblocks_11_attn_in_proj_weight_to_fp16 = const()[name = tensor<string, []>("model_visual_transformer_resblocks_11_attn_in_proj_weight_to_fp16"), val = tensor<fp16, [2304, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(160745216)))]; | |
| tensor<fp16, [2304]> model_visual_transformer_resblocks_11_attn_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_visual_transformer_resblocks_11_attn_in_proj_bias_to_fp16"), val = tensor<fp16, [2304]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(164284224)))]; | |
| tensor<fp16, [50, 1, 768]> query_cast_fp16 = transpose(perm = query_perm_0, x = x_55_cast_fp16)[name = tensor<string, []>("transpose_7")]; | |
| tensor<fp16, [50, 1, 2304]> linear_44_cast_fp16 = linear(bias = model_visual_transformer_resblocks_11_attn_in_proj_bias_to_fp16, weight = model_visual_transformer_resblocks_11_attn_in_proj_weight_to_fp16, x = query_cast_fp16)[name = tensor<string, []>("linear_44_cast_fp16")]; | |
| tensor<int32, [4]> concat_12 = const()[name = tensor<string, []>("concat_12"), val = tensor<int32, [4]>([50, 1, 3, 768])]; | |
| tensor<fp16, [50, 1, 3, 768]> var_1169_cast_fp16 = reshape(shape = concat_12, x = linear_44_cast_fp16)[name = tensor<string, []>("op_1169_cast_fp16")]; | |
| tensor<int32, [1]> var_1170_axes_0 = const()[name = tensor<string, []>("op_1170_axes_0"), val = tensor<int32, [1]>([0])]; | |
| tensor<fp16, [1, 50, 1, 3, 768]> var_1170_cast_fp16 = expand_dims(axes = var_1170_axes_0, x = var_1169_cast_fp16)[name = tensor<string, []>("op_1170_cast_fp16")]; | |
| tensor<int32, [5]> var_1171_perm_0 = const()[name = tensor<string, []>("op_1171_perm_0"), val = tensor<int32, [5]>([-2, 1, 2, 0, 4])]; | |
| tensor<int32, [1]> var_1172_axes_0 = const()[name = tensor<string, []>("op_1172_axes_0"), val = tensor<int32, [1]>([-2])]; | |
| tensor<fp16, [3, 50, 1, 1, 768]> var_1171_cast_fp16 = transpose(perm = var_1171_perm_0, x = var_1170_cast_fp16)[name = tensor<string, []>("transpose_6")]; | |
| tensor<fp16, [3, 50, 1, 768]> var_1172_cast_fp16 = squeeze(axes = var_1172_axes_0, x = var_1171_cast_fp16)[name = tensor<string, []>("op_1172_cast_fp16")]; | |
| tensor<int32, [4]> q_67_begin_0 = const()[name = tensor<string, []>("q_67_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; | |
| tensor<int32, [4]> q_67_end_0 = const()[name = tensor<string, []>("q_67_end_0"), val = tensor<int32, [4]>([1, 50, 1, 768])]; | |
| tensor<bool, [4]> q_67_end_mask_0 = const()[name = tensor<string, []>("q_67_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])]; | |
| tensor<bool, [4]> q_67_squeeze_mask_0 = const()[name = tensor<string, []>("q_67_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])]; | |
| tensor<fp16, [50, 1, 768]> q_67_cast_fp16 = slice_by_index(begin = q_67_begin_0, end = q_67_end_0, end_mask = q_67_end_mask_0, squeeze_mask = q_67_squeeze_mask_0, x = var_1172_cast_fp16)[name = tensor<string, []>("q_67_cast_fp16")]; | |
| tensor<int32, [4]> k_67_begin_0 = const()[name = tensor<string, []>("k_67_begin_0"), val = tensor<int32, [4]>([1, 0, 0, 0])]; | |
| tensor<int32, [4]> k_67_end_0 = const()[name = tensor<string, []>("k_67_end_0"), val = tensor<int32, [4]>([2, 50, 1, 768])]; | |
