legacy_backward.yaml 61.5 KB
Newer Older
1
- backward_op : abs_double_grad
Z
zyfncg 已提交
2 3 4 5 6 7 8 9 10
  forward : abs_grad (Tensor x, Tensor grad_out) -> Tensor(grad_x)
  args : (Tensor x, Tensor grad_x_grad)
  output : Tensor(grad_out_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [x]
  kernel :
    func : abs_double_grad

11
- backward_op : abs_grad
Z
zyfncg 已提交
12 13 14 15 16 17 18 19 20 21
  forward : abs (Tensor x) -> Tensor(out)
  args : (Tensor x, Tensor out_grad)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [x]
  kernel :
    func : abs_grad
  backward : abs_double_grad

22
- backward_op : add_double_grad
Z
zyfncg 已提交
23 24 25 26 27 28 29 30 31 32 33 34
  forward : add_grad (Tensor x, Tensor y, Tensor grad_out, int axis = -1) -> Tensor(grad_x), Tensor(grad_y)
  args : (Tensor y, Tensor grad_out, Tensor grad_x_grad, Tensor grad_y_grad, int axis = -1)
  output : Tensor(grad_out_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [grad_out]
  kernel :
    func : add_double_grad
  optional : grad_x_grad, grad_y_grad
  backward : add_triple_grad
  inplace : (grad_x_grad -> grad_out_grad)

35
- backward_op : add_grad
Z
zyfncg 已提交
36 37 38 39 40 41 42 43 44 45 46 47
  forward : add (Tensor x, Tensor y) -> Tensor(out)
  args : (Tensor x, Tensor y, Tensor out_grad, int axis = -1)
  output : Tensor(x_grad), Tensor(y_grad)
  infer_meta :
    func : GeneralBinaryGradInferMeta
    param : [x, y]
  kernel :
    func : add_grad
  no_need_buffer : x, y
  backward : add_double_grad
  inplace : (out_grad -> x_grad)

48
- backward_op : add_triple_grad
Z
zyfncg 已提交
49 50 51 52 53 54 55 56 57 58
  forward : add_double_grad (Tensor y, Tensor grad_out, Tensor grad_grad_x, Tensor grad_grad_y, int axis = -1) -> Tensor(grad_grad_out)
  args : (Tensor grad_grad_x, Tensor grad_grad_y, Tensor grad_grad_out_grad, int axis = -1)
  output : Tensor(grad_grad_x_grad), Tensor(grad_grad_y_grad)
  infer_meta :
    func : GeneralBinaryGradInferMeta
    param : [grad_grad_x, grad_grad_y]
  kernel :
    func : add_triple_grad
  inplace : (grad_grad_out_grad -> grad_grad_x_grad)

59
- backward_op : affine_grid_grad
60 61
  forward : affine_grid (Tensor input, IntArray outputShape, bool align_corners=true) -> Tensor(output)
  args : (Tensor input, Tensor output_grad, IntArray outputShape, bool align_corners=true)
62 63 64 65 66 67 68
  output : Tensor(input_grad)
  infer_meta :
    func : AffineGridGradInferMeta
    param : [output_grad, outputShape, align_corners]
  kernel :
    func : affine_grid_grad
    param : [output_grad, outputShape, align_corners]
69
  no_need_buffer : input
70

71
- backward_op : amax_grad
72 73
  forward: amax (Tensor x,  int64_t[] axis={},  bool keepdim=false) -> Tensor(out)
  args : (Tensor x, Tensor out, Tensor out_grad, int64_t[] axis={},  bool keepdim=false, bool reduce_all=false)
74 75 76 77 78 79 80
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param: [x]
  kernel :
    func : amax_grad

81
- backward_op : amin_grad
82 83
  forward: amin (Tensor x,  int64_t[] axis={},  bool keepdim=false) -> Tensor(out)
  args : (Tensor x, Tensor out, Tensor out_grad, int64_t[] axis={},  bool keepdim=false, bool reduce_all=false)
84 85 86 87 88 89 90
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param: [x]
  kernel :
    func : amin_grad

91
- backward_op : assign_grad
Z
zyfncg 已提交
92 93 94
  forward : assign (Tensor x) -> Tensor(out)
  args : (Tensor out_grad)
  output : Tensor(x_grad)
95
  invoke : assign(out_grad)
Z
zyfncg 已提交
96

97
- backward_op : assign_out__grad
Z
zyfncg 已提交
98 99 100 101 102 103 104 105 106
  forward : assign_out_ (Tensor x, Tensor output) -> Tensor(out)
  args : (Tensor out_grad)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
  kernel :
    func : assign
  inplace : (out_grad -> x_grad)

107
- backward_op : batch_norm_double_grad
108 109
  forward : batch_norm_grad (Tensor x, Tensor scale, Tensor bias, Tensor out_mean, Tensor out_variance, Tensor saved_mean, Tensor saved_variance, Tensor reserve_space, Tensor grad_out, float momentum, float epsilon, str data_layout, bool is_test, bool use_global_stats, bool trainable_statistics) -> Tensor(grad_x), Tensor(grad_scale), Tensor(grad_bias)
  args : (Tensor x, Tensor scale, Tensor out_mean, Tensor out_variance, Tensor saved_mean, Tensor saved_variance, Tensor grad_out,  Tensor grad_x_grad, Tensor grad_scale_grad, Tensor grad_bias_grad, float momentum, float epsilon, str data_layout, bool is_test, bool use_global_stats, bool trainable_statistics)
Z
zyfncg 已提交
110 111 112 113 114 115 116
  output : Tensor(x_grad), Tensor(scale_grad), Tensor(grad_out_grad)
  infer_meta :
    func : GeneralTernaryGradInferMeta
    param : [x, scale, x]
  kernel :
    func : batch_norm_grad_grad
    data_type : x
117
  optional : out_mean, out_variance, grad_x_grad, grad_scale_grad, grad_bias_grad
Z
zyfncg 已提交
118 119
  inplace : (grad_out -> grad_out_grad)

120
- backward_op : batch_norm_grad
121 122
  forward : batch_norm (Tensor x, Tensor mean, Tensor variance, Tensor scale, Tensor bias, bool is_test, float momentum, float epsilon, str data_layout, bool use_global_stats, bool trainable_statistics) -> Tensor(out), Tensor(mean_out), Tensor(variance_out), Tensor(saved_mean), Tensor(saved_variance), Tensor(reserve_space)
  args : (Tensor x, Tensor scale, Tensor bias, Tensor mean_out, Tensor variance_out, Tensor saved_mean, Tensor saved_variance, Tensor reserve_space, Tensor out_grad, float momentum, float epsilon, str data_layout, bool is_test, bool use_global_stats, bool trainable_statistics)
Z
zyfncg 已提交
123 124 125 126 127 128 129 130 131 132
  output : Tensor(x_grad), Tensor(scale_grad), Tensor(bias_grad)
  infer_meta :
    func : GeneralTernaryGradInferMeta
    param : [x, scale, bias]
  kernel :
    func : batch_norm_grad
    data_type : out_grad
  optional : mean_out, variance_out, reserve_space
  backward : batch_norm_double_grad

133
- backward_op : bce_loss_grad
Z
zyfncg 已提交
134 135 136 137 138 139 140 141 142 143
  forward : bce_loss (Tensor input, Tensor label) -> Tensor(out)
  args : (Tensor input, Tensor label, Tensor out_grad)
  output : Tensor(input_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [input]
  kernel :
    func : bce_loss_grad
  inplace : (out_grad -> input_grad)

144
- backward_op : bicubic_interp_grad
145 146 147 148 149 150 151 152 153 154 155
  forward : bicubic_interp (Tensor x, Tensor out_size, Tensor[] size_tensor, Tensor scale_tensor, str data_layout, int out_d, int out_h, int out_w, float[] scale, str interp_method, bool align_corners, int align_mode) -> Tensor(output)
  args : (Tensor x, Tensor out_size, Tensor[] size_tensor, Tensor scale_tensor, Tensor output_grad, str data_layout, int out_d, int out_h, int out_w, float[] scale, str interp_method, bool align_corners, int align_mode)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param: [x]
  optional: out_size, size_tensor, scale_tensor
  kernel :
    func : bicubic_interp_grad
    data_type : output_grad

156
- backward_op : bilinear_interp_grad
157 158 159 160 161 162 163 164 165 166 167
  forward : bilinear_interp (Tensor x, Tensor out_size, Tensor[] size_tensor, Tensor scale_tensor, str data_layout, int out_d, int out_h, int out_w, float[] scale, str interp_method, bool align_corners, int align_mode) -> Tensor(output)
  args : (Tensor x, Tensor out_size, Tensor[] size_tensor, Tensor scale_tensor, Tensor output_grad, str data_layout, int out_d, int out_h, int out_w, float[] scale, str interp_method, bool align_corners, int align_mode)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param: [x]
  optional: out_size, size_tensor, scale_tensor
  kernel :
    func : bilinear_interp_grad
    data_type : output_grad

168
- backward_op : bilinear_tensor_product_grad
169 170 171 172 173 174 175 176
  forward : bilinear_tensor_product (Tensor x, Tensor y, Tensor weight, Tensor bias) -> Tensor(out)
  args : (Tensor x, Tensor y, Tensor weight, Tensor out_grad)
  output : Tensor(x_grad), Tensor(y_grad), Tensor(weight_grad), Tensor(bias_grad)
  infer_meta :
    func : BilinearTensorProductGradInferMeta
  kernel :
    func : bilinear_tensor_product_grad

177
- backward_op : broadcast_tensors_grad
178 179 180
  forward : broadcast_tensors (Tensor[] input) -> Tensor[](out)
  args : (Tensor[] input, Tensor[] out_grad)
  output : Tensor[](input_grad)
181 182
  infer_meta :
    func : UnchangedMultiInferMeta
183
    param : [input]
184 185 186
  kernel :
    func : broadcast_tensors_grad
    param : [out_grad]
187
  no_need_buffer : input
188

189
- backward_op : cast_grad
190
  forward : cast (Tensor x, DataType dtype) -> Tensor(out)
Z
zyfncg 已提交
191 192
  args : (Tensor x, Tensor out_grad)
  output : Tensor(x_grad)
193
  invoke : cast (out_grad, x.dtype())
Z
zyfncg 已提交
194 195
  no_need_buffer : x

196
- backward_op : concat_double_grad
Z
zyfncg 已提交
197 198 199
  forward : concat_grad (Tensor[] x, Tensor grad_out, Scalar axis) -> Tensor[](grad_x)
  args : (Tensor[] grad_x_grad, Scalar axis = 0)
  output : Tensor(grad_out_grad)
200
  invoke : concat(grad_x_grad, axis)
Z
zyfncg 已提交
201

