legacy_backward.yaml 78.3 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 : addmm_grad
60
  forward : addmm (Tensor input, Tensor x, Tensor y, float beta, float alpha) -> Tensor(out)
Z
zyfncg 已提交
61 62 63 64 65 66 67 68
  args : (Tensor input, Tensor x, Tensor y, Tensor out_grad, float alpha, float beta)
  output : Tensor(input_grad), Tensor(x_grad), Tensor(y_grad)
  infer_meta :
    func : GeneralTernaryGradInferMeta
    param : [input, x, y]
  kernel :
    func : addmm_grad

69
- backward_op : affine_grid_grad
70 71 72 73 74 75 76 77 78 79 80
  forward : affine_grid (Tensor input, IntArray outputShape, bool use_cudnn=true, bool align_corners=true) -> Tensor(output)
  args : (Tensor output_grad, IntArray outputShape, bool use_cudnn=true, bool align_corners=true)
  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]
    use_gpudnn: use_cudnn

81
- backward_op : amax_grad
82 83
  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)
84 85 86 87 88 89 90
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param: [x]
  kernel :
    func : amax_grad

91
- backward_op : amin_grad
92 93
  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)
94 95 96 97 98 99 100
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param: [x]
  kernel :
    func : amin_grad

101
- backward_op : assign_grad
Z
zyfncg 已提交
102 103 104
  forward : assign (Tensor x) -> Tensor(out)
  args : (Tensor out_grad)
  output : Tensor(x_grad)
105
  invoke : assign(out_grad)
Z
zyfncg 已提交
106

107
- backward_op : assign_out__grad
Z
zyfncg 已提交
108 109 110 111 112 113 114 115 116
  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)

117
- backward_op : batch_norm_double_grad
118 119
  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
120 121 122 123 124 125 126 127 128 129
  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
  optional : out_mean, out_variance
  inplace : (grad_out -> grad_out_grad)

130
- backward_op : batch_norm_grad
131 132
  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
133 134 135 136 137 138 139 140 141 142
  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

143
- backward_op : bce_loss_grad
Z
zyfncg 已提交
144 145 146 147 148 149 150 151 152 153
  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)

154
- backward_op : bicubic_interp_grad
155 156 157 158 159 160 161 162 163 164 165
  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

166
- backward_op : bilinear_interp_grad
167 168 169 170 171 172 173 174 175 176 177
  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

178
- backward_op : bilinear_tensor_product_grad
179 180 181 182 183 184 185 186
  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

187
- backward_op : broadcast_tensors_grad
188 189 190
  forward : broadcast_tensors (Tensor[] input) -> Tensor[](out)
  args : (Tensor[] input, Tensor[] out_grad)
  output : Tensor[](input_grad)
191 192
  infer_meta :
    func : UnchangedMultiInferMeta
193
    param : [input]
194 195 196
  kernel :
    func : broadcast_tensors_grad
    param : [out_grad]
197
  no_need_buffer : input
198

199
- backward_op : cast_grad
200
  forward : cast (Tensor x, DataType dtype) -> Tensor(out)
Z
zyfncg 已提交
201 202
  args : (Tensor x, Tensor out_grad)
  output : Tensor(x_grad)
203
  invoke : cast (out_grad, x.dtype())
Z
zyfncg 已提交
204 205
  no_need_buffer : x

206
- backward_op : clip_double_grad
Z
zyfncg 已提交
207 208 209 210 211 212 213 214 215
  forward : clip_grad (Tensor x, Tensor grad_out, Scalar min = 0., Scalar max = 0.) -> Tensor(grad_x)
  args : (Tensor x, Tensor grad_x_grad, Scalar min = 0., Scalar max = 0.)
  output : Tensor(grad_out_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [x]
  kernel :
    func : clip_grad

216
- backward_op : clip_grad
Z
zyfncg 已提交
217 218 219 220 221 222 223 224 225 226 227
  forward : clip (Tensor x, Scalar min, Scalar max) -> Tensor(out)
  args : (Tensor x, Tensor out_grad, Scalar min = 0., Scalar max = 0.)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [x]
  kernel :
    func : clip_grad
  backward : clip_double_grad
  inplace : (out_grad -> x_grad)

228
- backward_op : complex_grad
229 230 231
  forward : complex (Tensor real, Tensor imag) -> Tensor(out)
  args : (Tensor real, Tensor imag, Tensor out_grad)
  output : Tensor(real_grad), Tensor(imag_grad)
232 233 234 235
  infer_meta :
    func : ComplexGradInferMeta
  kernel :
    func : complex_grad
236
    data_type : real
237

238
- backward_op : concat_double_grad
Z
zyfncg 已提交
239 240 241
  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)
242
  invoke : concat(grad_x_grad, axis)
Z
zyfncg 已提交
243

244
- backward_op : concat_grad
Z
zyfncg 已提交
245 246 247 248 249 250 251 252 253 254 255
  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

256
- backward_op : conj_grad
Z
zyfncg 已提交
257 258 259 260 261 262 263 264 265
  forward : conj (Tensor x) -> Tensor(out)
  args : (Tensor out_grad)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param: [out_grad]
  kernel :
    func : conj

266
- backward_op : conv2d_grad
267 268
  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
269
  output : Tensor(input_grad), Tensor(filter_grad)
Z
zyfncg 已提交
270 271 272 273 274 275
  infer_meta :
    func : GeneralBinaryGradInferMeta
    param : [input, filter]
  kernel :
    func : conv2d_grad
    use_gpudnn : true
Z
zyfncg 已提交
276 277
  backward : conv2d_grad_grad

278
- backward_op : conv2d_grad_grad
279 280
  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
281 282 283 284 285 286 287 288 289
  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
    use_gpudnn : true
  optional : grad_input_grad, grad_filter_grad

290
- backward_op : conv2d_transpose_double_grad
291 292
  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
293 294 295 296 297 298 299
  output : Tensor(x_grad), Tensor(filter_grad), Tensor(grad_out_grad)
  infer_meta :
    func : Conv2dTransposeDoubleGradInferMeta
  kernel :
    func : conv2d_transpose_grad_grad
    use_gpudnn : true

300
- backward_op : conv2d_transpose_grad
301 302
  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
303 304
  output : Tensor(x_grad), Tensor(filter_grad)
  infer_meta :
305
    func : Conv2dTransposeGradInferMeta
Z
zyfncg 已提交
306 307 308 309 310
  kernel :
    func : conv2d_transpose_grad
    use_gpudnn : true
  backward : conv2d_transpose_double_grad

311 312 313 314 315 316 317 318 319 320 321 322
- 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
    use_gpudnn : true
  optional : grad_input_grad, grad_filter_grad

