OpInfos are Python classes containing structured metadata, test directives, and sample inputs for PyTorch's operators. This data is used to automatically generate a variety of tests, including tests that operators work with a variety of systems, like autograd, forward autograd, torchscript, FX, and NNC correctly.

For more details on OpInfos and how to add them see the Github wiki article here. If you have more questions ping @mruberry directly (like in a comment responding to this issue).

Priority Functions to OpInfo

• nn.functional.batch_norm (OpInfo for nn.functional.batch_norm #63218)
• nn.functional.layer_norm (OpInfo for nn.functional.layer_norm #63276)
• nn.functional.interpolate (OpInfo for nn.functional.interpolate #61956)
• nn.functional.cross_entropy (OpInfo for nn.functional.cross_entropy #63547)
• nn.functional.conv2d (OpInfo: nn.functional.conv2d #63517)
• nn.functional.dropout (add OpInfo for torch.nn.functional.dropout #62315)
• nn.functional.linear (nn.functional.linear OpInfo #61971)
• nn.functional.max_pool2d (Add Maxpool to shape analysis / Opinfo #63530)
• nn.functional.nll_loss (add OpInfo for torch.nn.functional.nll_loss #63854)
• nn.functional.cosine_similarity (Add OpInfo for nn.functional.cosine_similarity #62959)
• nn.functional.embedding (Opinfo: embedding, add shape analysis #63959)
• cat (see tests in test_autograd.py) (Support torch.concat alias, add cat OpInfo & remove OpInfo test_out skips {cat, stack, hstack, vtack, dstack} #62560)
• block_diag (see tests in test_autograd.py) (Remove run_functional_checks from test_autograd and create necessary OpInfos #64993)
• broadcast_tensors (see tests in test_autograd.py) (Remove run_functional_checks from test_autograd and create necessary OpInfos #64993)
• lobpcg (see tests in test_autograd.py)
• torch.linalg.tensor_solve (see tests in test_linalg.py)
• igamma (see tests in test_autograd.py)
• as_strided (see tests in test_autograd.py)
• unbind (see tests in test_autograd.py)
• pdist (see tests in test_autograd.py)
• argsort (add OpInfo for torch.argsort #65454)
• repeat_interleave (add OpInfo for torch.repeat_interleave #65455)
• nn.functional.sigmoid (note that an OpInfo for torch.sigmoid exists)
• nn.functional.tanh (note than an OpInfo for torch.tanh exists)
• torch.nn.functional.conv1d (OpInfo for nn.functional.conv1d #67747)

OpInfo Backlog

• binary_cross_entropy_with_logits
• l1_loss (fix torch.nn.functional.l1_loss for complex inputs #65681)
• binary_cross_entropy
• elu
• upsample
• fold
• affine_grid
• max_pool1d ([OpInfo Hackathon] Parcel (1/2): OpInfo for max_pool1d, max_pool3d and max_poolNd_with_indices #67005)
• torch
• threshold
• smooth_l1_loss
• pairwise_distance (add OpInfo for torch.nn.functional.pairwise_distance #65460)
• logsigmoid
• adaptive_max_pool2d
• pixel_shuffle (add OpInfo for torch.nn.pixel_shuffle #65467)
• avg_pool3d Opinfos for avg_pooling #64214
• bilinear
• gumbel_softmax
• max_unpool2d (Improved error messages for max_unpool{}d operators #67328)
• kl_div (add OpInfo for torch.nn.functional.kl_div #65469)
• ctc_loss
• layer_norm (OpInfo for nn.functional.layer_norm #63276)
• conv3d
• max_unpool3d (Improved error messages for max_unpool{}d operators #67328)
• selu
• glu
• hardsigmoid
• upsample_bilinear
• max_pool3d ([OpInfo Hackathon] Parcel (1/2): OpInfo for max_pool1d, max_pool3d and max_poolNd_with_indices #67005)
• adaptive_avg_pool3d Opinfos for avg_pooling #64214
• instance_norm
• embedding_bag
• upsample_nearest
• avg_pool1d Opinfos for avg_pooling #64214
• prelu
• celu
• dropout2d (OpInfo for nn.functional.dropout2d, revise sample inputs for dropout #67891)
• hinge_embedding_loss
• softsign
• max_unpool1d (Improved error messages for max_unpool{}d operators #67328)
• silu
• softshrink
• leaky_relu_
• softmin
• channel_shuffle
• multilabel_margin_loss
• dropout3d
• multi_margin_loss
• lp_pool2d
• conv_transpose1d
• triplet_margin_loss
• tanhshrink
• adaptive_max_pool1d
• cosine_embedding_loss
• multi_head_attention_forward
• max_pool1d_with_indices ([OpInfo Hackathon] Parcel (1/2): OpInfo for max_pool1d, max_pool3d and max_poolNd_with_indices #67005)
• poisson_nll_loss
• margin_ranking_loss
• soft_margin_loss
• adaptive_max_pool3d
• group_norm
• local_response_norm
• multilabel_soft_margin_loss
• relu_
• alpha_dropout
• nn.functional.alpha_dropout (OpInfo for nn.functional.alpha_dropout #67823)
• feature_alpha_dropout
• lp_pool1d
• adaptive_max_pool1d_with_indices
• adaptive_max_pool2d_with_indices
• adaptive_max_pool3d_with_indices
• fractional_max_pool2d
• fractional_max_pool2d_with_indices
• fractional_max_pool3d
• fractional_max_pool3d_with_indices
• max_pool2d_with_indices ([OpInfo Hackathon] Parcel (1/2): OpInfo for max_pool1d, max_pool3d and max_poolNd_with_indices #67005)
• max_pool3d_with_indices ([OpInfo Hackathon] Parcel (1/2): OpInfo for max_pool1d, max_pool3d and max_poolNd_with_indices #67005)
• handle_torch_function
• has_torch_function
• adaptive_avg_pool1d Opinfos for avg_pooling #64214
• pdist
• rrelu_
• elu_
• hardtanh_
• triplet_margin_with_distance_loss
• selu_
• pixel_unshuffle (add OpInfo for torch.nn.pixel_unshuffle #65468)
• conv_transpose3d
• gaussian_nll_loss
• celu_
• huber_loss
• mish
• threshold_
• logical_and (Adding OpInfo for logical_or, logical_and, logical_xor #67178)
• logical_or (Adding OpInfo for logical_or, logical_and, logical_xor #67178)
• logical_xor (Adding OpInfo for logical_or, logical_and, logical_xor #67178)

cc @ezyang @gchanan @zou3519 @bdhirsh @jbschlosser @anjali411 @dylanbespalko @mruberry @lezcano @VitalyFedyunin @walterddr @mattip @kshitij12345