202
- backward_op : concat_grad
Z
zyfncg 已提交
203 204 205 206 207 208 209 210 211 212 213
  forward : concat (Tensor[] x, Scalar axis) -> Tensor(out)
  args : (Tensor[] x, Tensor out_grad, Scalar axis = 0)
  output : Tensor[](x_grad){x.size()}
  infer_meta :
    func : UnchangedMultiInferMeta
    param : [x]
  kernel :
    func : concat_grad
  no_need_buffer : x
  backward : concat_double_grad

214
- backward_op : conv2d_grad
215 216
  forward : conv2d (Tensor input, Tensor filter, int[] strides, int[] paddings, str padding_algorithm, int[] dilations, int groups, str data_format) -> Tensor(out)
  args : (Tensor input, Tensor filter, Tensor out_grad,  int[] strides, int[] paddings, str padding_algorithm, int[] dilations, int groups, str data_format)
Z
zyfncg 已提交
217
  output : Tensor(input_grad), Tensor(filter_grad)
Z
zyfncg 已提交
218 219 220 221 222
  infer_meta :
    func : GeneralBinaryGradInferMeta
    param : [input, filter]
  kernel :
    func : conv2d_grad
Z
zyfncg 已提交
223 224
  backward : conv2d_grad_grad

225
- backward_op : conv2d_grad_grad
226 227
  forward : conv2d_grad (Tensor input, Tensor filter, Tensor grad_out,  int[] strides, int[] paddings, str padding_algorithm, int[] dilations, int groups, str data_format) -> Tensor(grad_input), Tensor(grad_filter)
  args : (Tensor input, Tensor filter, Tensor grad_out, Tensor grad_input_grad, Tensor grad_filter_grad, int[] strides, int[] paddings, str padding_algorithm, int[] dilations, int groups, str data_format)
Z
zyfncg 已提交
228 229 230 231 232 233 234 235
  output : Tensor(input_grad), Tensor(filter_grad), Tensor(grad_out_grad)
  infer_meta :
    func : GeneralTernaryGradInferMeta
    param: [input, filter, grad_out]
  kernel :
    func : conv2d_grad_grad
  optional : grad_input_grad, grad_filter_grad

236
- backward_op : conv2d_transpose_double_grad
237 238
  forward : conv2d_transpose_grad(Tensor x, Tensor filter, Tensor grad_out, int[] strides, int[] paddings, int[] output_padding, IntArray output_size, str padding_algorithm, int groups, int[] dilations, str data_format) -> Tensor(grad_x), Tensor(grad_filter)
  args : (Tensor x, Tensor filter, Tensor grad_out, Tensor grad_x_grad, Tensor grad_filter_grad, int[] strides, int[] paddings, int[] output_padding, IntArray output_size, str padding_algorithm, int groups, int[] dilations, str data_format)
Z
zyfncg 已提交
239 240 241 242 243 244
  output : Tensor(x_grad), Tensor(filter_grad), Tensor(grad_out_grad)
  infer_meta :
    func : Conv2dTransposeDoubleGradInferMeta
  kernel :
    func : conv2d_transpose_grad_grad

245
- backward_op : conv2d_transpose_grad
246 247
  forward : conv2d_transpose(Tensor x, Tensor filter, int[] strides, int[] paddings, int[] output_padding, IntArray output_size, str padding_algorithm, int groups, int[] dilations, str data_format) -> Tensor(out)
  args : (Tensor x, Tensor filter, Tensor out_grad, int[] strides, int[] paddings, int[] output_padding, IntArray output_size, str padding_algorithm, int groups, int[] dilations, str data_format)
Z
zyfncg 已提交
248 249
  output : Tensor(x_grad), Tensor(filter_grad)
  infer_meta :
250
    func : Conv2dTransposeGradInferMeta
Z
zyfncg 已提交
251 252 253 254
  kernel :
    func : conv2d_transpose_grad
  backward : conv2d_transpose_double_grad

255 256 257 258 259 260 261 262 263 264 265
- backward_op : conv3d_double_grad
  forward : conv3d_grad (Tensor input, Tensor filter, Tensor grad_out,  int[] strides, int[] paddings, str padding_algorithm, int groups, int[] dilations, str data_format) -> Tensor(grad_input), Tensor(grad_filter)
  args : (Tensor input, Tensor filter, Tensor grad_out, Tensor grad_input_grad, Tensor grad_filter_grad, int[] strides, int[] paddings, str padding_algorithm, int groups, int[] dilations, str data_format)
  output : Tensor(input_grad), Tensor(filter_grad), Tensor(grad_out_grad)
  infer_meta :
    func : GeneralTernaryGradInferMeta
    param: [input, filter, grad_out]
  kernel :
    func : conv3d_double_grad
  optional : grad_input_grad, grad_filter_grad

266
- backward_op : conv3d_grad
267 268
  forward : conv3d (Tensor input, Tensor filter, int[] strides, int[] paddings, str padding_algorithm, int groups, int[] dilations, str data_format) -> Tensor(out)
  args : (Tensor input, Tensor filter, Tensor out_grad,  int[] strides, int[] paddings, str padding_algorithm, int groups, int[] dilations, str data_format)
Z
zyfncg 已提交
269
  output : Tensor(input_grad), Tensor(filter_grad)
Z
zyfncg 已提交
270 271 272 273 274
  infer_meta :
    func : GeneralBinaryGradInferMeta
    param : [input, filter]
  kernel :
    func : conv3d_grad
275
  backward : conv3d_double_grad
Z
zyfncg 已提交
276

277
- backward_op : conv3d_transpose_grad
Z
zyfncg 已提交
278 279 280 281 282 283 284 285
  forward : conv3d_transpose(Tensor x, Tensor filter, int[] strides, int[] paddings, int[] output_padding, int[] output_size, str padding_algorithm, int groups, int[] dilations, str data_format) -> Tensor(out)
  args : (Tensor x, Tensor filter, Tensor out_grad, int[] strides, int[] paddings, int[] output_padding, int[] output_size, str padding_algorithm, int groups, int[] dilations, str data_format)
  output : Tensor(x_grad), Tensor(filter_grad)
  infer_meta :
    func : ConvTransposeGradInferMeta
  kernel :
    func : conv3d_transpose_grad

286
- backward_op : cross_entropy_with_softmax_grad
Z
zyfncg 已提交
287 288 289 290 291 292 293 294 295 296
  forward : cross_entropy_with_softmax (Tensor input, Tensor label, bool soft_label, bool use_softmax, bool numeric_stable_mode, int ignore_index, int axis) -> Tensor(softmax), Tensor(loss)
  args : (Tensor label, Tensor softmax, Tensor loss_grad, bool soft_label, bool use_softmax, bool numeric_stable_mode, int ignore_index, int axis)
  output : Tensor(input_grad)
  infer_meta :
    func : CrossEntropyWithSoftmaxGradInferMeta
  kernel :
    func : cross_entropy_with_softmax_grad
    data_type : softmax
  inplace : (softmax -> input_grad)

297
- backward_op : cumprod_grad
Z
zyfncg 已提交
298 299 300 301 302 303 304 305 306
  forward : cumprod (Tensor x, int dim) -> Tensor(out)
  args : (Tensor x, Tensor out, Tensor out_grad, int dim)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param: [x]
  kernel :
    func : cumprod_grad

307
- backward_op : cumsum_grad
W
WangZhen 已提交
308 309
  forward : cumsum(Tensor x, Scalar axis, bool flatten, bool exclusive, bool reverse) -> Tensor(out)
  args : (Tensor out_grad, Scalar axis, bool flatten, bool exclusive, bool reverse)
Z
zyfncg 已提交
310 311 312
  output : Tensor(x_grad)
  invoke : cumsum(out_grad, axis, flatten, exclusive, !reverse)

313
- backward_op : deformable_conv_grad
Z
zyfncg 已提交
314 315 316 317 318 319 320 321 322 323
  forward : deformable_conv(Tensor x, Tensor offset, Tensor filter, Tensor mask, int[] strides, int[] paddings, int[] dilations, int deformable_groups, int groups, int im2col_step) -> Tensor(out)
  args : (Tensor x, Tensor offset, Tensor filter, Tensor mask, Tensor out_grad, int[] strides, int[] paddings, int[] dilations, int deformable_groups, int groups, int im2col_step)
  output : Tensor(x_grad), Tensor(offset_grad), Tensor(filter_grad), Tensor(mask_grad)
  infer_meta :
    func : DeformableConvGradInferMeta
  kernel :
    func : deformable_conv_grad
    data_type : x
  optional : mask

324
- backward_op : depthwise_conv2d_double_grad
325
  forward : depthwise_conv2d_grad (Tensor input, Tensor filter, Tensor grad_out, int[] strides, int[] paddings, str padding_algorithm, int groups, int[] dilations, str data_format) -> Tensor(grad_input), Tensor(grad_filter)
326 327 328 329 330 331 332 333 334
  args : (Tensor input, Tensor filter, Tensor grad_out, Tensor grad_input_grad, Tensor grad_filter_grad, int[] strides, int[] paddings, str padding_algorithm, int groups, int[] dilations, str data_format)
  output : Tensor(input_grad), Tensor(filter_grad), Tensor(grad_out_grad)
  infer_meta :
    func : GeneralTernaryGradInferMeta
    param: [input, filter, grad_out]
  kernel :
    func : depthwise_conv2d_double_grad
  optional : grad_input_grad, grad_filter_grad

335
- backward_op : depthwise_conv2d_grad
336 337
  forward : depthwise_conv2d (Tensor input, Tensor filter, int[] strides, int[] paddings, str padding_algorithm, int groups, int[] dilations, str data_format) -> Tensor(out)
  args : (Tensor input, Tensor filter, Tensor out_grad, int[] strides, int[] paddings, str padding_algorithm, int groups, int[] dilations, str data_format)
Z
zyfncg 已提交
338 339 340 341 342 343
  output : Tensor(input_grad), Tensor(filter_grad)
  infer_meta :
    func : GeneralBinaryGradInferMeta
    param : [input, filter]
  kernel :
    func : depthwise_conv2d_grad
344 345
    param : [input, filter, out_grad, strides, paddings, padding_algorithm, groups, dilations, data_format]
  backward : depthwise_conv2d_double_grad
Z
zyfncg 已提交
346

