Port to new registration API (part 1)

[ezyang]

Stack from ghstack:

• Put TORCH_LIBRARY in torch/library.h; add custom class API #36742 Add custom class API to TORCH_LIBRARY API.
• Port to new registration API (part 1) #36389 Port to new registration API (part 1)

This is a roll up of a bunch of small PRs for ease of landing.

• Update reference to RegisterOperators in error message in Convolution. Port to new registration API (part 1) #36389
• Port Resize to use new registration API. Port Resize to use new registration API. #36390
• Port detach/detach to new registration API. Port detach/detach to new registration API #36512
• Add explicit schema for quantized conv/conv_prepack (fixes quantized::conv2d and variants don't have explicit schema #36511). Add explicit schema for quantized conv/conv_prepack #36513
• Add a centralized TORCH_LIBRARY declaration for quantized and xnnpack ops (fixes Meaning of _quantized namespace isn't documented #36510). Add a centralized TORCH_LIBRARY declaration for quantized and xnnpack ops. #36520
• Functionalize qadd and register with new registration API. Functionalize qadd and register with new registration API. #36527
• Update quantized README for registering operators with new API. Update quantized README for registering operators with new API. #36531
• Convert qbatch_norm to new operator registration API. Convert qbatch_norm to new operator registration API. #36534
• Convert qclamp to new operator registration API. Convert qclamp to new operator registration API. #36535
• Functionalize qconcat and register with new registration API. Functionalize qconcat and register with new registration API #36536

Differential Revision: D20964193

[dr-ci]

💊 Build failures summary and remediations

As of commit 388f0c0 (more details on the Dr. CI page):

---

💚 💚 Looks good so far! There are no failures yet. 💚 💚

---

This comment was automatically generated by Dr. CI (expand for details).

Follow this link to opt-out of these comments for your Pull Requests.

Please report bugs/suggestions on the GitHub issue tracker.

See how this bot performed.

This comment has been revised 74 times.

[dzhulgakov]

what's going on here? why do you need to put it into BC test?

[ezyang]

Oh, this is mistake from when I typoed conv_prepack for an registration, and that lit up the bc check, and then when I fixed the typo I didn't fix the registration

[ezyang]

I need this in the BC test because I made a (technical) BC breaking change.

Previously, conv_prepack was registered with inferred schema:

-        .op("quantized::conv_prepack", // conv_prepack is deprecated, please use
-                                       // conv2d_prepack for 2D conv.

With this PR, it is now done with explicit schema:

+  // conv_prepack is deprecated, please use conv2d_prepack for 2D conv.
+  m.def("conv_prepack(Tensor weight, Tensor? bias, int[] stride, int[] padding, int[] dilation, int groups) -> Tensor");

Moving from implicit to explicit is considered bc breaking by the bc script, thus I must whitelist it.

[kostmo]

Appears to have caused the following in master:

Apr 17 00:15:27 ======================================================================
Apr 17 00:15:27 ERROR [0.374s]: test_batch_norm2d (__main__.TestQuantizedOps)
Apr 17 00:15:27 ----------------------------------------------------------------------
Apr 17 00:15:27 Traceback (most recent call last):
Apr 17 00:15:27   File "quantization/test_quantized.py", line 1534, in test_batch_norm2d
Apr 17 00:15:27     min_side=1, max_side=32),
Apr 17 00:15:27   File "/var/lib/jenkins/.local/lib/python3.7/site-packages/hypothesis/core.py", line 1116, in wrapped_test
Apr 17 00:15:27     raise the_error_hypothesis_found
Apr 17 00:15:27   File "quantization/test_quantized.py", line 1554, in test_batch_norm2d
Apr 17 00:15:27     qy = torch.ops.quantized.batch_norm2d(qx, weight, bias, mean, var, eps, Y_scale, Y_zero_point)
Apr 17 00:15:27   File "/opt/python/nightly/lib/python3.7/site-packages/torch/_ops.py", line 61, in __getattr__
Apr 17 00:15:27     op = torch._C._jit_get_operation(qualified_op_name)

Reverting.

[facebook-github-bot]

@ezyang merged this pull request in 6615886.

[ezyang]

Stunning, the local ci was all green! XD