323
- backward_op : conv3d_grad
324 325
  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
326
  output : Tensor(input_grad), Tensor(filter_grad)
Z
zyfncg 已提交
327 328 329 330 331 332
  infer_meta :
    func : GeneralBinaryGradInferMeta
    param : [input, filter]
  kernel :
    func : conv3d_grad
    use_gpudnn : true
333
  backward : conv3d_double_grad
Z
zyfncg 已提交
334

335
- backward_op : conv3d_transpose_grad
Z
zyfncg 已提交
336 337 338 339 340 341 342 343 344
  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
    use_gpudnn : true

345
- backward_op : crop_grad
346 347 348 349
  forward : crop_tensor (Tensor x, IntArray shape, IntArray offsets) -> Tensor(out)
  args : (Tensor x, Tensor out_grad, IntArray offsets)
  output : Tensor(x_grad)
  infer_meta :
350
    func : CropGradInferMeta
351
  kernel :
352
    func : crop_grad
353 354
    data_type : x

355
- backward_op : cross_entropy_with_softmax_grad
Z
zyfncg 已提交
356 357 358 359 360 361 362 363 364 365
  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)

366
- backward_op : cumprod_grad
Z
zyfncg 已提交
367 368 369 370 371 372 373 374 375
  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

376
- backward_op : cumsum_grad
W
WangZhen 已提交
377 378
  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
379 380 381
  output : Tensor(x_grad)
  invoke : cumsum(out_grad, axis, flatten, exclusive, !reverse)

382
- backward_op : deformable_conv_grad
Z
zyfncg 已提交
383 384 385 386 387 388 389 390 391 392
  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

393 394 395 396 397 398 399 400 401 402 403
- backward_op : depthwise_conv2d_double_grad
  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, bool use_gpudnn) -> 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 : depthwise_conv2d_double_grad
  optional : grad_input_grad, grad_filter_grad

404
- backward_op : depthwise_conv2d_grad
405 406
  forward : depthwise_conv2d (Tensor input, Tensor filter, int[] strides, int[] paddings, str padding_algorithm, int groups, int[] dilations, str data_format, bool use_gpudnn) -> Tensor(out)
  args : (Tensor input, Tensor filter, Tensor out_grad, int[] strides, int[] paddings, str padding_algorithm, int groups, int[] dilations, str data_format, bool use_gpudnn)
Z
zyfncg 已提交
407 408 409 410 411 412
  output : Tensor(input_grad), Tensor(filter_grad)
  infer_meta :
    func : GeneralBinaryGradInferMeta
    param : [input, filter]
  kernel :
    func : depthwise_conv2d_grad
413
    param : [input, filter, out_grad, strides, paddings, padding_algorithm, groups, dilations, data_format]
Z
zyfncg 已提交
414
    use_gpudnn : use_gpudnn
415
  backward : depthwise_conv2d_double_grad
Z
zyfncg 已提交
416

417
- backward_op : depthwise_conv2d_transpose_grad
418 419
  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
420 421
  output : Tensor(x_grad), Tensor(filter_grad)
  infer_meta :
422
    func : Conv2dTransposeGradInferMeta
Z
zyfncg 已提交
423 424 425
  kernel :
    func : depthwise_conv2d_transpose_grad

426
- backward_op : divide_double_grad
Z
zyfncg 已提交
427 428 429 430 431 432 433 434 435 436 437 438
  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)

439
- backward_op : divide_grad
Z
zyfncg 已提交
440 441 442 443 444 445 446 447 448 449
  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

450
- backward_op : dropout_grad
451 452
  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
453 454 455 456 457 458 459
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [out_grad]
  kernel :
    func : dropout_grad

460
- backward_op : eigvalsh_grad
461 462 463 464 465 466 467 468 469 470 471
  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

472
- backward_op : einsum_grad
Z
zyfncg 已提交
473 474 475 476 477 478 479 480 481
  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

482
- backward_op : elementwise_pow_grad
Z
zyfncg 已提交
483 484 485 486 487 488 489 490 491
  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

492
- backward_op : embedding_grad
Z
zyfncg 已提交
493 494 495 496 497
  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)

498
- backward_op : expand_as_grad
Z
zyfncg 已提交
499 500 501 502 503 504 505 506 507 508
  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

509
- backward_op : expand_double_grad
Z
zyfncg 已提交
510 511 512
  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)
513
  invoke : expand(grad_x_grad, shape)
Z
zyfncg 已提交
514

515
- backward_op : expand_grad
Z
zyfncg 已提交
516 517 518 519 520 521 522 523 524 525 526
  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

527
- backward_op : exponential__grad
528
  forward : exponential_ (Tensor x, float lam) -> Tensor(out)
529 530 531 532
  args : (Tensor out_grad)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
533
  invoke : zeros_like(out_grad)
534

535
- backward_op : fill_diagonal_grad
Z
zhiboniu 已提交
536 537 538 539 540 541 542
  forward : fill_diagonal (Tensor x, float value, int offset, bool wrap) -> Tensor(out)
  args : (Tensor out_grad, float value, int offset, bool wrap)
  output : Tensor(x_grad)
  infer_meta :
    func : FillDiagonalGradInferMeta
  kernel :
    func : fill_diagonal_grad
Z
zhiboniu 已提交
543

544
- backward_op : fill_diagonal_tensor_grad
Z
zhiboniu 已提交
545 546 547 548 549 550 551
  forward : fill_diagonal_tensor (Tensor x, Tensor y, int64_t offset, int dim1, int dim2) -> Tensor(out)
  args : (Tensor out_grad, int64_t offset, int dim1, int dim2)
  output : Tensor(x_grad)
  infer_meta :
    func : FillDiagonalTensorGradInferMeta
  kernel :
    func : fill_diagonal_tensor_grad
552 553
  inplace : (out_grad -> x_grad)

554
- backward_op : fill_grad
555 556 557 558 559 560 561 562 563 564
  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)

565
- backward_op : flatten_grad
Z
zyfncg 已提交
566 567 568 569 570 571 572 573 574 575 576 577 578
  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)

579
- backward_op : fmax_grad
Z
zyfncg 已提交
580 581 582 583 584 585 586 587 588
  forward : fmax(Tensor x, Tensor y, int axis) -> Tensor(out)
  args : (Tensor x, Tensor y, Tensor out_grad, int axis)
  output : Tensor(x_grad), Tensor(y_grad)
  infer_meta :
    func : GeneralBinaryGradInferMeta
    param: [x, y]
  kernel :
    func : fmax_grad

589
- backward_op : fmin_grad
Z
zyfncg 已提交
590 591 592 593 594 595 596 597 598
  forward : fmin(Tensor x, Tensor y, int axis) -> Tensor(out)
  args : (Tensor x, Tensor y, Tensor out_grad, int axis)
  output : Tensor(x_grad), Tensor(y_grad)
  infer_meta :
    func : GeneralBinaryGradInferMeta
    param: [x, y]
  kernel :
    func : fmin_grad