347
- backward_op : depthwise_conv2d_transpose_grad
348 349
  forward : depthwise_conv2d_transpose(Tensor x, Tensor filter, int[] strides, int[] paddings, int[] output_padding, IntArray output_size, str padding_algorithm, int groups, int[] dilations, str data_format) -> Tensor(out)
  args : (Tensor x, Tensor filter, Tensor out_grad, int[] strides, int[] paddings, int[] output_padding, IntArray output_size, str padding_algorithm, int groups, int[] dilations, str data_format)
Z
zyfncg 已提交
350 351
  output : Tensor(x_grad), Tensor(filter_grad)
  infer_meta :
352
    func : Conv2dTransposeGradInferMeta
Z
zyfncg 已提交
353 354 355
  kernel :
    func : depthwise_conv2d_transpose_grad

356
- backward_op : divide_double_grad
Z
zyfncg 已提交
357 358 359 360 361 362 363 364 365 366 367 368
  forward : divide_grad (Tensor x, Tensor y, Tensor out, Tensor grad_out, int axis = -1) -> Tensor(grad_x), Tensor(grad_y)
  args : (Tensor y, Tensor out, Tensor grad_x, Tensor grad_x_grad, Tensor grad_y_grad, int axis = -1)
  output : Tensor(y_grad), Tensor(out_grad), Tensor(grad_out_grad)
  infer_meta :
    func : GeneralTernaryGradInferMeta
    param : [y, grad_x, grad_x]
  kernel :
    func : divide_double_grad
    data_type : out
  optional : grad_x_grad, grad_y_grad
  inplace : (grad_x_grad -> grad_out_grad)

369
- backward_op : divide_grad
Z
zyfncg 已提交
370 371 372 373 374 375 376 377 378 379
  forward : divide (Tensor x, Tensor y) -> Tensor(out)
  args : (Tensor x, Tensor y, Tensor out, Tensor out_grad, int axis = -1)
  output : Tensor(x_grad), Tensor(y_grad)
  infer_meta :
    func : GeneralBinaryGradInferMeta
    param : [x, y]
  kernel :
    func : divide_grad
  backward : divide_double_grad

380
- backward_op : dropout_grad
381 382
  forward : dropout (Tensor x, Tensor seed_tensor, Scalar p, bool is_test, str mode, int seed, bool fix_seed) -> Tensor(out), Tensor(mask)
  args : (Tensor mask, Tensor out_grad, Scalar p, bool is_test, str mode)
Z
zyfncg 已提交
383 384 385 386 387 388 389
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [out_grad]
  kernel :
    func : dropout_grad

390
- backward_op : eigvalsh_grad
391 392 393 394 395 396 397 398 399 400 401
  forward : eigvalsh (Tensor x, str uplo, bool is_test) -> Tensor(eigenvalues), Tensor(eigenvectors)
  args : (Tensor eigenvectors, Tensor eigenvalues_grad, str uplo, bool is_test)
  output : Tensor(x_grad)
  infer_meta :
    func : EigvalshGradInferMeta
  kernel :
    func : eigvalsh_grad
    data_type : eigenvectors
  data_transform :
    skip_transform : eigenvalues_grad

402
- backward_op : einsum_grad
Z
zyfncg 已提交
403 404 405 406 407 408 409 410 411
  forward : einsum (Tensor[] x, str equation) -> Tensor(out), Tensor[](inner_cache), Tensor[](x_shape)
  args : (Tensor[] x_shape, Tensor[] inner_cache, Tensor out_grad, str equation)
  output : Tensor[](x_grad){x.size()}
  infer_meta :
    func : UnchangedMultiInferMeta
    param : [x_shape]
  kernel :
    func : einsum_grad

412
- backward_op : elementwise_pow_grad
Z
zyfncg 已提交
413 414 415 416 417 418 419 420 421
  forward : elementwise_pow(Tensor x, Tensor y) -> Tensor(out)
  args : (Tensor x, Tensor y, Tensor out_grad, int axis=-1)
  output : Tensor(x_grad), Tensor(y_grad)
  infer_meta :
    func : GeneralBinaryGradInferMeta
    param: [x, y]
  kernel :
    func : elementwise_pow_grad

422
- backward_op : embedding_grad
Z
zyfncg 已提交
423 424 425 426
  forward : embedding (Tensor x, Tensor weight, int64_t padding_idx=-1, bool sparse=false) -> Tensor(out)
  args : (Tensor x, Tensor weight, Tensor out_grad, int64_t padding_idx=-1, bool sparse=false)
  output : Tensor(weight_grad)
  invoke : embedding_grad_impl(x, weight, out_grad, padding_idx, sparse, weight_grad)
W
wanghuancoder 已提交
427
  no_need_buffer : weight
Z
zyfncg 已提交
428

429
- backward_op : expand_as_grad
Z
zyfncg 已提交
430 431 432 433 434 435 436 437 438 439
  forward : expand_as (Tensor x, Tensor y, int[] target_shape) -> Tensor(out)
  args : (Tensor x, Tensor out_grad, int[] target_shape)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [x]
  kernel :
    func : expand_as_grad
  no_need_buffer : x

440
- backward_op : expand_double_grad
Z
zyfncg 已提交
441 442 443
  forward : expand_grad (Tensor x, Tensor grad_out, IntArray shape) -> Tensor(grad_x)
  args : (Tensor grad_x_grad, IntArray shape)
  output : Tensor(grad_out_grad)
444
  invoke : expand(grad_x_grad, shape)
Z
zyfncg 已提交
445

446
- backward_op : expand_grad
Z
zyfncg 已提交
447 448 449 450 451 452 453 454 455 456 457
  forward : expand (Tensor x, IntArray shape) -> Tensor(out)
  args : (Tensor x, Tensor out_grad, IntArray shape)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [x]
  kernel :
    func : expand_grad
  no_need_buffer : x
  backward : expand_double_grad

458
- backward_op : exponential__grad
459
  forward : exponential_ (Tensor x, float lam) -> Tensor(out)
460 461 462 463
  args : (Tensor out_grad)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
464
  invoke : zeros_like(out_grad)
465

466
- backward_op : fill_grad
467 468 469 470 471 472 473 474 475 476
  forward : fill (Tensor x, Scalar value) -> Tensor(out)
  args : (Tensor out_grad, Scalar value)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [out_grad]
  kernel :
    func : fill_grad
  inplace : (out_grad -> x_grad)

477
- backward_op : flatten_grad
Z
zyfncg 已提交
478 479 480 481 482 483 484 485 486 487 488 489 490
  forward : flatten(Tensor x, int start_axis, int stop_axis) -> Tensor(out), Tensor(xshape)
  args : (Tensor xshape, Tensor out_grad)
  output : Tensor(x_grad)
  infer_meta :
    func :  KernelWithXShapeInferMeta
    param : [xshape]
  kernel :
    func : flatten_grad
    data_type: out_grad
    backend: out_grad
    layout: out_grad
  inplace : (out_grad -> x_grad)

491
- backward_op : fmax_grad
492 493
  forward : fmax(Tensor x, Tensor y) -> Tensor(out)
  args : (Tensor x, Tensor y, Tensor out_grad, int axis = -1)
Z
zyfncg 已提交
494 495 496 497 498 499 500
  output : Tensor(x_grad), Tensor(y_grad)
  infer_meta :
    func : GeneralBinaryGradInferMeta
    param: [x, y]
  kernel :
    func : fmax_grad

501
- backward_op : fmin_grad
502 503
  forward : fmin(Tensor x, Tensor y) -> Tensor(out)
  args : (Tensor x, Tensor y, Tensor out_grad, int axis = -1)
Z
zyfncg 已提交
504 505 506 507 508 509 510
  output : Tensor(x_grad), Tensor(y_grad)
  infer_meta :
    func : GeneralBinaryGradInferMeta
    param: [x, y]
  kernel :
    func : fmin_grad

511
- backward_op : frobenius_norm_grad
Z
zyfncg 已提交
512 513 514 515 516 517 518 519 520
  forward : frobenius_norm(Tensor x, int64_t[] axis,  bool keep_dim,  bool reduce_all) -> Tensor(out)
  args : (Tensor x, Tensor out, Tensor out_grad, int64_t[] axis,  bool keep_dim,  bool reduce_all)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [x]
  kernel :
    func : frobenius_norm_grad

521
- backward_op : gather_grad
Z
zyfncg 已提交
522 523 524 525 526 527 528 529 530 531 532
  forward : gather(Tensor x, Tensor index, Scalar axis=0) -> Tensor(out)
  args : (Tensor x, Tensor index, Tensor out_grad, Scalar axis=0, bool overwrite=false)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param: [x]
  kernel :
    data_type: x
    func : gather_grad
  no_need_buffer : x

533
- backward_op : group_norm_grad
Z
zyfncg 已提交
534 535 536 537 538 539 540 541 542 543 544 545
  forward : group_norm (Tensor x, Tensor scale, Tensor bias, float epsilon, int groups, str data_layout) -> Tensor(y), Tensor(mean), Tensor(variance)
  args : (Tensor x, Tensor scale, Tensor bias, Tensor y, Tensor mean, Tensor variance, Tensor y_grad, float epsilon, int groups, str data_layout)
  output : Tensor(x_grad), Tensor(scale_grad), Tensor(bias_grad)
  infer_meta :
    func : GeneralTernaryGradInferMeta
    param : [y, scale, bias]
  kernel :
    func : group_norm_grad
    data_type : y_grad
  optional: scale, bias
  inplace : (y_grad -> x_grad)

546
- backward_op : hardswish_grad
547 548
  forward : hardswish (Tensor x) -> Tensor(out)
  args : (Tensor x, Tensor out_grad, float threshold = 6.0, float scale = 6.0, float offset = 3.0)
Z
zyfncg 已提交
549 550 551 552 553
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [x]
  kernel :
Z
zyfncg 已提交
554
    func : hardswish_grad
Z
zyfncg 已提交
555 556
  inplace : (out_grad -> x_grad)

557
- backward_op : hsigmoid_loss_grad
558 559
  forward : hsigmoid_loss (Tensor x, Tensor label, Tensor w, Tensor bias, Tensor path, Tensor code, int num_classes, bool remote_prefetch, bool is_sparse) -> Tensor(out), Tensor(pre_out), Tensor(w_out)
  args : (Tensor x, Tensor w, Tensor label, Tensor path, Tensor code, Tensor bias, Tensor pre_out, Tensor out_grad, int num_classes, bool remote_prefetch, bool is_sparse)
560 561 562 563 564 565
  output : Tensor(x_grad), Tensor(w_grad), Tensor(bias_grad)
  infer_meta :
    func : GeneralTernaryGradInferMeta
    param : [x ,w, bias]
  optional: path, code, bias
  kernel :
566
    func : hsigmoid_loss_grad
567

