Centralize addition of "always on" dispatch keys.

[ezyang]

Stack from ghstack:

• Add XLAPreAutograd key for XLA use cases that need custom autograd. #32788 Add XLAPreAutograd key for XLA use cases that need custom autograd.
• Stop using dispatchTypeId to do checks for tensor list unwrap. #32787 Stop using dispatchTypeId to do checks for tensor list unwrap.
• Centralize addition of "always on" dispatch keys. #32734 Centralize addition of "always on" dispatch keys.
• Make DispatchKeyGuards accept DispatchKey::Undefined #32729 Make DispatchKeyGuards accept DispatchKey::Undefined
• Rename DispatchKey::UndefinedTensorId to Undefined #32728 Rename DispatchKey::UndefinedTensorId to Undefined

VariableTensorId is the only key with this treatment today,
but BackendSelect and CompoundOp are coming soon.

Signed-off-by: Edward Z. Yang ezyang@fb.com

Differential Revision: D19628091

[kostmo]

💊 CircleCI build failures summary and remediations

As of commit fd13557:

• 1/1 failures introduced in this PR

Detailed failure analysis

One may explore the probable reasons each build failed interactively on the Dr. CI website.

🕵️ 1 new failure recognized by patterns

The following build failures do not appear to be due to upstream breakage:

pytorch_macos_10_13_py3_test (1/1)

Step: "Test" (full log | pattern match details)

Jan 30 09:19:53 RuntimeError: distributed/test_distributed failed!

Jan 30 09:19:53 Ran 94 tests in 345.913s 
Jan 30 09:19:53  
Jan 30 09:19:53 FAILED (failures=1, skipped=20) 
Jan 30 09:19:53  
Jan 30 09:19:53 Generating XML reports... 
Jan 30 09:19:53 Traceback (most recent call last): 
Jan 30 09:19:53   File "test/run_test.py", line 457, in <module> 
Jan 30 09:19:53     main() 
Jan 30 09:19:53   File "test/run_test.py", line 450, in main 
Jan 30 09:19:53     raise RuntimeError(message) 
Jan 30 09:19:53 RuntimeError: distributed/test_distributed failed! 
Jan 30 09:19:53 + cleanup 
Jan 30 09:19:53 + retcode=1 
Jan 30 09:19:53 + set +x 

---

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.

This comment has been revised 11 times.

[izdeby]

This PR made me think of a centralized place for fallback registrations as the number of DispatchKeys that are available is growing. Do we need a mechanism like that? OR its too soon to make this call as there are only 2 keys like that - variable and backendSelect?

[ezyang]

This PR made me think of a centralized place for fallback registrations as the number of DispatchKeys that are available is growing. Do we need a mechanism like that? OR its too soon to make this call as there are only 2 keys like that - variable and backendSelect?

Yeah, I dunno. It's not strictly necessary to centralize registrations; e.g., CPU/CUDA are registered in CPUType/CUDAType, and then Variable is registered in VariableType. But I agree it's not that clear what the most obvious place to put registrations is.

[facebook-github-bot]

@ezyang merged this pull request in 690d41f.

[facebook-github-bot]

@ezyang merged this pull request in 690d41f.