599
- backward_op : frame_grad
C
Charles-hit 已提交
600 601 602 603 604 605 606 607 608
  forward : frame(Tensor x, int frame_length, int hop_length, int axis) -> Tensor(out)
  args : (Tensor x, Tensor out_grad, int frame_length, int hop_length, int axis)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [x]
  kernel :
    func : frame_grad

609
- backward_op : frobenius_norm_grad
Z
zyfncg 已提交
610 611 612 613 614 615 616 617 618
  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

619
- backward_op : gather_grad
Z
zyfncg 已提交
620 621 622 623 624 625 626 627 628 629 630
  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

631
- backward_op : gather_nd_grad
Z
zyfncg 已提交
632 633 634 635 636 637 638 639 640 641
  forward : gather_nd (Tensor x, Tensor index) -> Tensor(out)
  args : (Tensor x, Tensor index, Tensor out_grad)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [x]
  kernel :
    func : gather_nd_grad
  no_need_buffer : x

642
- backward_op : gelu_grad
Z
zyfncg 已提交
643 644 645 646 647 648 649 650 651
  forward : gelu(Tensor x,  bool approximate) -> Tensor(out)
  args : (Tensor x, Tensor out_grad,  bool approximate)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param: [x]
  kernel :
    func : gelu_grad

652
- backward_op : grid_sample_grad
W
Wang Bojun 已提交
653 654 655
  forward : grid_sample (Tensor x, Tensor grid, str mode, str padding_mode, bool align_corners) -> Tensor(out)
  args : (Tensor x, Tensor grid, Tensor out_grad, str mode, str padding_mode, bool align_corners)
  output : Tensor(x_grad), Tensor(grid_grad)
656
  infer_meta :
W
Wang Bojun 已提交
657 658
    func : GeneralBinaryGradInferMeta
    param : [x, grid]
659
  kernel :
W
Wang Bojun 已提交
660 661 662
    func : grid_sample_grad
    data_type : x

663
- backward_op : group_norm_grad
Z
zyfncg 已提交
664 665 666 667 668 669 670 671 672 673 674 675
  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)

676
- backward_op : gumbel_softmax_grad
Z
zyfncg 已提交
677 678 679 680 681 682 683 684 685
  forward : gumbel_softmax (Tensor x, float temperature, bool hard, int axis) -> Tensor(out)
  args : (Tensor out, Tensor out_grad, int axis)
  output : Tensor(x_grad)
  infer_meta :
    func : GumbelSoftmaxGradInferMeta
    param : [out, out_grad, axis]
  kernel :
    func : gumbel_softmax_grad

686 687
- backward_op : hardswish_grad
  forward : hardswish (Tensor x, float threshold = 6.0, float scale = 6.0, float offset = 3.0) -> Tensor(out)
Z
zyfncg 已提交
688 689 690 691 692 693 694 695 696
  args : (Tensor x, Tensor out_grad, float threshold, float scale, float offset)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [x]
  kernel :
    func : hard_swish_grad
  inplace : (out_grad -> x_grad)

697 698 699 700 701 702 703 704 705 706 707
- backward_op : hardtanh_grad
  forward : hardtanh (Tensor x, float t_min, float t_max) -> Tensor(out)
  args : (Tensor x, Tensor out_grad, float t_min, float t_max)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [x]
  kernel :
    func : hard_tanh_grad
  inplace : (out_grad -> x_grad)

708 709
- backward_op : hsigmoid_loss_grad
  forward : hsigmoid_loss (Tensor x, Tensor w, Tensor label, Tensor path, Tensor code, Tensor bias, int num_classes, bool remote_prefetch, int trainer_id, int64_t[] height_sections, str[] epmap, str[] table_names, bool is_sparse) -> Tensor(out), Tensor(pre_out), Tensor(w_out)
710 711 712 713 714 715 716
  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, int trainer_id, int64_t[] height_sections, str[] epmap, str[] table_names, bool is_sparse)
  output : Tensor(x_grad), Tensor(w_grad), Tensor(bias_grad)
  infer_meta :
    func : GeneralTernaryGradInferMeta
    param : [x ,w, bias]
  optional: path, code, bias
  kernel :
717
    func : hsigmoid_loss_grad
718

719
- backward_op : huber_loss_grad
Z
zyfncg 已提交
720 721 722 723 724 725 726 727 728
  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

729
- backward_op : imag_grad
Z
zyfncg 已提交
730 731 732 733 734
  forward : imag (Tensor x) -> Tensor(out)
  args : (Tensor out_grad)
  output : Tensor(x_grad)
  invoke : imag_grad_impl(out_grad, x_grad)

735
- backward_op : index_add_grad
L
Li Min 已提交
736 737 738 739 740 741 742 743 744 745
  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)

746
- backward_op : index_sample_grad
Z
zyfncg 已提交
747 748 749 750 751 752 753 754 755 756 757
  forward : index_sample (Tensor x, Tensor index) -> Tensor(out)
  args : (Tensor x, Tensor index, Tensor out_grad)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [x]
  kernel :
    func : index_sample_grad
    data_type : out_grad
  no_need_buffer : x

758
- backward_op : index_select_grad
759 760
  forward : index_select(Tensor x, Tensor index,  int axis) -> Tensor(out)
  args : (Tensor x, Tensor index, Tensor out_grad,  int axis)
Z
zyfncg 已提交
761 762 763 764 765 766 767 768 769
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [x]
  kernel :
    func : index_select_grad
    data_type : x
  no_need_buffer : x

770
- backward_op : instance_norm_double_grad
Z
zyfncg 已提交
771 772 773 774 775 776 777 778 779 780
  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

781
- backward_op : instance_norm_grad
Z
zyfncg 已提交
782 783 784 785 786 787 788 789 790 791 792
  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

793
- backward_op : inverse_grad
794 795 796 797 798 799 800 801
  forward : inverse(Tensor x) -> Tensor(out)
  args : (Tensor out, Tensor out_grad)
  output : Tensor(x_grad)
  infer_meta:
    func : InverseGradInferMeta
  kernel :
    func : inverse_grad

802
- backward_op : kldiv_loss_grad
Z
zyfncg 已提交
803 804 805 806 807 808 809 810 811 812
  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

813
- backward_op : kron_grad
Z
zyfncg 已提交
814 815 816 817 818 819 820 821 822 823
  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

824
- backward_op : kthvalue_grad
Z
zyfncg 已提交
825 826 827 828 829 830 831 832 833
  forward : kthvalue(Tensor x, int k, int axis, bool keepdim) -> Tensor(out), Tensor(indices)
  args : (Tensor x, Tensor indices, Tensor out_grad, int k, int axis, bool keepdim)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param: [x]
  kernel :
    func : kthvalue_grad