568
- backward_op : huber_loss_grad
Z
zyfncg 已提交
569 570 571 572 573 574 575 576 577
  forward : huber_loss (Tensor input, Tensor label, float delta) -> Tensor(out), Tensor(residual)
  args : (Tensor residual, Tensor out_grad, float delta)
  output : Tensor(input_grad), Tensor(label_grad)
  infer_meta :
    func : GeneralBinaryGradInferMeta
    param : [residual, residual]
  kernel :
    func : huber_loss_grad

578
- backward_op : index_add_grad
L
Li Min 已提交
579 580 581 582 583 584 585 586 587 588
  forward : index_add(Tensor x, Tensor index,  Tensor add_value, int axis) -> Tensor(out)
  args : (Tensor index, Tensor add_value, Tensor out_grad, int axis)
  output : Tensor(x_grad), Tensor(add_value_grad)
  infer_meta :
    func : IndexAddGradInferMeta
  kernel :
    func : index_add_grad
    data_type : out_grad
  inplace : (out_grad -> x_grad)

589
- backward_op : instance_norm_double_grad
Z
zyfncg 已提交
590 591 592 593 594 595 596 597 598 599
  forward : instance_norm_grad(Tensor x, Tensor fwd_scale, Tensor saved_mean, Tensor saved_variance, Tensor grad_y, float epsilon) -> Tensor(grad_x), Tensor(grad_scale), Tensor(grad_bias)
  args : (Tensor x, Tensor fwd_scale, Tensor saved_mean, Tensor saved_variance, Tensor grad_y, Tensor grad_x_grad, Tensor grad_scale_grad, Tensor grad_bias_grad, float epsilon)
  output : Tensor(x_grad), Tensor(fwd_scale_grad), Tensor(grad_y_grad)
  infer_meta :
    func : InstanceNormDoubleGradInferMeta
  kernel :
    func : instance_norm_double_grad
    data_type : x
  optional : fwd_scale, grad_x_grad, grad_scale_grad, grad_bias_grad

600
- backward_op : instance_norm_grad
Z
zyfncg 已提交
601 602 603 604 605 606 607 608 609 610 611
  forward : instance_norm(Tensor x, Tensor scale, Tensor bias, float epsilon) -> Tensor(y), Tensor(saved_mean), Tensor(saved_variance)
  args : (Tensor x, Tensor scale, Tensor saved_mean, Tensor saved_variance, Tensor y_grad, float epsilon)
  output : Tensor(x_grad), Tensor(scale_grad), Tensor(bias_grad)
  infer_meta :
    func : InstanceNormGradInferMeta
  kernel :
    func : instance_norm_grad
    data_type : x
  optional : scale
  backward : instance_norm_double_grad

612
- backward_op : kldiv_loss_grad
Z
zyfncg 已提交
613 614 615 616 617 618 619 620 621 622
  forward : kldiv_loss(Tensor x, Tensor label, str reduction) -> Tensor(out)
  args : (Tensor x, Tensor label, Tensor out_grad, str reduction)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param: [x]
  kernel :
    func : kldiv_loss_grad
  no_need_buffer : x

623
- backward_op : kron_grad
Z
zyfncg 已提交
624 625 626 627 628 629 630 631 632 633
  forward : kron (Tensor x, Tensor y) -> Tensor(out)
  args : (Tensor x, Tensor y, Tensor out_grad)
  output : Tensor(x_grad), Tensor(y_grad)
  infer_meta :
    func : GeneralBinaryGradInferMeta
    param : [x, y]
  kernel :
    func : kron_grad
    data_type : out_grad

634
- backward_op : layer_norm_grad
635 636
  forward : layer_norm (Tensor x, Tensor scale, Tensor bias, float epsilon, int begin_norm_axis) -> Tensor(out), Tensor(mean), Tensor(variance)
  args : (Tensor x,  Tensor scale, Tensor bias, Tensor mean, Tensor variance, Tensor out_grad, float epsilon, int begin_norm_axis)
Z
zyfncg 已提交
637 638 639 640 641 642 643 644 645 646
  output : Tensor(x_grad), Tensor(scale_grad), Tensor(bias_grad)
  infer_meta :
    func : LayerNormGradInferMeta
    param : [x, scale, bias]
  kernel :
    func : layer_norm_grad
    data_type : out_grad
  no_need_buffer : bias
  optional : scale, bias

647
- backward_op : linear_interp_grad
648
  forward : linear_interp (Tensor x, Tensor out_size, Tensor[] size_tensor, Tensor scale_tensor, str data_layout, int out_d, int out_h, int out_w, float[] scale, str interp_method, bool align_corners, int align_mode) -> Tensor(output)
649 650 651 652 653 654 655
  args : (Tensor x, Tensor out_size, Tensor[] size_tensor, Tensor scale_tensor, Tensor output_grad, str data_layout, int out_d, int out_h, int out_w, float[] scale, str interp_method, bool align_corners, int align_mode)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param: [x]
  optional: out_size, size_tensor, scale_tensor
  kernel :
656
    func : linear_interp_grad
657 658
    data_type : output_grad

659
- backward_op : log_softmax_grad
Z
zyfncg 已提交
660 661 662 663 664 665 666 667 668
  forward : log_softmax(Tensor x,  int axis) -> Tensor(out)
  args : (Tensor out, Tensor out_grad,  int axis)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param: [out]
  kernel :
    func : log_softmax_grad

669
- backward_op : logcumsumexp_grad
Z
zyfncg 已提交
670 671 672 673 674 675 676 677 678
  forward : logcumsumexp(Tensor x, int axis, bool flatten, bool exclusive, bool reverse) -> Tensor(out)
  infer_meta :
    func : UnchangedInferMeta
    param : [x]
  args : (Tensor x, Tensor out, Tensor out_grad, int axis, bool flatten, bool exclusive, bool reverse)
  output : Tensor(x_grad)
  kernel :
    func : logcumsumexp_grad

679
- backward_op : logsumexp_grad
Z
zyfncg 已提交
680 681 682 683 684 685 686 687 688
  forward : logsumexp(Tensor x, int64_t[] axis,  bool keepdim,  bool reduce_all) -> Tensor(out)
  args : (Tensor x, Tensor out, Tensor out_grad, int64_t[] axis,  bool keepdim,  bool reduce_all)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param: [x]
  kernel :
    func : logsumexp_grad

689
- backward_op : lu_grad
L
Lin Manhui 已提交
690 691 692 693 694 695 696 697
  forward : lu (Tensor x, bool pivot) -> Tensor(out), Tensor(pivots), Tensor(infos)
  args : (Tensor x, Tensor out, Tensor pivots, Tensor out_grad, bool pivot)
  output : Tensor(x_grad)
  infer_meta :
    func : LUGradInferMeta
  kernel :
    func : lu_grad

698
- backward_op : margin_cross_entropy_grad
699 700 701 702 703 704 705 706 707 708
  forward : margin_cross_entropy (Tensor logits, Tensor label, bool return_softmax, int ring_id, int rank, int nranks, float margin1, float margin2, float margin3, float scale) -> Tensor(softmax), Tensor(loss)
  args : (Tensor logits, Tensor label, Tensor softmax, Tensor loss_grad, bool return_softmax, int ring_id, int rank, int nranks, float margin1, float margin2, float margin3, float scale)
  output : Tensor(logits_grad)
  infer_meta :
    func : MarginCrossEntropyGradInferMeta
  kernel :
    func : margin_cross_entropy_grad
    data_type : softmax
  inplace : (softmax -> logits_grad)

709
- backward_op : matmul_double_grad
Z
zyfncg 已提交
710 711 712 713 714 715 716 717 718 719 720
  forward : matmul_grad (Tensor x, Tensor y, Tensor grad_out, bool transpose_x=false, bool transpose_y=false) -> Tensor(grad_x), Tensor(grad_y)
  args : (Tensor x, Tensor y, Tensor grad_out, Tensor grad_x_grad, Tensor grad_y_grad, bool transpose_x=false, bool transpose_y=false)
  output : Tensor(x_grad), Tensor(y_grad), Tensor(grad_out_grad)
  infer_meta :
    func : GeneralTernaryGradInferMeta
    param : [x, y, grad_out]
  kernel :
    func : matmul_double_grad
  backward : matmul_triple_grad
  optional : grad_x_grad, grad_y_grad

721
- backward_op : matmul_grad
Z
zyfncg 已提交
722 723 724 725 726 727 728 729 730 731
  forward : matmul (Tensor x, Tensor y, bool transpose_x=false, bool transpose_y=false) -> Tensor(out)
  args : (Tensor x, Tensor y, Tensor out_grad, bool transpose_x=false, bool transpose_y=false)
  output : Tensor(x_grad), Tensor(y_grad)
  infer_meta :
    func : GeneralBinaryGradInferMeta
    param : [x, y]
  kernel :
    func : matmul_grad
  backward : matmul_double_grad

732
- backward_op : matmul_triple_grad
Z
zyfncg 已提交
733 734 735 736 737 738 739 740
  forward : matmul_double_grad (Tensor x, Tensor y, Tensor fwd_grad_out, Tensor fwd_grad_grad_x, Tensor fwd_grad_grad_y, bool transpose_x=false, bool transpose_y=false) -> Tensor(grad_x), Tensor(grad_y), Tensor(grad_grad_out)
  args : (Tensor x, Tensor y, Tensor fwd_grad_out, Tensor fwd_grad_grad_x, Tensor fwd_grad_grad_y, Tensor grad_x_grad, Tensor grad_y_grad, Tensor grad_grad_out_grad, bool transpose_x=false, bool transpose_y=false)
  output : Tensor(x_grad), Tensor(y_grad), Tensor(fwd_grad_out_grad), Tensor(fwd_grad_grad_x_grad), Tensor(fwd_grad_grad_y_grad)
  infer_meta :
    func : GeneralQuinaryGradInferMeta
    param : [x, y, fwd_grad_out, fwd_grad_grad_x, fwd_grad_grad_y]
  kernel :
    func : matmul_triple_grad
741
  optional : fwd_grad_grad_x, fwd_grad_grad_y, grad_x_grad, grad_y_grad, grad_grad_out_grad
Z
zyfncg 已提交
742