| tensor<bool, [4]> k_67_end_mask_0 = const()[name = tensor<string, []>("k_67_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])]; | |
| tensor<bool, [4]> k_67_squeeze_mask_0 = const()[name = tensor<string, []>("k_67_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])]; | |
| tensor<fp16, [50, 1, 768]> k_67_cast_fp16 = slice_by_index(begin = k_67_begin_0, end = k_67_end_0, end_mask = k_67_end_mask_0, squeeze_mask = k_67_squeeze_mask_0, x = var_1172_cast_fp16)[name = tensor<string, []>("k_67_cast_fp16")]; | |
| tensor<int32, [4]> v_67_begin_0 = const()[name = tensor<string, []>("v_67_begin_0"), val = tensor<int32, [4]>([2, 0, 0, 0])]; | |
| tensor<int32, [4]> v_67_end_0 = const()[name = tensor<string, []>("v_67_end_0"), val = tensor<int32, [4]>([3, 50, 1, 768])]; | |
| tensor<bool, [4]> v_67_end_mask_0 = const()[name = tensor<string, []>("v_67_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])]; | |
| tensor<bool, [4]> v_67_squeeze_mask_0 = const()[name = tensor<string, []>("v_67_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])]; | |
| tensor<fp16, [50, 1, 768]> v_67_cast_fp16 = slice_by_index(begin = v_67_begin_0, end = v_67_end_0, end_mask = v_67_end_mask_0, squeeze_mask = v_67_squeeze_mask_0, x = var_1172_cast_fp16)[name = tensor<string, []>("v_67_cast_fp16")]; | |
| tensor<int32, [3]> var_1180 = const()[name = tensor<string, []>("op_1180"), val = tensor<int32, [3]>([50, 12, 64])]; | |
| tensor<fp16, [50, 12, 64]> var_1181_cast_fp16 = reshape(shape = var_1180, x = q_67_cast_fp16)[name = tensor<string, []>("op_1181_cast_fp16")]; | |
| tensor<int32, [3]> q_69_perm_0 = const()[name = tensor<string, []>("q_69_perm_0"), val = tensor<int32, [3]>([1, 0, 2])]; | |
| tensor<int32, [3]> var_1187 = const()[name = tensor<string, []>("op_1187"), val = tensor<int32, [3]>([50, 12, 64])]; | |
| tensor<fp16, [50, 12, 64]> var_1188_cast_fp16 = reshape(shape = var_1187, x = k_67_cast_fp16)[name = tensor<string, []>("op_1188_cast_fp16")]; | |
| tensor<int32, [3]> k_69_perm_0 = const()[name = tensor<string, []>("k_69_perm_0"), val = tensor<int32, [3]>([1, 0, 2])]; | |
| tensor<int32, [3]> var_1194 = const()[name = tensor<string, []>("op_1194"), val = tensor<int32, [3]>([50, 12, 64])]; | |
| tensor<fp16, [50, 12, 64]> var_1195_cast_fp16 = reshape(shape = var_1194, x = v_67_cast_fp16)[name = tensor<string, []>("op_1195_cast_fp16")]; | |
| tensor<int32, [3]> v_69_perm_0 = const()[name = tensor<string, []>("v_69_perm_0"), val = tensor<int32, [3]>([1, 0, 2])]; | |
| tensor<int32, [4]> var_1198 = const()[name = tensor<string, []>("op_1198"), val = tensor<int32, [4]>([1, 12, 50, 64])]; | |
| tensor<fp16, [12, 50, 64]> q_69_cast_fp16 = transpose(perm = q_69_perm_0, x = var_1181_cast_fp16)[name = tensor<string, []>("transpose_5")]; | |
| tensor<fp16, [1, 12, 50, 64]> q_cast_fp16 = reshape(shape = var_1198, x = q_69_cast_fp16)[name = tensor<string, []>("q_cast_fp16")]; | |
| tensor<int32, [4]> var_1200 = const()[name = tensor<string, []>("op_1200"), val = tensor<int32, [4]>([1, 12, 50, 64])]; | |
| tensor<fp16, [12, 50, 64]> k_69_cast_fp16 = transpose(perm = k_69_perm_0, x = var_1188_cast_fp16)[name = tensor<string, []>("transpose_4")]; | |
| tensor<fp16, [1, 12, 50, 64]> k_cast_fp16 = reshape(shape = var_1200, x = k_69_cast_fp16)[name = tensor<string, []>("k_cast_fp16")]; | |