834
- backward_op : label_smooth_grad
Z
zyfncg 已提交
835 836 837 838 839 840 841 842 843
  forward : label_smooth (Tensor label, Tensor prior_dist, float epsilon) -> Tensor(out)
  args : (Tensor out_grad, float epsilon)
  output : Tensor(label_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [out_grad]
  kernel :
    func : label_smooth_grad

844
- backward_op : layer_norm_grad
Z
zyfncg 已提交
845 846 847 848 849 850 851 852 853 854 855 856
  forward : layer_norm (Tensor x, Tensor scale, Tensor bias, float epsilon, int begin_norm_axis, bool is_test) -> 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, bool is_test)
  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

857
- backward_op : lerp_grad
Z
zyfncg 已提交
858 859 860 861 862 863 864 865 866
  forward : lerp (Tensor x, Tensor y, Tensor weight) -> Tensor(out)
  args : (Tensor x, Tensor y, Tensor weight, Tensor out, Tensor out_grad)
  output : Tensor(x_grad), Tensor(y_grad)
  infer_meta :
    func : GeneralBinaryGradInferMeta
    param : [x, y]
  kernel :
    func : lerp_grad

867
- backward_op : linear_interp_grad
868
  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)
869 870 871 872 873 874 875
  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 :
876
    func : linear_interp_grad
877 878
    data_type : output_grad

879
- backward_op : log_loss_grad
Z
zyfncg 已提交
880 881 882 883 884 885 886 887 888
  forward : log_loss (Tensor input, Tensor label, float epsilon) -> Tensor(out)
  args : (Tensor input, Tensor label, Tensor out_grad, float epsilon)
  output : Tensor(input_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [input]
  kernel :
    func : log_loss_grad

889
- backward_op : log_softmax_grad
Z
zyfncg 已提交
890 891 892 893 894 895 896 897 898
  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

899
- backward_op : logcumsumexp_grad
Z
zyfncg 已提交
900 901 902 903 904 905 906 907 908
  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

909
- backward_op : logsumexp_grad
Z
zyfncg 已提交
910 911 912 913 914 915 916 917 918
  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

919
- backward_op : lu_grad
L
Lin Manhui 已提交
920 921 922 923 924 925 926 927
  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

928
- backward_op : lu_unpack_grad
929 930
  forward : lu_unpack (Tensor x, Tensor y, bool unpack_ludata, bool unpack_pivots) -> Tensor(pmat), Tensor(l), Tensor(u)
  args : (Tensor x, Tensor y, Tensor l, Tensor u, Tensor pmat, Tensor l_grad, Tensor u_grad, bool unpack_ludata, bool unpack_pivots)
931 932 933 934 935 936
  output : Tensor(x_grad)
  infer_meta :
    func : LUUnpackGradInferMeta
  kernel :
    func : lu_unpack_grad

937
- backward_op : margin_cross_entropy_grad
938 939 940 941 942 943 944 945 946 947
  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)

948
- backward_op : masked_select_grad
Z
zyfncg 已提交
949 950 951 952 953 954 955 956 957 958 959
  forward : masked_select (Tensor x, Tensor mask) -> Tensor(out)
  args : (Tensor x, Tensor mask, Tensor out_grad)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [x]
  kernel :
    func : masked_select_grad
    data_type : x
  no_need_buffer : x

960
- backward_op : matmul_double_grad
Z
zyfncg 已提交
961 962 963 964 965 966 967 968 969 970 971
  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

972
- backward_op : matmul_grad
Z
zyfncg 已提交
973 974 975 976 977 978 979 980 981 982
  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

983
- backward_op : matmul_triple_grad
Z
zyfncg 已提交
984 985 986 987 988 989 990 991 992 993
  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
  optional : grad_x_grad, grad_y_grad, grad_grad_out_grad

994
- backward_op : matrix_power_grad
Z
zyfncg 已提交
995 996 997 998 999 1000 1001 1002 1003
  forward : matrix_power (Tensor x, int n) -> Tensor(out)
  args : (Tensor x, Tensor out, Tensor out_grad, int n)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [x]
  kernel :
    func : matrix_power_grad

1004
- backward_op : max_grad
1005 1006
  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
1007 1008 1009 1010 1011 1012 1013
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param: [x]
  kernel :
    func : max_grad

1014
- backward_op : max_pool2d_with_index_grad
Z
zyfncg 已提交
1015 1016 1017 1018 1019 1020 1021 1022
  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

1023
- backward_op : max_pool3d_with_index_grad
Z
zyfncg 已提交
1024 1025 1026 1027 1028 1029 1030 1031
  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

1032
- backward_op : maximum_grad
Z
zyfncg 已提交
1033 1034 1035 1036 1037 1038 1039 1040 1041
  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

1042
- backward_op : maxout_grad
Z
zyfncg 已提交
1043 1044 1045 1046 1047 1048 1049 1050 1051
  forward : maxout(Tensor x, int groups, int axis) -> Tensor(out)
  args : (Tensor x, Tensor out, Tensor out_grad, int groups, int axis)
  output : Tensor(x_grad)
  infer_meta :
    func : GeneralUnaryGradInferMeta
    param: [x]
  kernel :
    func : maxout_grad

1052
- backward_op : mean_all_grad
Z
zyfncg 已提交
1053 1054 1055 1056 1057 1058 1059 1060 1061
  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

1062
- backward_op : mean_double_grad
1063 1064
  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
1065
  output : Tensor(grad_out_grad)
1066
  invoke : mean(grad_x_grad, axis, keepdim)
Z
zyfncg 已提交
1067

1068
- backward_op : mean_grad
1069 1070
  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
1071 1072 1073 1074 1075 1076 1077 1078 1079
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param: [x]
  kernel :
    func : mean_grad
  backward : mean_double_grad
  no_need_buffer : x

1080
- backward_op : meshgrid_grad
Z
zyfncg 已提交
1081 1082 1083 1084 1085 1086 1087 1088
  forward : meshgrid (Tensor[] inputs) -> Tensor[](outputs)
  args : (Tensor[] inputs, Tensor[] outputs_grad)
  output : Tensor[](inputs_grad){inputs.size()}
  infer_meta :
    func : MeshgridGradInferMeta
  kernel :
    func : meshgrid_grad

1089
- backward_op : min_grad
1090 1091
  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
1092 1093 1094 1095 1096 1097 1098
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param: [x]
  kernel :
    func : min_grad

1099
- backward_op : minimum_grad
Z
zyfncg 已提交
1100 1101 1102 1103 1104 1105 1106 1107 1108
  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

1109
- backward_op : mish_grad
Z
zyfncg 已提交
1110 1111 1112 1113 1114 1115 1116 1117 1118 1119
  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)