743
- backward_op : max_grad
744 745
  forward: max (Tensor x,  IntArray axis={},  bool keepdim=false) -> Tensor(out)
  args : (Tensor x, Tensor out, Tensor out_grad, IntArray axis={}, bool keepdim=false, bool reduce_all=false)
Z
zyfncg 已提交
746 747 748 749 750 751 752
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param: [x]
  kernel :
    func : max_grad

753
- backward_op : max_pool2d_with_index_grad
Z
zyfncg 已提交
754 755 756 757 758 759 760 761
  forward : max_pool2d_with_index(Tensor x, int[] kernel_size, int[] strides, int[] paddings, bool global_pooling, bool adaptive) -> Tensor(out), Tensor(mask)
  args : (Tensor x, Tensor mask, Tensor out_grad, int[] kernel_size, int[] strides, int[] paddings, bool global_pooling, bool adaptive)
  output : Tensor(x_grad)
  infer_meta :
    func : MaxPoolWithIndexGradInferMeta
  kernel :
    func : max_pool2d_with_index_grad

762
- backward_op : max_pool3d_with_index_grad
Z
zyfncg 已提交
763 764 765 766 767 768 769 770
  forward : max_pool3d_with_index(Tensor x, int[] kernel_size, int[] strides, int[] paddings, bool global_pooling, bool adaptive) -> Tensor(out), Tensor(mask)
  args : (Tensor x, Tensor mask, Tensor out_grad, int[] kernel_size, int[] strides, int[] paddings, bool global_pooling, bool adaptive)
  output : Tensor(x_grad)
  infer_meta :
    func : MaxPoolWithIndexGradInferMeta
  kernel :
    func : max_pool3d_with_index_grad

771
- backward_op : maximum_grad
Z
zyfncg 已提交
772 773 774 775 776 777 778 779 780
  forward : maximum(Tensor x, Tensor y) -> Tensor(out)
  args : (Tensor x, Tensor y, Tensor out_grad, int axis=-1)
  output : Tensor(x_grad), Tensor(y_grad)
  infer_meta :
    func : GeneralBinaryGradInferMeta
    param: [x, y]
  kernel :
    func : maximum_grad

781
- backward_op : mean_all_grad
Z
zyfncg 已提交
782 783 784 785 786 787 788 789 790
  forward : mean_all(Tensor x) -> Tensor(out)
  args : (Tensor x, Tensor out_grad)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param: [x]
  kernel :
    func : mean_all_grad

791
- backward_op : mean_double_grad
792 793
  forward: mean_grad (Tensor x, Tensor grad_out, IntArray axis={},  bool keepdim=false, bool reduce_all = false) -> Tensor(grad_x)
  args : (Tensor grad_x_grad, IntArray axis={},  bool keepdim=false)
Z
zyfncg 已提交
794
  output : Tensor(grad_out_grad)
795
  invoke : mean(grad_x_grad, axis, keepdim)
Z
zyfncg 已提交
796

797
- backward_op : mean_grad
798 799
  forward: mean (Tensor x,  IntArray axis={},  bool keepdim=false) -> Tensor(out)
  args : (Tensor x, Tensor out_grad, IntArray axis={},  bool keepdim=false, bool reduce_all=false)
Z
zyfncg 已提交
800 801 802 803 804 805 806 807 808
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param: [x]
  kernel :
    func : mean_grad
  backward : mean_double_grad
  no_need_buffer : x

809
- backward_op : min_grad
810 811
  forward: min (Tensor x,  IntArray axis={},  bool keepdim=false) -> Tensor(out)
  args : (Tensor x, Tensor out, Tensor out_grad, IntArray axis={}, bool keepdim=false, bool reduce_all=false)
Z
zyfncg 已提交
812 813 814 815 816 817 818
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param: [x]
  kernel :
    func : min_grad

819
- backward_op : minimum_grad
Z
zyfncg 已提交
820 821 822 823 824 825 826 827 828
  forward : minimum(Tensor x, Tensor y) -> Tensor(out)
  args : (Tensor x, Tensor y, Tensor out_grad, int axis=-1)
  output : Tensor(x_grad), Tensor(y_grad)
  infer_meta :
    func : GeneralBinaryGradInferMeta
    param: [x, y]
  kernel :
    func : minimum_grad

829
- backward_op : mish_grad
Z
zyfncg 已提交
830 831 832 833 834 835 836 837 838 839
  forward : mish (Tensor x, float threshold) -> Tensor(out)
  args : (Tensor x, Tensor out_grad, float threshold)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [x]
  kernel :
    func : mish_grad
  inplace : (out_grad -> x_grad)

840
- backward_op : multiply_double_grad
Z
zyfncg 已提交
841 842 843 844 845 846 847 848 849 850 851 852
  forward : multiply_grad (Tensor x, Tensor y, Tensor grad_out, int axis = -1) -> Tensor(grad_x), Tensor(grad_y)
  args : (Tensor x, Tensor y, Tensor grad_out, Tensor grad_x_grad, Tensor grad_y_grad, int axis = -1)
  output : Tensor(x_grad), Tensor(y_grad), Tensor(grad_out_grad)
  infer_meta :
    func : GeneralTernaryGradInferMeta
    param : [x, y, grad_out]
  kernel :
    func : multiply_double_grad
  optional : grad_x_grad, grad_y_grad
  backward : multiply_triple_grad
  inplace : (grad_x_grad -> grad_out_grad)

853
- backward_op : multiply_grad
Z
zyfncg 已提交
854 855 856 857 858 859 860 861 862 863
  forward : multiply (Tensor x, Tensor y) -> Tensor(out)
  args : (Tensor x, Tensor y, Tensor out_grad, int axis = -1)
  output : Tensor(x_grad), Tensor(y_grad)
  infer_meta :
    func : GeneralBinaryGradInferMeta
    param : [x, y]
  kernel :
    func : multiply_grad
  backward : multiply_double_grad

864
- backward_op : multiply_triple_grad
Z
zyfncg 已提交
865 866 867 868 869
  forward : multiply_double_grad (Tensor x, Tensor y, Tensor fwd_grad_out, Tensor fwd_grad_grad_x, Tensor fwd_grad_grad_y, int aixs = -1) -> Tensor(grad_x), Tensor(grad_y), Tensor(grad_grad_out)
  args : (Tensor x, Tensor y, Tensor fwd_grad_out, Tensor fwd_grad_grad_x, Tensor fwd_grad_grad_y, Tensor grad_x_grad, Tensor grad_y_grad, Tensor grad_grad_out_grad, int axis = -1)
  output : Tensor(x_grad), Tensor(y_grad), Tensor(fwd_grad_out_grad), Tensor(fwd_grad_grad_x_grad), Tensor(fwd_grad_grad_y_grad)
  infer_meta :
    func : GeneralQuinaryGradInferMeta
870
    param : [x, y, fwd_grad_out, fwd_grad_grad_x, fwd_grad_grad_y]
Z
zyfncg 已提交
871 872
  kernel :
    func : multiply_triple_grad
873
  optional : fwd_grad_grad_x, fwd_grad_grad_y, grad_x_grad, grad_y_grad, grad_grad_out_grad
Z
zyfncg 已提交
874

875
- backward_op : nearest_interp_grad
876 877 878 879 880 881 882 883 884 885 886
  forward : nearest_interp (Tensor x, Tensor out_size, Tensor[] size_tensor, Tensor scale_tensor, str data_layout, int out_d, int out_h, int out_w, float[] scale, str interp_method, bool align_corners, int align_mode) -> Tensor(output)
  args : (Tensor x, Tensor out_size, Tensor[] size_tensor, Tensor scale_tensor, Tensor output_grad, str data_layout, int out_d, int out_h, int out_w, float[] scale, str interp_method, bool align_corners, int align_mode)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param: [x]
  optional: out_size, size_tensor, scale_tensor
  kernel :
    func : nearest_interp_grad
    data_type : output_grad

887
- backward_op : norm_grad
Z
zyfncg 已提交
888 889 890 891 892 893 894 895 896
  forward : norm (Tensor x, int axis, float epsilon, bool is_test) -> Tensor(out), Tensor(norm)
  args : (Tensor x, Tensor norm, Tensor out_grad, int axis, float epsilon, bool is_test)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [x]
  kernel :
    func : norm_grad

897
- backward_op : p_norm_grad
Z
zyfncg 已提交
898 899 900 901 902 903 904 905 906
  forward : p_norm(Tensor x,  float porder,  int axis,  float epsilon,  bool keepdim,  bool asvector=false) -> Tensor(out)
  args : (Tensor x, Tensor out, Tensor out_grad,  float porder,  int axis,  float epsilon,  bool keepdim,  bool asvector)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param: [x]
  kernel :
    func : p_norm_grad

907
- backward_op : pad3d_double_grad
Z
zyfncg 已提交
908 909 910 911 912 913 914 915
  forward : pad3d_grad(Tensor x, Tensor grad_out, IntArray paddings, str mode, float pad_value, str data_format) -> Tensor(grad_x)
  args : (Tensor grad_x_grad, IntArray paddings, str mode, float pad_value, str data_format)
  output : Tensor(grad_out_grad)
  infer_meta :
    func : Pad3dInferMeta
  kernel :
    func : pad3d

916
- backward_op : pad3d_grad
Z
zyfncg 已提交
917 918 919 920 921 922 923 924 925 926 927
  forward : pad3d(Tensor x, IntArray paddings, str mode,  float pad_value, str data_format) -> Tensor(out)
  args : (Tensor x, Tensor out_grad, IntArray paddings, str mode,  float pad_value, str data_format)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param: [x]
  kernel :
    func : pad3d_grad
  no_need_buffer : x
  backward : pad3d_double_grad

928
- backward_op : pad_double_grad
929 930
  forward : pad_grad(Tensor x, Tensor grad_out, int[] paddings, Scalar pad_value) -> Tensor(grad_x)
  args : (Tensor grad_x_grad, int[] paddings, Scalar pad_value)
Z
zyfncg 已提交
931 932 933 934 935 936
  output : Tensor(grad_out_grad)
  infer_meta :
    func : PadInferMeta
  kernel :
    func : pad

937
- backward_op : pad_grad
938 939
  forward : pad(Tensor x, int[] paddings, Scalar pad_value) -> Tensor(out)
  args : (Tensor x, Tensor out_grad, int[] paddings, Scalar pad_value)
Z
zyfncg 已提交
940 941 942 943 944 945 946 947 948 949
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param: [x]
  kernel :
    func : pad_grad
    param: [out_grad, paddings, pad_value]
  no_need_buffer : x
  backward : pad_double_grad