| tensor<int32, [4]> var_1202 = const()[name = tensor<string, []>("op_1202"), val = tensor<int32, [4]>([1, 12, 50, 64])]; | |
| tensor<fp16, [12, 50, 64]> v_69_cast_fp16 = transpose(perm = v_69_perm_0, x = var_1195_cast_fp16)[name = tensor<string, []>("transpose_3")]; | |
| tensor<fp16, [1, 12, 50, 64]> v_cast_fp16 = reshape(shape = var_1202, x = v_69_cast_fp16)[name = tensor<string, []>("v_cast_fp16")]; | |
| tensor<fp16, []> mul_23_y_0_to_fp16 = const()[name = tensor<string, []>("mul_23_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)]; | |
| tensor<fp16, [1, 12, 50, 64]> mul_23_cast_fp16 = mul(x = q_cast_fp16, y = mul_23_y_0_to_fp16)[name = tensor<string, []>("mul_23_cast_fp16")]; | |
| tensor<bool, []> matmul_11_transpose_y_0 = const()[name = tensor<string, []>("matmul_11_transpose_y_0"), val = tensor<bool, []>(true)]; | |
| tensor<bool, []> matmul_11_transpose_x_0 = const()[name = tensor<string, []>("matmul_11_transpose_x_0"), val = tensor<bool, []>(false)]; | |
| tensor<fp16, [1, 12, 50, 50]> matmul_11_cast_fp16 = matmul(transpose_x = matmul_11_transpose_x_0, transpose_y = matmul_11_transpose_y_0, x = mul_23_cast_fp16, y = k_cast_fp16)[name = tensor<string, []>("matmul_11_cast_fp16")]; | |
| tensor<int32, []> softmax_11_axis_0 = const()[name = tensor<string, []>("softmax_11_axis_0"), val = tensor<int32, []>(-1)]; | |
| tensor<fp16, [1, 12, 50, 50]> softmax_11_cast_fp16 = softmax(axis = softmax_11_axis_0, x = matmul_11_cast_fp16)[name = tensor<string, []>("softmax_11_cast_fp16")]; | |
| tensor<bool, []> attn_output_89_transpose_x_0 = const()[name = tensor<string, []>("attn_output_89_transpose_x_0"), val = tensor<bool, []>(false)]; | |
| tensor<bool, []> attn_output_89_transpose_y_0 = const()[name = tensor<string, []>("attn_output_89_transpose_y_0"), val = tensor<bool, []>(false)]; | |
| tensor<fp16, [1, 12, 50, 64]> attn_output_89_cast_fp16 = matmul(transpose_x = attn_output_89_transpose_x_0, transpose_y = attn_output_89_transpose_y_0, x = softmax_11_cast_fp16, y = v_cast_fp16)[name = tensor<string, []>("attn_output_89_cast_fp16")]; | |
| tensor<int32, [4]> var_1205 = const()[name = tensor<string, []>("op_1205"), val = tensor<int32, [4]>([2, 0, 1, 3])]; | |
| tensor<int32, [2]> var_1210 = const()[name = tensor<string, []>("op_1210"), val = tensor<int32, [2]>([50, 768])]; | |
| tensor<fp16, [50, 1, 12, 64]> var_1206_cast_fp16 = transpose(perm = var_1205, x = attn_output_89_cast_fp16)[name = tensor<string, []>("transpose_2")]; | |
| tensor<fp16, [50, 768]> attn_output_91_cast_fp16 = reshape(shape = var_1210, x = var_1206_cast_fp16)[name = tensor<string, []>("attn_output_91_cast_fp16")]; | |
| tensor<fp16, [768, 768]> model_visual_transformer_resblocks_11_attn_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("model_visual_transformer_resblocks_11_attn_out_proj_weight_to_fp16"), val = tensor<fp16, [768, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(164288896)))]; | |
| tensor<fp16, [768]> model_visual_transformer_resblocks_11_attn_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_visual_transformer_resblocks_11_attn_out_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(165468608)))]; | |