1120
- backward_op : mode_grad
Z
zyfncg 已提交
1121 1122 1123 1124 1125 1126 1127 1128 1129
  forward : mode(Tensor x,  int axis,  bool keepdim) -> Tensor(out), Tensor(indices)
  args : (Tensor x, Tensor indices, Tensor out_grad,  int axis,  bool keepdim)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param: [x]
  kernel :
    func : mode_grad

1130
- backward_op : multi_dot_grad
Z
zyfncg 已提交
1131 1132 1133 1134 1135 1136 1137 1138
  forward : multi_dot (Tensor[] x) -> Tensor(out)
  args : (Tensor[] x, Tensor out_grad)
  output : Tensor[](x_grad) {x.size()}
  infer_meta :
    func : MultiDotGradInferMeta
  kernel :
    func : multi_dot_grad

1139
- backward_op : multiplex_grad
1140 1141 1142
  forward : multiplex (Tensor[] inputs, Tensor index) -> Tensor(out)
  args : (Tensor[] inputs, Tensor index, Tensor out_grad)
  output : Tensor[](inputs_grad){inputs.size()}
Z
zyfncg 已提交
1143 1144
  infer_meta :
    func : MultiplexGradInferMeta
1145
    param : [index, out_grad]
Z
zyfncg 已提交
1146 1147
  kernel :
    func : multiplex_grad
1148
    param : [index, out_grad]
Z
zyfncg 已提交
1149

1150
- backward_op : multiply_double_grad
Z
zyfncg 已提交
1151 1152 1153 1154 1155 1156 1157 1158 1159 1160 1161 1162
  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)

1163
- backward_op : multiply_grad
Z
zyfncg 已提交
1164 1165 1166 1167 1168 1169 1170 1171 1172 1173
  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

1174
- backward_op : multiply_triple_grad
Z
zyfncg 已提交
1175 1176 1177 1178 1179 1180 1181 1182 1183 1184
  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
    param : [x, y, fwd_grad_out, x, y]
  kernel :
    func : multiply_triple_grad
  optional : fwd_grad_grad_x, fwd_grad_grad_y, grad_grad_out_grad

1185
- backward_op : nearest_interp_grad
1186 1187 1188 1189 1190 1191 1192 1193 1194 1195 1196
  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

1197
- backward_op : nll_loss_grad
Z
zyfncg 已提交
1198 1199 1200 1201 1202 1203 1204 1205 1206 1207
  forward : nll_loss (Tensor input, Tensor label, Tensor weight, int64_t ignore_index, str reduction) -> Tensor(out), Tensor(total_weight)
  args : (Tensor input, Tensor label, Tensor weight, Tensor total_weight, Tensor out_grad, int64_t ignore_index, str reduction)
  output : Tensor(input_grad)
  infer_meta :
    func : NllLossGradInferMeta
  kernel :
    func : nll_loss_grad
    data_type : input
  optional : weight

1208
- backward_op : norm_grad
Z
zyfncg 已提交
1209 1210 1211 1212 1213 1214 1215 1216 1217
  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

1218
- backward_op : overlap_add_grad
1219 1220 1221 1222 1223 1224 1225 1226 1227
  forward : overlap_add(Tensor x, int hop_length, int axis) -> Tensor(out)
  args : (Tensor x, Tensor out_grad, int hop_length, int axis)
  output : Tensor(x_grad)
  infer_meta :
    func : OverlapAddGradInferMeta
  kernel :
    func : overlap_add_grad
    data_type : x

1228
- backward_op : p_norm_grad
Z
zyfncg 已提交
1229 1230 1231 1232 1233 1234 1235 1236 1237
  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

1238
- backward_op : pad3d_double_grad
Z
zyfncg 已提交
1239 1240 1241 1242 1243 1244 1245 1246
  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

1247
- backward_op : pad3d_grad
Z
zyfncg 已提交
1248 1249 1250 1251 1252 1253 1254 1255 1256 1257 1258
  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

1259
- backward_op : pad_double_grad
1260 1261
  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
1262 1263 1264 1265 1266 1267
  output : Tensor(grad_out_grad)
  infer_meta :
    func : PadInferMeta
  kernel :
    func : pad

1268
- backward_op : pad_grad
1269 1270
  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
1271 1272 1273 1274 1275 1276 1277 1278 1279 1280
  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

1281
- backward_op : pixel_shuffle_grad
Z
zyfncg 已提交
1282 1283 1284 1285 1286 1287 1288 1289
  forward : pixel_shuffle (Tensor x, int upscale_factor, str data_format) -> Tensor(out)
  args : (Tensor out_grad, int upscale_factor, str data_format)
  output : Tensor(x_grad)
  infer_meta :
    func : PixelShuffleGradInferMeta
  kernel :
    func : pixel_shuffle_grad

1290
- backward_op : pool2d_double_grad
1291 1292
  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, bool use_gpudnn) -> Tensor(grad_x)
  args : (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, bool use_gpudnn)
Z
zyfncg 已提交
1293 1294
  output : Tensor(grad_out_grad)
  infer_meta :
1295
    func : Pool2DInferMeta
1296
    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
1297 1298
  kernel :
    func : pool2d_double_grad
1299 1300
    param : [grad_x_grad, kernel_size, strides, paddings, ceil_mode, exclusive, data_format, pooling_type, global_pooling, adaptive, padding_algorithm]
    use_gpudnn : use_gpudnn
Z
zyfncg 已提交
1301

1302
- backward_op : pool2d_grad
1303 1304
  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, bool use_gpudnn) -> 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, bool use_gpudnn)
Z
zyfncg 已提交
1305 1306
  output : Tensor(x_grad)
  infer_meta :
1307 1308
    func : UnchangedInferMeta
    param: [x]
Z
zyfncg 已提交
1309 1310
  kernel :
    func : pool2d_grad
1311 1312
    param : [x, out, out_grad, kernel_size, strides, paddings, ceil_mode, exclusive, data_format, pooling_type, global_pooling, adaptive, padding_algorithm]
    use_gpudnn : use_gpudnn
Z
zyfncg 已提交
1313 1314
  backward : pool2d_double_grad

1315
- backward_op : pool3d_grad
1316 1317
  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, bool use_gpudnn) -> 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, bool use_gpudnn)
Z
zyfncg 已提交
1318 1319
  output : Tensor(x_grad)
  infer_meta :
1320 1321
    func : UnchangedInferMeta
    param: [x]
Z
zyfncg 已提交
1322 1323
  kernel :
    func : pool3d_grad
1324 1325
    param : [x, out, out_grad, kernel_size, strides, paddings, ceil_mode, exclusive, data_format, pooling_type, global_pooling, adaptive, padding_algorithm]
    use_gpudnn : use_gpudnn
Z
zyfncg 已提交
1326