950
- backward_op : pool2d_double_grad
951 952
  forward : pool2d_grad(Tensor x, Tensor out, Tensor grad_out, IntArray kernel_size, int[] strides, int[] paddings, bool ceil_mode, bool exclusive, str data_format, str pooling_type, bool global_pooling, bool adaptive, str padding_algorithm) -> Tensor(grad_x)
  args : (Tensor x, Tensor grad_x_grad, IntArray kernel_size, int[] strides, int[] paddings, bool ceil_mode, bool exclusive, str data_format, str pooling_type, bool global_pooling, bool adaptive, str padding_algorithm)
Z
zyfncg 已提交
953 954
  output : Tensor(grad_out_grad)
  infer_meta :
955
    func : Pool2DInferMeta
956
    param : [grad_x_grad, kernel_size, strides, paddings, ceil_mode, exclusive, data_format, pooling_type, global_pooling, adaptive, padding_algorithm]
Z
zyfncg 已提交
957 958
  kernel :
    func : pool2d_double_grad
959
    param : [grad_x_grad, kernel_size, strides, paddings, ceil_mode, exclusive, data_format, pooling_type, global_pooling, adaptive, padding_algorithm]
960
  no_need_buffer : x
Z
zyfncg 已提交
961

962
- backward_op : pool2d_grad
963 964
  forward : pool2d(Tensor x, IntArray kernel_size, int[] strides, int[] paddings, bool ceil_mode, bool exclusive, str data_format, str pooling_type, bool global_pooling, bool adaptive, str padding_algorithm) -> Tensor(out)
  args : (Tensor x, Tensor out, Tensor out_grad, IntArray kernel_size, int[] strides, int[] paddings, bool ceil_mode, bool exclusive, str data_format, str pooling_type, bool global_pooling, bool adaptive, str padding_algorithm)
Z
zyfncg 已提交
965 966
  output : Tensor(x_grad)
  infer_meta :
967 968
    func : UnchangedInferMeta
    param: [x]
Z
zyfncg 已提交
969 970
  kernel :
    func : pool2d_grad
971
    param : [x, out, out_grad, kernel_size, strides, paddings, ceil_mode, exclusive, data_format, pooling_type, global_pooling, adaptive, padding_algorithm]
Z
zyfncg 已提交
972 973
  backward : pool2d_double_grad

974
- backward_op : pool3d_grad
975 976
  forward : pool3d(Tensor x, int[] kernel_size, int[] strides, int[] paddings, bool ceil_mode, bool exclusive, str data_format, str pooling_type, bool global_pooling, bool adaptive, str padding_algorithm) -> Tensor(out)
  args : (Tensor x, Tensor out, Tensor out_grad, int[] kernel_size, int[] strides, int[] paddings, bool ceil_mode, bool exclusive, str data_format, str pooling_type, bool global_pooling, bool adaptive, str padding_algorithm)
Z
zyfncg 已提交
977 978
  output : Tensor(x_grad)
  infer_meta :
979 980
    func : UnchangedInferMeta
    param: [x]
Z
zyfncg 已提交
981 982
  kernel :
    func : pool3d_grad
983
    param : [x, out, out_grad, kernel_size, strides, paddings, ceil_mode, exclusive, data_format, pooling_type, global_pooling, adaptive, padding_algorithm]
Z
zyfncg 已提交
984

C
Charles-hit 已提交
985 986 987 988 989 990
- backward_op : pow_double_grad
  forward : pow_grad(Tensor x, Tensor grad_out, Scalar y) -> Tensor(grad_x)
  args : (Tensor x, Tensor grad_out, Tensor grad_x_grad, Scalar y)
  output : Tensor(x_grad), Tensor(grad_out_grad)
  infer_meta :
    func : GeneralBinaryGradInferMeta
C
Charles-hit 已提交
991
    param: [x, grad_out]
C
Charles-hit 已提交
992 993
  kernel :
    func : pow_double_grad
C
Charles-hit 已提交
994
  backward : pow_triple_grad
C
Charles-hit 已提交
995 996
  inplace : (grad_x_grad -> x_grad)

997
- backward_op : pow_grad
998 999
  forward : pow(Tensor x, Scalar y) -> Tensor(out)
  args : (Tensor x, Tensor out_grad, Scalar y=-1)
Z
zyfncg 已提交
1000 1001 1002 1003 1004 1005
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param: [x]
  kernel :
    func : pow_grad
C
Charles-hit 已提交
1006
  backward: pow_double_grad
Z
zyfncg 已提交
1007 1008
  inplace : (out_grad -> x_grad)

C
Charles-hit 已提交
1009 1010 1011 1012 1013 1014 1015 1016 1017 1018
- backward_op : pow_triple_grad
  forward : pow_double_grad(Tensor x, Tensor grad_out, Tensor grad_grad_x, Scalar y) -> Tensor(grad_x), Tensor(grad_grad_out)
  args : (Tensor x, Tensor grad_out, Tensor grad_grad_x, Tensor grad_x_grad, Tensor grad_grad_out_grad, Scalar y)
  output : Tensor(x_grad), Tensor(grad_out_grad), Tensor(grad_grad_x_grad)
  infer_meta :
    func : GeneralTernaryGradInferMeta
    param: [x, grad_out, grad_grad_x]
  kernel :
    func : pow_triple_grad

1019
- backward_op : prelu_grad
Z
zyfncg 已提交
1020 1021 1022 1023 1024 1025 1026 1027 1028
  forward : prelu(Tensor x, Tensor alpha, str data_format, str mode) -> Tensor(out)
  args : (Tensor x, Tensor alpha, Tensor out_grad, str data_format, str mode)
  output : Tensor(x_grad), Tensor(alpha_grad)
  infer_meta :
    func : GeneralBinaryGradInferMeta
    param: [x, alpha]
  kernel :
    func : prelu_grad

1029 1030 1031 1032 1033 1034 1035 1036 1037 1038
- backward_op : prod_grad
  forward : prod (Tensor x, IntArray dims, bool keep_dim, bool reduce_all) -> Tensor(out)
  args : (Tensor x, Tensor out, Tensor out_grad, IntArray dims,  bool keep_dim, bool reduce_all)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [x]
  kernel :
    func : prod_grad

1039
- backward_op : psroi_pool_grad
Z
zyfncg 已提交
1040 1041 1042 1043 1044 1045 1046 1047 1048 1049 1050
  forward : psroi_pool (Tensor x, Tensor boxes, Tensor boxes_num, int pooled_height, int pooled_width, int output_channels, float spatial_scale) -> Tensor(out)
  args : (Tensor x, Tensor boxes, Tensor boxes_num, Tensor out_grad, int pooled_height, int pooled_width, int output_channels, float spatial_scale)
  output : Tensor(x_grad)
  infer_meta :
    func : GeneralUnaryGradInferMeta
    param : [x]
  kernel :
    func : psroi_pool_grad
    data_type : x
  optional : boxes_num

1051
- backward_op : relu6_grad
1052 1053
  forward : relu6 (Tensor x) -> Tensor(out)
  args : (Tensor out, Tensor out_grad, float threshold = 6)
1054 1055 1056 1057 1058 1059 1060 1061
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [out]
  kernel :
    func : relu6_grad
  inplace : (out_grad -> x_grad)

1062
- backward_op : repeat_interleave_grad
1063 1064
  forward : repeat_interleave(Tensor x, int repeats, int axis) -> Tensor(out)
  args : (Tensor x, Tensor out_grad, int repeats, int axis)
S
seemingwang 已提交
1065 1066 1067 1068 1069 1070 1071
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [x]
  kernel :
    func : repeat_interleave_grad

1072
- backward_op : repeat_interleave_with_tensor_index_grad
1073 1074
  forward : repeat_interleave_with_tensor_index(Tensor x, Tensor repeats, int axis) -> Tensor(out)
  args : (Tensor x, Tensor repeats, Tensor out_grad, int axis)
S
seemingwang 已提交
1075 1076 1077 1078 1079 1080 1081 1082
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [x]
  kernel :
    func : repeat_interleave_with_tensor_index_grad
    data_type : x

1083
- backward_op : reshape_double_grad
Z
zyfncg 已提交
1084 1085 1086 1087 1088 1089 1090 1091 1092 1093 1094
  forward : reshape_grad (Tensor xshape, Tensor grad_out) -> Tensor(grad_x)
  args : (Tensor grad_out, Tensor grad_x_grad)
  output : Tensor(grad_out_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [grad_out]
  kernel :
    func : reshape_double_grad
  no_need_buffer : grad_out
  inplace : (grad_x_grad -> grad_out_grad)

1095
- backward_op : reshape_grad
Z
zyfncg 已提交
1096 1097 1098 1099 1100 1101 1102 1103 1104 1105 1106 1107 1108 1109 1110
  forward : reshape (Tensor x, IntArray shape) -> Tensor(out), Tensor(xshape)
  args : (Tensor xshape, Tensor out_grad)
  output : Tensor(x_grad)
  infer_meta :
    func : KernelWithXShapeInferMeta
    param : [xshape]
  kernel :
    func : reshape_grad
    param : [out_grad]
    data_type: out_grad
    backend: out_grad
    layout: out_grad
  backward : reshape_double_grad
  inplace : (out_grad -> x_grad)

1111
- backward_op : reverse_grad
1112 1113
  forward : reverse (Tensor x, IntArray axis) -> Tensor(out)
  args : (Tensor out_grad, IntArray axis)
W
wanghuancoder 已提交
1114 1115 1116
  output : Tensor(x_grad)
  invoke : reverse(out_grad, axis)

Y
YuanRisheng 已提交
1117 1118 1119 1120 1121 1122 1123 1124 1125 1126 1127 1128
- backward_op : rnn_grad
  forward : rnn (Tensor x, Tensor[] pre_state, Tensor[] weight_list, Tensor sequence_length, Tensor dropout_state_in, float dropout_prob, bool is_bidirec, int input_size, int hidden_size, int num_layers, str mode, int seed, bool is_test) -> Tensor(out), Tensor(dropout_state_out), Tensor[](state), Tensor(reserve)
  args : (Tensor x, Tensor[] pre_state, Tensor[] weight_list, Tensor sequence_length, Tensor out, Tensor dropout_state_out, Tensor reserve, Tensor out_grad, Tensor[] state_grad, float dropout_prob, bool is_bidirec, int input_size, int hidden_size, int num_layers, str mode, int seed, bool is_test)
  output : Tensor(x_grad), Tensor[](pre_state_grad){pre_state.size()}, Tensor[](weight_list_grad){weight_list.size()}
  infer_meta :
    func : RnnGradInferMeta
    param : [x, pre_state, weight_list]
  kernel :
    func : rnn_grad
    data_type: out_grad
  optional : sequence_length

