grad transform failing tests tracking issue

[zou3519]

Current fails for:

Calls internal_new_from_data: (#65)

• [x] __getitem__
• [x] __rpow__ (straight up calls torch.tensor)
• [x] torch.tensor
• [x] Tensor.new()

Data pointer accessed by helper function (#65)

• [ ] linalg.cholesky (linalg_cholesky calls linalg_cholesky_ex (prim) and does error checking)
• [ ] linalg.inv (linalg_inv calls linalg_inv_ex (prim) and does error checking)
• [ ] linalg.matrix_power (can call inv)

The norm problem (#14); AKA: CompositeImplicitAutograd op calls an "out= variant" that calls raw native::resize_ on tensors.

• [ ] linalg.matrix_norm
• [ ] linalg.norm
• [ ] nanquantile
• [ ] quantile

Requires an integer tensor for the "splits" argument...

• [ ] tensor_split

Test by uncommenting out https://github.com/zou3519/functorch/blob/ae97def8eb8508418053a1a7c81371b9b44dcc3d/test/test_grad.py#L49. I haven't investigated the problems yet.

Miscellaneous non-OpInfo problems (test_torch.py)

• [ ] Tensor.numpy
• [ ] Tensor.tolist
• [ ] copy.copy
• [ ] to_dlpack
• [ ] repeat_interleave
• [ ] Tensor.map_
• [ ] Tensor.map2_
• [ ] pickle.dumps
• [ ] printing
• [ ] torch._sobol_engine_initializestate
• [ ] assigning to Tensor.data

Miscellaneous non-OpInfo problems (test_nn.py)

• [ ] F.ctc_loss
• [ ] F.max_pool1d (testing artifact)
• [ ] Lazy modules

Miscellaneous non-OpInfo problems (test_linalg.py)

• [ ] linalg.cond (https://gist.github.com/zou3519/2c6a820bc3ef0b3a9c3107d2c26fb064)
• [ ] linalg.eigvals
• [ ] linalg.eigvalsh
• [ ] torch.inverse
• [ ] torch.kron
• [ ] linalg.matrix_rank
• [ ] linalg.nuclear_norm
• [ ] linalg.pinv
• [ ] torch.svd
• [ ] linalg.tensorinv
• [ ] linalg.tensorsolve

Miscellaneous non-OpInfo problems (test_tensor_creation.py)

• [ ] expected = torch.tensor(list(product([a], b, c)), device=device) https://gist.github.com/zou3519/742b36c9eabb41449f049e6bc3ab1410

Miscellaneous non-OpInfo problems test_unary_ufuncs.py

• [ ] conj

https://docs.google.com/spreadsheets/d/18sv-cKBqMGVCNdclFk5jB9LmQJGzb_eNAE9O2-oep3Q/edit?usp=sharing

[zou3519]

From discussion with Horace:

internal_new_from_data: allowing data_ptr access for TensorWrapper that doesn't wrap BatchedTensor sounds fine

What does XLA do? Do they not support this?

TODO: bring up with composability group somehow