C
Charles-hit 已提交
1327 1328 1329 1330 1331 1332
- 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
1333
    param: [x, grad_out]
C
Charles-hit 已提交
1334 1335
  kernel :
    func : pow_double_grad
C
Charles-hit 已提交
1336
  backward : pow_triple_grad
C
Charles-hit 已提交
1337 1338
  inplace : (grad_x_grad -> x_grad)

1339
- backward_op : pow_grad
1340 1341
  forward : pow(Tensor x, Scalar y) -> Tensor(out)
  args : (Tensor x, Tensor out_grad, Scalar y=-1)
Z
zyfncg 已提交
1342 1343 1344 1345 1346 1347
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param: [x]
  kernel :
    func : pow_grad
C
Charles-hit 已提交
1348
  backward: pow_double_grad
Z
zyfncg 已提交
1349 1350
  inplace : (out_grad -> x_grad)

C
Charles-hit 已提交
1351 1352 1353 1354 1355 1356 1357 1358 1359 1360
- 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

1361
- backward_op : prelu_grad
Z
zyfncg 已提交
1362 1363 1364 1365 1366 1367 1368 1369 1370
  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

1371 1372 1373 1374 1375 1376 1377 1378 1379 1380
- 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

1381
- backward_op : psroi_pool_grad
Z
zyfncg 已提交
1382 1383 1384 1385 1386 1387 1388 1389 1390 1391 1392 1393
  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

# output is optional
1394
- backward_op : put_along_axis_grad
1395 1396
  forward : put_along_axis (Tensor arr, Tensor indices, Tensor value, int axis, str reduce) -> Tensor(out)
  args : (Tensor arr, Tensor indices, Tensor out_grad, int axis, str reduce)
1397
  output : Tensor(arr_grad), Tensor(value_grad)
Z
zyfncg 已提交
1398 1399
  infer_meta :
    func : GeneralBinaryGradInferMeta
1400
    param : [arr, indices]
Z
zyfncg 已提交
1401 1402 1403
  kernel :
    func : put_along_axis_grad

1404
- backward_op : qr_grad
Y
Yulong Ao 已提交
1405 1406 1407 1408 1409 1410 1411 1412 1413
  forward : qr (Tensor x, str mode) -> Tensor(q), Tensor(r)
  args : (Tensor x, Tensor q, Tensor r, Tensor q_grad, Tensor r_grad, str mode)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [x]
  kernel :
    func : qr_grad

1414
- backward_op : real_grad
Z
zyfncg 已提交
1415 1416 1417 1418 1419
  forward : real (Tensor x) -> Tensor(out)
  args : (Tensor out_grad)
  output : Tensor(x_grad)
  invoke : real_grad_impl(out_grad, x_grad)

1420
- backward_op : relu6_grad
1421 1422 1423 1424 1425 1426 1427 1428 1429 1430
  forward : relu6 (Tensor x, float threshold) -> Tensor(out)
  args : (Tensor out, Tensor out_grad, float threshold)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [out]
  kernel :
    func : relu6_grad
  inplace : (out_grad -> x_grad)

1431
- backward_op : renorm_grad
S
seemingwang 已提交
1432 1433 1434 1435 1436 1437 1438 1439 1440
  forward : renorm (Tensor x, float p, int axis, float max_norm) -> Tensor(out)
  args : (Tensor x, Tensor out_grad, float p, int axis, float max_norm)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [out_grad]
  kernel :
    func : renorm_grad

1441
- backward_op : repeat_interleave_grad
1442 1443
  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
1444 1445 1446 1447 1448 1449 1450
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [x]
  kernel :
    func : repeat_interleave_grad

1451
- backward_op : repeat_interleave_with_tensor_index_grad
1452 1453
  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
1454 1455 1456 1457 1458 1459 1460 1461
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [x]
  kernel :
    func : repeat_interleave_with_tensor_index_grad
    data_type : x

1462
- backward_op : reshape_double_grad
Z
zyfncg 已提交
1463 1464 1465 1466 1467 1468 1469 1470 1471 1472 1473
  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)

1474
- backward_op : reshape_grad
Z
zyfncg 已提交
1475 1476 1477 1478 1479 1480 1481 1482 1483 1484 1485 1486 1487 1488 1489
  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)

1490
- backward_op : reverse_array_grad
1491 1492
  forward : reverse_array (Tensor[] x, IntArray axis) -> Tensor[](out)
  args : (Tensor[] out_grad, IntArray axis)
W
wanghuancoder 已提交
1493 1494 1495 1496 1497 1498
  output : Tensor[](x_grad){out_grad.size()}
  infer_meta :
    func : ReverseArrayInferMeta
  kernel :
    func : reverse

1499
- backward_op : reverse_grad
1500 1501
  forward : reverse (Tensor x, IntArray axis) -> Tensor(out)
  args : (Tensor out_grad, IntArray axis)
W
wanghuancoder 已提交
1502 1503 1504
  output : Tensor(x_grad)
  invoke : reverse(out_grad, axis)

Y
YuanRisheng 已提交
1505 1506 1507 1508 1509 1510 1511 1512 1513 1514 1515 1516
- 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

1517
- backward_op : roi_align_grad
Z
zyfncg 已提交
1518 1519 1520 1521 1522 1523 1524 1525 1526 1527 1528 1529
  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

1530
- backward_op : roi_pool_grad
Z
zyfncg 已提交
1531 1532 1533 1534 1535 1536 1537 1538 1539 1540 1541
  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

1542
- backward_op : roll_grad
Z
zyfncg 已提交
1543 1544 1545 1546 1547 1548 1549 1550 1551 1552 1553
  forward : roll(Tensor x, IntArray shifts, int64_t[] axis) -> Tensor(out)
  args : (Tensor x, Tensor out_grad, IntArray shifts, int64_t[] axis)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [x]
  kernel :
    func : roll_grad
    data_type : x
  no_need_buffer : x

1554
- backward_op : scale_grad
Z
zyfncg 已提交
1555
  forward : scale (Tensor x, Scalar scale, float bias, bool bias_after_scale) -> Tensor(out)
1556
  args : (Tensor out_grad, Scalar scale=1.0, bool bias_after_scale=true)
Z
zyfncg 已提交
1557 1558 1559
  output : Tensor(x_grad)
  invoke : scale(out_grad, scale, 0.0, bias_after_scale)

1560
- backward_op : scatter_grad
Z
zyfncg 已提交
1561 1562 1563 1564 1565 1566 1567 1568 1569 1570
  forward : scatter (Tensor x, Tensor index, Tensor updates, bool overwrite) -> Tensor(out)
  args : (Tensor index, Tensor updates, Tensor out_grad, bool overwrite)
  output : Tensor(x_grad), Tensor(updates_grad)
  infer_meta :
    func : ScatterGradInferMeta
    param : [index, updates, out_grad, overwrite]
  kernel :
    func : scatter_grad
  no_need_buffer : updates