1129
- backward_op : roi_align_grad
Z
zyfncg 已提交
1130 1131 1132 1133 1134 1135 1136 1137 1138 1139 1140 1141
  forward : roi_align (Tensor x, Tensor boxes, Tensor boxes_num, int pooled_height, int pooled_width, float spatial_scale, int sampling_ratio, bool aligned) -> Tensor(out)
  args : (Tensor x, Tensor boxes, Tensor boxes_num, Tensor out_grad, int pooled_height, int pooled_width, float spatial_scale, int sampling_ratio, bool aligned)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [x]
  kernel :
    func : roi_align_grad
    data_type : boxes
  no_need_buffer : x
  optional : boxes_num

1142
- backward_op : roi_pool_grad
Z
zyfncg 已提交
1143 1144 1145 1146 1147 1148 1149 1150 1151 1152 1153
  forward : roi_pool (Tensor x, Tensor boxes, Tensor boxes_num, int pooled_height, int pooled_width, float spatial_scale) -> Tensor(out), Tensor(arg_max)
  args : (Tensor x, Tensor boxes, Tensor boxes_num, Tensor arg_max, Tensor out_grad, int pooled_height, int pooled_width, float spatial_scale)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [x]
  kernel :
    func : roi_pool_grad
    data_type : x
  optional : boxes_num

1154
- backward_op : scale_grad
Z
zyfncg 已提交
1155
  forward : scale (Tensor x, Scalar scale, float bias, bool bias_after_scale) -> Tensor(out)
1156
  args : (Tensor out_grad, Scalar scale=1.0, bool bias_after_scale=true)
Z
zyfncg 已提交
1157 1158 1159
  output : Tensor(x_grad)
  invoke : scale(out_grad, scale, 0.0, bias_after_scale)

1160
- backward_op : segment_pool_grad
Z
zyfncg 已提交
1161 1162 1163 1164 1165 1166 1167 1168 1169 1170 1171
  forward : segment_pool (Tensor x, Tensor segment_ids, str pooltype) -> Tensor(out), Tensor(summed_ids)
  args : (Tensor x, Tensor segment_ids, Tensor out, Tensor summed_ids, Tensor out_grad, str pooltype)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [x]
  kernel :
    func : segment_pool_grad
    data_type : x
  optional : summed_ids

1172 1173 1174 1175 1176 1177 1178 1179 1180 1181 1182 1183 1184 1185 1186 1187 1188 1189 1190 1191 1192 1193 1194 1195
- backward_op : send_u_recv_grad
  forward : send_u_recv (Tensor x, Tensor src_index, Tensor dst_index, str reduce_op = "SUM", IntArray out_size = {0}) -> Tensor(out), Tensor(dst_count)
  args : (Tensor x, Tensor src_index, Tensor dst_index, Tensor out, Tensor dst_count, Tensor out_grad, str reduce_op = "SUM")
  output : Tensor(x_grad)
  infer_meta :
    func : GeneralUnaryGradInferMeta
    param : [x]
  kernel :
    func : send_u_recv_grad
    data_type : out_grad
  optional: out, dst_count

- backward_op : send_ue_recv_grad
  forward : send_ue_recv (Tensor x, Tensor y, Tensor src_index, Tensor dst_index, str message_op, str reduce_op, IntArray out_size) -> Tensor(out), Tensor(dst_count)
  args : (Tensor x, Tensor y, Tensor src_index, Tensor dst_index, Tensor out, Tensor dst_count, Tensor out_grad, str message_op, str reduce_op)
  output : Tensor(x_grad), Tensor(y_grad)
  infer_meta :
    func : GeneralBinaryGradInferMeta
    param : [x, y]
  kernel :
    func : send_ue_recv_grad
    data_type : out_grad
  optional: out, dst_count

1196
- backward_op : sigmoid_cross_entropy_with_logits_grad
Z
zyfncg 已提交
1197 1198 1199 1200 1201 1202 1203 1204 1205 1206
  forward : sigmoid_cross_entropy_with_logits (Tensor x, Tensor label, bool normalize, int ignore_index) -> Tensor(out)
  args : (Tensor x, Tensor label, Tensor out_grad, bool normalize, int ignore_index)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [x]
  kernel :
    func : sigmoid_cross_entropy_with_logits_grad
  inplace : (out_grad -> x_grad)

1207 1208 1209 1210 1211 1212
- backward_op : sign_grad
  forward : sign (Tensor x) -> Tensor(out)
  args : (Tensor out_grad)
  output : Tensor(x_grad)
  invoke : scale(out_grad, 0.0, 0.0, true)

1213
- backward_op : slice_double_grad
1214 1215 1216
  forward : slice_grad (Tensor input, Tensor grad_out, int64_t[] axes, IntArray starts, IntArray ends, int64_t[] infer_flags, int64_t[] decrease_axis) -> Tensor(grad_input)
  args : (Tensor grad_input_grad, int64_t[] axes, IntArray starts, IntArray ends, int64_t[] infer_flags, int64_t[] decrease_axis)
  output : Tensor(grad_out_grad)
1217
  invoke : slice(grad_input_grad, axes, starts, ends, infer_flags, decrease_axis)
1218

1219
- backward_op : slice_grad
Z
zyfncg 已提交
1220 1221 1222 1223 1224 1225 1226 1227
  forward : slice (Tensor input, int64_t[] axes, IntArray starts, IntArray ends, int64_t[] infer_flags, int64_t[] decrease_axis) -> Tensor(out)
  args : (Tensor input, Tensor out_grad, int64_t[] axes, IntArray starts, IntArray ends, int64_t[] infer_flags, int64_t[] decrease_axis)
  output : Tensor(input_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [input]
  kernel :
    func : slice_grad
1228
  backward : slice_double_grad
Z
zyfncg 已提交
1229 1230
  no_need_buffer : input

1231
- backward_op : slogdet_grad
1232 1233 1234 1235 1236 1237 1238
  forward : slogdet (Tensor x) -> Tensor(out)
  args : (Tensor x, Tensor out, Tensor out_grad)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [x]
  kernel :
1239
    func : slogdet_grad
1240

1241
- backward_op : softmax_grad
Z
zyfncg 已提交
1242 1243 1244 1245 1246 1247 1248 1249 1250
  forward : softmax (Tensor x, int axis) -> Tensor(out)
  args : (Tensor out, Tensor out_grad, int axis)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [out]
  kernel :
    func : softmax_grad

1251
- backward_op : spectral_norm_grad
1252 1253 1254 1255 1256 1257 1258 1259 1260
  forward : spectral_norm (Tensor weight, Tensor u, Tensor v, int dim, int power_iters, float eps) -> Tensor(out)
  args : (Tensor weight, Tensor u, Tensor v, Tensor out_grad, int dim, int power_iters, float eps)
  output : Tensor(weight_grad)
  infer_meta :
    func : SpectralNormGradInferMeta
  kernel :
    func : spectral_norm_grad
    data_type : out_grad

1261
- backward_op : split_grad
Z
zyfncg 已提交
1262 1263 1264 1265
  forward : split (Tensor x, IntArray num_or_sections, Scalar axis) -> Tensor[](out)
  args : (Tensor[] out_grad, Scalar axis = -1)
  output : Tensor(x_grad)
  invoke : concat( out_grad, axis)
C
Charles-hit 已提交
1266

1267
- backward_op : split_with_num_grad
C
Charles-hit 已提交
1268 1269 1270 1271
  forward : split_with_num (Tensor x, int num, Scalar axis) -> Tensor[](out)
  args : (Tensor[] out_grad, Scalar axis = -1)
  output : Tensor(x_grad)
  invoke : concat( out_grad, axis)
Z
zyfncg 已提交
1272

1273
- backward_op : squared_l2_norm_grad
1274 1275 1276 1277 1278 1279 1280 1281 1282
  forward : squared_l2_norm(Tensor x) -> Tensor(out)
  args : (Tensor x, Tensor out_grad)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param: [x]
  kernel :
    func : squared_l2_norm_grad

1283
- backward_op : stack_grad
Z
zyfncg 已提交
1284 1285 1286 1287 1288 1289 1290 1291 1292 1293 1294
  forward : stack (Tensor[] x, int axis) -> Tensor(out)
  args : (Tensor[] x, Tensor out_grad, int axis)
  output : Tensor[](x_grad){x.size()}
  infer_meta :
    func : StackGradInferMeta
    param: [out_grad, axis]
  kernel :
    func : stack_grad
    param : [out_grad, axis]
  no_need_buffer : x

1295
- backward_op : strided_slice_grad
Z
zyfncg 已提交
1296 1297 1298 1299 1300 1301 1302 1303 1304 1305
  forward : strided_slice (Tensor x, int[] axes, IntArray starts, IntArray ends, IntArray strides) -> Tensor(out)
  args : (Tensor x, Tensor out_grad, int[] axes, IntArray starts, IntArray ends, IntArray strides)
  output : Tensor(x_grad)
  infer_meta :
    func : GeneralUnaryGradInferMeta
    param : [x]
  kernel :
    func : strided_slice_grad
  no_need_buffer : x

1306
- backward_op : subtract_double_grad
Z
zyfncg 已提交
1307 1308 1309 1310 1311 1312 1313 1314 1315 1316 1317 1318
  forward : subtract_grad (Tensor x, Tensor y, Tensor grad_out, int axis = -1) -> Tensor(grad_x), Tensor(grad_y)
  args : (Tensor y, Tensor grad_out, Tensor grad_x_grad, Tensor grad_y_grad, int axis = -1)
  output : Tensor(grad_out_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [grad_out]
  kernel :
    func : subtract_double_grad
  optional : grad_x_grad, grad_y_grad
  no_need_buffer : y, grad_out
  inplace : (grad_x_grad -> grad_out_grad)

1319
- backward_op : subtract_grad
Z
zyfncg 已提交
1320 1321 1322 1323 1324 1325 1326 1327 1328 1329 1330 1331
  forward : subtract (Tensor x, Tensor y) -> Tensor(out)
  args : (Tensor x, Tensor y, Tensor out_grad, int axis = -1)
  output : Tensor(x_grad), Tensor(y_grad)
  infer_meta :
    func : GeneralBinaryGradInferMeta
    param : [x, y]
  kernel :
    func : subtract_grad
  no_need_buffer : x, y
  backward : subtract_double_grad
  inplace : (out_grad -> x_grad)