| tensor<fp16, [50, 768]> linear_45_cast_fp16 = linear(bias = model_visual_transformer_resblocks_11_attn_out_proj_bias_to_fp16, weight = model_visual_transformer_resblocks_11_attn_out_proj_weight_to_fp16, x = attn_output_91_cast_fp16)[name = tensor<string, []>("linear_45_cast_fp16")]; | |
| tensor<int32, [3]> var_1214 = const()[name = tensor<string, []>("op_1214"), val = tensor<int32, [3]>([50, 1, 768])]; | |
| tensor<fp16, [50, 1, 768]> attn_output_cast_fp16 = reshape(shape = var_1214, x = linear_45_cast_fp16)[name = tensor<string, []>("attn_output_cast_fp16")]; | |
| tensor<int32, [3]> var_1216_perm_0 = const()[name = tensor<string, []>("op_1216_perm_0"), val = tensor<int32, [3]>([1, 0, 2])]; | |
| tensor<fp16, [1, 50, 768]> var_1216_cast_fp16 = transpose(perm = var_1216_perm_0, x = attn_output_cast_fp16)[name = tensor<string, []>("transpose_1")]; | |
| tensor<fp16, [1, 50, 768]> input_115_cast_fp16 = add(x = input_113_cast_fp16, y = var_1216_cast_fp16)[name = tensor<string, []>("input_115_cast_fp16")]; | |
| tensor<int32, [1]> x_57_axes_0 = const()[name = tensor<string, []>("x_57_axes_0"), val = tensor<int32, [1]>([-1])]; | |
| tensor<fp16, [768]> model_visual_transformer_resblocks_11_ln_2_weight_to_fp16 = const()[name = tensor<string, []>("model_visual_transformer_resblocks_11_ln_2_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(165470208)))]; | |
| tensor<fp16, [768]> model_visual_transformer_resblocks_11_ln_2_bias_to_fp16 = const()[name = tensor<string, []>("model_visual_transformer_resblocks_11_ln_2_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(165471808)))]; | |
| tensor<fp16, [1, 50, 768]> x_57_cast_fp16 = layer_norm(axes = x_57_axes_0, beta = model_visual_transformer_resblocks_11_ln_2_bias_to_fp16, epsilon = var_13_to_fp16, gamma = model_visual_transformer_resblocks_11_ln_2_weight_to_fp16, x = input_115_cast_fp16)[name = tensor<string, []>("x_57_cast_fp16")]; | |
| tensor<fp16, [3072, 768]> model_visual_transformer_resblocks_11_mlp_c_fc_weight_to_fp16 = const()[name = tensor<string, []>("model_visual_transformer_resblocks_11_mlp_c_fc_weight_to_fp16"), val = tensor<fp16, [3072, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(165473408)))]; | |
| tensor<fp16, [3072]> model_visual_transformer_resblocks_11_mlp_c_fc_bias_to_fp16 = const()[name = tensor<string, []>("model_visual_transformer_resblocks_11_mlp_c_fc_bias_to_fp16"), val = tensor<fp16, [3072]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(170192064)))]; | |
| tensor<fp16, [1, 50, 3072]> linear_46_cast_fp16 = linear(bias = model_visual_transformer_resblocks_11_mlp_c_fc_bias_to_fp16, weight = model_visual_transformer_resblocks_11_mlp_c_fc_weight_to_fp16, x = x_57_cast_fp16)[name = tensor<string, []>("linear_46_cast_fp16")]; | |
| tensor<string, []> input_121_mode_0 = const()[name = tensor<string, []>("input_121_mode_0"), val = tensor<string, []>("EXACT")]; | |
| tensor<fp16, [1, 50, 3072]> input_121_cast_fp16 = gelu(mode = input_121_mode_0, x = linear_46_cast_fp16)[name = tensor<string, []>("input_121_cast_fp16")]; | |
| tensor<fp16, [768, 3072]> model_visual_transformer_resblocks_11_mlp_c_proj_weight_to_fp16 = const()[name = tensor<string, []>("model_visual_transformer_resblocks_11_mlp_c_proj_weight_to_fp16"), val = tensor<fp16, [768, 3072]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(170198272)))]; | |