1571
- backward_op : scatter_nd_add_grad
Z
zyfncg 已提交
1572 1573 1574 1575 1576 1577 1578 1579 1580 1581
  forward : scatter_nd_add (Tensor x, Tensor index, Tensor updates) -> Tensor(out)
  args : (Tensor index, Tensor updates, Tensor out_grad)
  output : Tensor(x_grad), Tensor(updates_grad)
  infer_meta :
    func : ScatterNdAddGradInferMeta
    param : [index, updates, out_grad]
  kernel :
    func : scatter_nd_add_grad
  no_need_buffer : updates

1582
- backward_op : segment_pool_grad
Z
zyfncg 已提交
1583 1584 1585 1586 1587 1588 1589 1590 1591 1592 1593
  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

1594
- backward_op : selu_grad
Z
zyfncg 已提交
1595 1596 1597 1598 1599 1600 1601 1602 1603
  forward : selu (Tensor x, float scale, float alpha) -> Tensor(out)
  args : (Tensor out, Tensor out_grad, float scale, float alpha)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [out]
  kernel :
    func : selu_grad

1604 1605 1606 1607 1608 1609 1610 1611 1612 1613 1614 1615 1616 1617 1618 1619 1620 1621 1622 1623 1624 1625 1626 1627
- 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

1628
- backward_op : sigmoid_cross_entropy_with_logits_grad
Z
zyfncg 已提交
1629 1630 1631 1632 1633 1634 1635 1636 1637 1638
  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)

1639 1640 1641 1642 1643 1644
- 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)

1645
- backward_op : slice_double_grad
1646 1647 1648
  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)
1649
  invoke : slice(grad_input_grad, axes, starts, ends, infer_flags, decrease_axis)
1650

1651
- backward_op : slice_grad
Z
zyfncg 已提交
1652 1653 1654 1655 1656 1657 1658 1659
  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
1660
  backward : slice_double_grad
Z
zyfncg 已提交
1661 1662
  no_need_buffer : input

1663
- backward_op : slogdet_grad
1664 1665 1666 1667 1668 1669 1670 1671 1672
  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 :
    func : slogdeterminant_grad

1673
- backward_op : softmax_grad
Z
zyfncg 已提交
1674 1675 1676 1677 1678 1679 1680 1681 1682 1683
  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
    use_gpudnn : true

1684
- backward_op : spectral_norm_grad
1685 1686 1687 1688 1689 1690 1691 1692 1693
  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

1694
- backward_op : split_grad
Z
zyfncg 已提交
1695 1696 1697 1698
  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
1699

1700
- backward_op : split_with_num_grad
C
Charles-hit 已提交
1701 1702 1703 1704
  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
1705

1706
- backward_op : squared_l2_norm_grad
1707 1708 1709 1710 1711 1712 1713 1714 1715
  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

1716
- backward_op : squeeze_double_grad
1717 1718
  forward : squeeze_grad(Tensor xshape, Tensor grad_out, IntArray axis) -> Tensor(grad_x)
  args : (Tensor grad_x_grad, IntArray axis)
Z
zyfncg 已提交
1719
  output : Tensor(grad_out_grad)
1720
  invoke: squeeze(grad_x_grad, axis)
Z
zyfncg 已提交
1721

1722
- backward_op : squeeze_grad
1723 1724
  forward : squeeze(Tensor x, IntArray axis) -> Tensor(out), Tensor(xshape)
  args : (Tensor xshape, Tensor out_grad, IntArray axis)
Z
zyfncg 已提交
1725 1726 1727 1728 1729 1730 1731 1732 1733
  output : Tensor(x_grad)
  infer_meta :
    func : KernelWithXShapeInferMeta
    param: [xshape]
  kernel :
    func : squeeze_grad
  inplace : (out_grad -> x_grad)
  backward: squeeze_double_grad

1734
- backward_op : stack_grad
Z
zyfncg 已提交
1735 1736 1737 1738 1739 1740 1741 1742 1743 1744 1745
  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

1746
- backward_op : strided_slice_grad
Z
zyfncg 已提交
1747 1748 1749 1750 1751 1752 1753 1754 1755 1756
  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

1757
- backward_op : subtract_double_grad
Z
zyfncg 已提交
1758 1759 1760 1761 1762 1763 1764 1765 1766 1767 1768 1769
  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)

1770
- backward_op : subtract_grad
Z
zyfncg 已提交
1771 1772 1773 1774 1775 1776 1777 1778 1779 1780 1781 1782
  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)

1783
- backward_op : sum_double_grad
1784 1785
  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
1786
  output : Tensor(grad_out_grad)
1787
  invoke : sum(grad_x_grad, axis, grad_x_grad.dtype(), keepdim)
Z
zyfncg 已提交
1788

1789
- backward_op : sum_grad
1790 1791
  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
1792 1793 1794 1795 1796 1797 1798 1799 1800
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [x]
  kernel :
    func : sum_grad
  no_need_buffer : x
  backward : sum_double_grad

1801
- backward_op : svd_grad
1802 1803
  forward : svd (Tensor x, bool full_matrices) -> Tensor(u), Tensor(s), Tensor(vh)
  args : (Tensor x, Tensor u, Tensor vh, Tensor s, Tensor u_grad, Tensor vh_grad, Tensor s_grad, bool full_matrices)
1804 1805 1806 1807 1808 1809 1810 1811
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [x]
  kernel :
    func : svd_grad
  optional: u_grad, vh_grad, s_grad

1812
- backward_op : swish_grad
Z
zyfncg 已提交
1813 1814 1815 1816 1817 1818 1819 1820 1821 1822
  forward : swish (Tensor x, float beta=1.0) -> Tensor(out)
  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)

1823
- backward_op : sync_batch_norm_grad
1824 1825
  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)
1826 1827 1828 1829 1830 1831 1832
  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
1833
  optional : reserve_space
1834

1835
- backward_op : take_along_axis_grad
1836 1837 1838
  forward : take_along_axis (Tensor arr, Tensor indices, int axis) -> Tensor(out)
  args : (Tensor arr, Tensor indices, Tensor out_grad, int axis)
  output : Tensor(arr_grad)
Z
zyfncg 已提交
1839 1840
  infer_meta :
    func : UnchangedInferMeta
1841
    param : [arr]
Z
zyfncg 已提交
1842 1843 1844
  kernel :
    func : take_along_axis_grad

1845
- backward_op : temporal_shift_grad
C
ccrrong 已提交
1846 1847 1848 1849 1850 1851 1852 1853 1854
  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