1332
- backward_op : sum_double_grad
1333 1334
  forward : sum_grad (Tensor x, Tensor grad_out, IntArray axis, bool keepdim, bool reduce_all=false) -> Tensor(grad_x)
  args : (Tensor grad_x_grad, IntArray axis={}, bool keepdim=false)
Z
zyfncg 已提交
1335
  output : Tensor(grad_out_grad)
1336
  invoke : sum(grad_x_grad, axis, grad_x_grad.dtype(), keepdim)
Z
zyfncg 已提交
1337

1338
- backward_op : sum_grad
1339 1340
  forward : sum (Tensor x, IntArray axis={}, DataType dtype=DataType::UNDEFINED, bool keepdim=false) -> Tensor(out)
  args : (Tensor x, Tensor out_grad, IntArray axis, bool keepdim, bool reduce_all=false)
Z
zyfncg 已提交
1341 1342 1343 1344 1345 1346 1347 1348 1349
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [x]
  kernel :
    func : sum_grad
  no_need_buffer : x
  backward : sum_double_grad

1350
- backward_op : swish_grad
1351
  forward : swish (Tensor x) -> Tensor(out)
Z
zyfncg 已提交
1352 1353 1354 1355 1356 1357 1358 1359 1360
  args : (Tensor x, Tensor out_grad, float bete=1.0)
  output : Tensor(x_grad)
  infer_meta :
    func : GeneralUnaryGradInferMeta
    param : [x]
  kernel :
    func : swish_grad
  inplace : (out_grad -> x_grad)

1361
- backward_op : sync_batch_norm_grad
1362 1363
  forward : sync_batch_norm_ (Tensor x, Tensor mean, Tensor variance, Tensor scale, Tensor bias, bool is_test, float momentum, float epsilon, str data_layout, bool use_global_stats, bool trainable_statistics) -> Tensor(out), Tensor(mean_out), Tensor(variance_out), Tensor(saved_mean), Tensor(saved_variance), Tensor(reserve_space)
  args : (Tensor x, Tensor scale, Tensor bias, Tensor saved_mean, Tensor saved_variance, Tensor reserve_space, Tensor out_grad, float momentum, float epsilon, str data_layout, bool is_test, bool use_global_stats, bool trainable_statistics)
1364 1365 1366 1367 1368 1369 1370
  output : Tensor(x_grad), Tensor(scale_grad), Tensor(bias_grad)
  infer_meta :
    func : GeneralTernaryGradInferMeta
    param : [x, scale, bias]
  kernel :
    func : sync_batch_norm_grad
    data_type : out_grad
1371
  optional : reserve_space
1372

1373
- backward_op : temporal_shift_grad
C
ccrrong 已提交
1374 1375 1376 1377 1378 1379 1380 1381 1382
  forward : temporal_shift(Tensor x, int seg_num, float shift_ratio, str data_format_str) -> Tensor(out)
  args : (Tensor out_grad, int seg_num, float shift_ratio, str data_format_str)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [out_grad]
  kernel :
    func : temporal_shift_grad

1383
- backward_op : tile_double_grad
Z
zyfncg 已提交
1384 1385 1386
  forward : tile_grad (Tensor x, Tensor grad_out, IntArray repeat_times) -> Tensor(grad_x)
  args : (Tensor grad_x_grad, IntArray repeat_times)
  output : Tensor(grad_out_grad)
1387
  invoke : tile(grad_x_grad, repeat_times)
Z
zyfncg 已提交
1388

1389
- backward_op : tile_grad
Z
zyfncg 已提交
1390 1391 1392 1393 1394 1395 1396 1397 1398 1399 1400
  forward : tile (Tensor x, IntArray repeat_times) -> Tensor(out)
  args : (Tensor x, Tensor out_grad, IntArray repeat_times)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [x]
  kernel :
    func : tile_grad
  no_need_buffer : x
  backward : tile_double_grad

1401
- backward_op : transpose_double_grad
1402 1403
  forward : transpose_grad (Tensor grad_out, int[] perm) -> Tensor(grad_x)
  args : (Tensor grad_x_grad, int[] perm)
Z
zyfncg 已提交
1404
  output : Tensor(grad_out_grad)
1405
  invoke : transpose(grad_x_grad, perm)
Z
zyfncg 已提交
1406

1407
- backward_op : transpose_grad
1408 1409
  forward : transpose (Tensor x, int[] perm) -> Tensor(out)
  args : (Tensor out_grad, int[] perm)
Z
zyfncg 已提交
1410 1411 1412
  output : Tensor(x_grad)
  infer_meta :
    func : TransposeGradInferMeta
1413
    param : [out_grad, perm]
Z
zyfncg 已提交
1414 1415 1416 1417
  kernel :
    func : transpose_grad
  backward : transpose_double_grad

1418
- backward_op : triangular_solve_grad
Z
zyfncg 已提交
1419 1420 1421 1422 1423 1424 1425 1426 1427
  forward : triangular_solve (Tensor x, Tensor y, bool upper, bool tranpose, bool unitriangular) -> Tensor(out)
  args : (Tensor x, Tensor y, Tensor out, Tensor out_grad, bool upper, bool tranpose, bool unitriangular)
  output : Tensor(x_grad), Tensor(y_grad)
  infer_meta :
    func : GeneralBinaryGradInferMeta
    param : [x, y]
  kernel :
    func : triangular_solve_grad

1428
- backward_op : tril_grad
1429 1430
  forward : tril(Tensor x,  int diagonal) -> Tensor(out)
  args : (Tensor out_grad,  int diagonal)
Z
zyfncg 已提交
1431 1432 1433 1434 1435
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [out_grad]
  kernel :
1436
    func : tril_grad
Z
zyfncg 已提交
1437

1438
- backward_op : trilinear_interp_grad
1439 1440 1441 1442 1443 1444 1445 1446 1447 1448 1449
  forward : trilinear_interp (Tensor x, Tensor out_size, Tensor[] size_tensor, Tensor scale_tensor, str data_layout, int out_d, int out_h, int out_w, float[] scale, str interp_method, bool align_corners, int align_mode) -> Tensor(output)
  args : (Tensor x, Tensor out_size, Tensor[] size_tensor, Tensor scale_tensor, Tensor output_grad, str data_layout, int out_d, int out_h, int out_w, float[] scale, str interp_method, bool align_corners, int align_mode)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param: [x]
  optional: out_size, size_tensor, scale_tensor
  kernel :
    func : trilinear_interp_grad
    data_type : output_grad

1450 1451 1452 1453 1454 1455 1456 1457 1458 1459
- backward_op : triu_grad
  forward : triu(Tensor x,  int diagonal) -> Tensor(out)
  args : (Tensor out_grad,  int diagonal)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [out_grad]
  kernel :
    func : triu_grad

1460
- backward_op : unbind_grad
Z
zyfncg 已提交
1461 1462 1463 1464 1465
  forward : unbind (Tensor input, int axis) -> Tensor[](out)
  args : (Tensor[] out_grad, int axis)
  output : Tensor(input_grad)
  invoke : stack(out_grad, axis)

1466 1467
- backward_op : uniform_inplace_grad
  forward : uniform_inplace(Tensor x, float min, float max, int seed, int diag_num, int diag_step, float diag_val) -> Tensor(out)
1468 1469 1470 1471 1472
  args : (Tensor out_grad, float min, float max, int seed, int diag_num, int diag_step, float diag_val)
  output : Tensor(x_grad)
  infer_meta :
    func : UniformRandomInplaceGradInferMeta
  kernel :
1473
    func : uniform_inplace_grad
1474 1475
  inplace : (out_grad -> x_grad)

1476
- backward_op : warpctc_grad
1477
  forward : warpctc (Tensor logits, Tensor label, Tensor logits_length, Tensor labels_length, int blank, bool norm_by_times) -> Tensor(loss), Tensor(warpctcgrad)
Z
Zhong Hui 已提交
1478 1479 1480 1481 1482 1483 1484 1485 1486 1487
  args : (Tensor logits, Tensor logits_length, Tensor warpctcgrad, Tensor loss_grad, int blank, bool norm_by_times)
  output : Tensor(logits_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [logits]
  kernel :
    func : warpctc_grad
  optional : logits_length
  no_need_buffer : logits

1488 1489
- backward_op : yolo_loss_grad
  forward : yolo_loss(Tensor x, Tensor gt_box, Tensor gt_label, Tensor gt_score, int[] anchors, int[] anchor_mask, int class_num, float ignore_thresh, int downsample_ratio, bool use_label_smooth=true, float scale_x_y=1.0) -> Tensor(loss), Tensor(objectness_mask), Tensor(gt_match_mask)
1490 1491 1492
  args : (Tensor x, Tensor gt_box, Tensor gt_label, Tensor gt_score, Tensor objectness_mask, Tensor gt_match_mask, Tensor loss_grad, int[] anchors, int[] anchor_mask, int class_num, float ignore_thresh, int downsample_ratio, bool use_label_smooth=true, float scale_x_y=1.0)
  output : Tensor(x_grad), Tensor(gt_box_grad), Tensor(gt_label_grad), Tensor(gt_score_grad)
  infer_meta :
1493
    func : YoloLossGradInferMeta
1494
  kernel :
1495
    func : yolo_loss_grad
1496
  optional : gt_score
X
xiaoting 已提交
1497

1498
- backward_op: unpool3d_grad
X
xiaoting 已提交
1499 1500 1501 1502 1503 1504 1505 1506 1507 1508
  forward: unpool3d (Tensor x, Tensor indices, int[] ksize, int[] strides, int[] padding, int[] output_size, str data_format) -> Tensor(out)
  args: (Tensor x, Tensor indices, Tensor out, Tensor out_grad, int[] ksize, int[] strides, int[] padding, int[] output_size, str data_format)
  output: Tensor(x_grad)
  infer_meta:
    func: UnchangedInferMeta
    param : [x]
  kernel:
    func: unpool3d_grad
    data_type: x

1509
- backward_op: unpool_grad
1510 1511
  forward: unpool (Tensor x, Tensor indices, int[] ksize, int[] strides, int[] padding,  IntArray output_size, str data_format) -> Tensor(out)
  args: (Tensor x, Tensor indices, Tensor out, Tensor out_grad, int[] ksize, int[] strides, int[] padding, IntArray output_size, str data_format)
X
xiaoting 已提交
1512 1513 1514 1515 1516 1517 1518
  output: Tensor(x_grad)
  infer_meta:
    func: UnchangedInferMeta
    param : [x]
  kernel:
    func: unpool_grad
    data_type: x