| tensor<fp16, [768]> model_visual_transformer_resblocks_11_mlp_c_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_visual_transformer_resblocks_11_mlp_c_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(174916928)))]; | |
| tensor<fp16, [1, 50, 768]> linear_47_cast_fp16 = linear(bias = model_visual_transformer_resblocks_11_mlp_c_proj_bias_to_fp16, weight = model_visual_transformer_resblocks_11_mlp_c_proj_weight_to_fp16, x = input_121_cast_fp16)[name = tensor<string, []>("linear_47_cast_fp16")]; | |
| tensor<fp16, [1, 50, 768]> input_cast_fp16 = add(x = input_115_cast_fp16, y = linear_47_cast_fp16)[name = tensor<string, []>("input_cast_fp16")]; | |
| tensor<int32, [1]> x_59_axes_0 = const()[name = tensor<string, []>("x_59_axes_0"), val = tensor<int32, [1]>([-1])]; | |
| tensor<fp16, [768]> model_visual_ln_post_weight_to_fp16 = const()[name = tensor<string, []>("model_visual_ln_post_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(174918528)))]; | |
| tensor<fp16, [768]> model_visual_ln_post_bias_to_fp16 = const()[name = tensor<string, []>("model_visual_ln_post_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(174920128)))]; | |
| tensor<fp16, [1, 50, 768]> x_59_cast_fp16 = layer_norm(axes = x_59_axes_0, beta = model_visual_ln_post_bias_to_fp16, epsilon = var_13_to_fp16, gamma = model_visual_ln_post_weight_to_fp16, x = input_cast_fp16)[name = tensor<string, []>("x_59_cast_fp16")]; | |
| tensor<int32, [3]> pooled_begin_0 = const()[name = tensor<string, []>("pooled_begin_0"), val = tensor<int32, [3]>([0, 0, 0])]; | |
| tensor<int32, [3]> pooled_end_0 = const()[name = tensor<string, []>("pooled_end_0"), val = tensor<int32, [3]>([1, 1, 768])]; | |
| tensor<bool, [3]> pooled_end_mask_0 = const()[name = tensor<string, []>("pooled_end_mask_0"), val = tensor<bool, [3]>([true, false, true])]; | |
| tensor<bool, [3]> pooled_squeeze_mask_0 = const()[name = tensor<string, []>("pooled_squeeze_mask_0"), val = tensor<bool, [3]>([false, true, false])]; | |
| tensor<fp16, [1, 768]> pooled_cast_fp16 = slice_by_index(begin = pooled_begin_0, end = pooled_end_0, end_mask = pooled_end_mask_0, squeeze_mask = pooled_squeeze_mask_0, x = x_59_cast_fp16)[name = tensor<string, []>("pooled_cast_fp16")]; | |
| tensor<fp16, [512, 768]> transpose_0_to_fp16 = const()[name = tensor<string, []>("transpose_0_to_fp16"), val = tensor<fp16, [512, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(174921728)))]; | |
| tensor<fp16, [512]> var_1240_bias_0_to_fp16 = const()[name = tensor<string, []>("op_1240_bias_0_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(175708224)))]; | |
| tensor<fp16, [1, 512]> var_1240_cast_fp16 = linear(bias = var_1240_bias_0_to_fp16, weight = transpose_0_to_fp16, x = pooled_cast_fp16)[name = tensor<string, []>("op_1240_cast_fp16")]; | |
| tensor<string, []> var_1240_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_1240_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")]; | |
| tensor<fp32, [1, 512]> var_1240 = cast(dtype = var_1240_cast_fp16_to_fp32_dtype_0, x = var_1240_cast_fp16)[name = tensor<string, []>("cast_134")]; | |
| } -> (var_1240); | |
| } |