1855
- backward_op : tile_double_grad
Z
zyfncg 已提交
1856 1857 1858
  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)
1859
  invoke : tile(grad_x_grad, repeat_times)
Z
zyfncg 已提交
1860

1861
- backward_op : tile_grad
Z
zyfncg 已提交
1862 1863 1864 1865 1866 1867 1868 1869 1870 1871 1872
  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

1873 1874
- backward_op : topk_grad
  forward : topk (Tensor x, Scalar k, int axis = -1, bool largest = true, bool sorted = true) -> Tensor(out), Tensor(indices)
Z
zyfncg 已提交
1875 1876 1877 1878 1879 1880 1881 1882
  args : (Tensor x, Tensor indices, Tensor out_grad, Scalar k = -1, int axis = -1, bool largest = true, bool sorted = true)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [x]
  kernel :
    func : top_k_grad

1883
- backward_op : transpose_double_grad
1884 1885
  forward : transpose_grad (Tensor grad_out, int[] perm) -> Tensor(grad_x)
  args : (Tensor grad_x_grad, int[] perm)
Z
zyfncg 已提交
1886
  output : Tensor(grad_out_grad)
1887
  invoke : transpose(grad_x_grad, perm)
Z
zyfncg 已提交
1888

1889
- backward_op : transpose_grad
1890 1891
  forward : transpose (Tensor x, int[] perm) -> Tensor(out)
  args : (Tensor out_grad, int[] perm)
Z
zyfncg 已提交
1892 1893 1894
  output : Tensor(x_grad)
  infer_meta :
    func : TransposeGradInferMeta
1895
    param : [out_grad, perm]
Z
zyfncg 已提交
1896 1897 1898 1899
  kernel :
    func : transpose_grad
  backward : transpose_double_grad

1900
- backward_op : triangular_solve_grad
Z
zyfncg 已提交
1901 1902 1903 1904 1905 1906 1907 1908 1909
  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

1910 1911
- backward_op : tril_grad
  forward : tril(Tensor x,  int diagonal,  bool lower) -> Tensor(out)
Z
zyfncg 已提交
1912 1913 1914 1915 1916 1917
  args : (Tensor out_grad,  int diagonal,  bool lower)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [out_grad]
  kernel :
1918
    func : tril_grad
Z
zyfncg 已提交
1919

1920
- backward_op : trilinear_interp_grad
1921 1922 1923 1924 1925 1926 1927 1928 1929 1930 1931
  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

1932
- backward_op : unbind_grad
Z
zyfncg 已提交
1933 1934 1935 1936 1937
  forward : unbind (Tensor input, int axis) -> Tensor[](out)
  args : (Tensor[] out_grad, int axis)
  output : Tensor(input_grad)
  invoke : stack(out_grad, axis)

1938
- backward_op : unfold_grad
Z
zyfncg 已提交
1939 1940 1941 1942 1943 1944 1945 1946 1947 1948
  forward : unfold (Tensor x, int[] kernel_sizes, int[] strides, int[] paddings, int[] dilations) -> Tensor(out)
  args : (Tensor x, Tensor out_grad, int[] kernel_sizes, int[] strides, int[] paddings, int[] dilations)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [x]
  kernel :
    func : unfold_grad
  no_need_buffer : x

1949 1950
- 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)
1951 1952 1953 1954 1955
  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 :
1956
    func : uniform_inplace_grad
1957 1958
  inplace : (out_grad -> x_grad)

1959
- backward_op : unsqueeze_double_grad
Z
zyfncg 已提交
1960 1961 1962 1963 1964
  forward : unsqueeze_grad(Tensor xshape, Tensor grad_out, IntArray axes) -> Tensor(grad_x)
  args : (Tensor grad_x_grad, IntArray axes)
  output : Tensor(grad_out_grad)
  invoke : unsqueeze(grad_x_grad, axes)

1965
- backward_op : unsqueeze_grad
Z
zyfncg 已提交
1966 1967 1968 1969 1970 1971 1972 1973 1974 1975 1976 1977
  forward : unsqueeze(Tensor x, IntArray axes) -> Tensor(out), Tensor(xshape)
  args : (Tensor xshape, Tensor out_grad, IntArray axes)
  output : Tensor(x_grad)
  infer_meta :
    func : KernelWithXShapeInferMeta
    param: [xshape]
  kernel :
    func : unsqueeze_grad
    param: [xshape, out_grad]
  inplace : (out_grad -> x_grad)
  backward : unsqueeze_double_grad

1978
- backward_op : unstack_grad
1979 1980 1981 1982 1983 1984 1985 1986 1987
  forward : unstack (Tensor x, int axis, int num) -> Tensor[](out)
  args : (Tensor[] out_grad, int axis)
  output : Tensor(x_grad)
  infer_meta :
    func : UnStackGradInferMeta
    param : [out_grad, axis]
  kernel :
    func : unstack_grad

1988
- backward_op : warpctc_grad
1989
  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
1990 1991 1992 1993 1994 1995 1996 1997 1998 1999
  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

2000
- backward_op : where_grad
Z
zyfncg 已提交
2001 2002 2003 2004 2005 2006 2007 2008 2009
  forward : where (Tensor condition, Tensor x, Tensor y) -> Tensor(out)
  args : (Tensor condition, Tensor x, Tensor y, Tensor out_grad)
  output : Tensor(x_grad), Tensor(y_grad)
  infer_meta :
    func : GeneralBinaryGradInferMeta
    param : [x, y]
  kernel :
    func : where_grad
  no_need_buffer : x, y
2010

2011 2012
- 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)
2013 2014 2015
  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 :
2016
    func : YoloLossGradInferMeta
2017
  kernel :
2018
    func : yolo_loss_grad
2019
  optional : gt_score
X
xiaoting 已提交
2020

2021
- backward_op: fold_grad
X
xiaoting 已提交
2022 2023 2024 2025 2026 2027 2028 2029 2030 2031
  forward: fold (Tensor x, int[] output_sizes, int[] kernel_sizes, int[] strides, int[] paddings, int[] dilations) -> Tensor(out)
  args: (Tensor x, Tensor out_grad, int[] output_sizes, int[] kernel_sizes, int[] strides, int[] paddings, int[] dilations)
  output: Tensor(x_grad)
  infer_meta:
    func: UnchangedInferMeta
    param : [x]
  kernel:
    func: fold_grad
  no_need_buffer : x

2032
- backward_op: unpool3d_grad
X
xiaoting 已提交
2033 2034 2035 2036 2037 2038 2039 2040 2041 2042
  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

2043
- backward_op: unpool_grad
2044 2045
  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
2046 2047 2048 2049 2050 2051 2052
  output: Tensor(x_grad)
  infer_meta:
    func: UnchangedInferMeta
    param : [x]
  kernel:
    func: unpool_grad
    data_type: x