Add shim.h C API to call dispatcher on our own aten ops #148832
Conversation
[ghstack-poisoned]
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/148832
Note: Links to docs will display an error until the docs builds have been completed. ⏳ 9 Pending, 2 Unrelated FailuresAs of commit 9f133f4 with merge base c983e11 ( BROKEN TRUNK - The following jobs failed but were present on the merge base:👉 Rebase onto the `viable/strict` branch to avoid these failures
This comment was automatically generated by Dr. CI and updates every 15 minutes. |
Attention! PyTorch one of the C-stable API file was changedYou MUST NOT change existing function declarations in this, as this header defines a stable C ABI. If you need to change the signature for a function, introduce a new v2 version of the function and modify code generation to target the new version of the function. Caused by: |
This PR still needs testing through some cpp extension [ghstack-poisoned]
This PR still needs testing through some cpp extension [ghstack-poisoned]
This PR still needs testing through some cpp extension [ghstack-poisoned]
This PR still needs testing through some cpp extension [ghstack-poisoned]
This PR still needs testing through some cpp extension [ghstack-poisoned]
This PR still needs testing through some cpp extension [ghstack-poisoned]
This PR still needs testing through some cpp extension [ghstack-poisoned]
test/test_cpp_extensions_aot.py
Outdated
| self.assertEqual(cpu_t, torch.ones_like(t, device="cpu")) | ||
|
|
||
| def _make_cuda_tensors(prior_mem): | ||
| cuda_t = libtorch_agnostic.ops.my_ones_like(t, "cuda") |
There was a problem hiding this comment.
note, passing in the device obj here does not work
| cuda_t = libtorch_agnostic.ops.my_ones_like(t, "cuda") | |
| cuda_t = libtorch_agnostic.ops.my_ones_like(t, device) |
There was a problem hiding this comment.
Why not? Device is not a supported type for the stable dispatcher registration?
test/cpp_extensions/libtorch_agnostic_extension/test/test_libtorch_agnostic.py
Outdated
Show resolved
Hide resolved
...p_extensions/libtorch_agnostic_extension/libtorch_agnostic/csrc/libtorch_agnostic_kernel.cpp
Outdated
Show resolved
Hide resolved
This PR still needs testing through some cpp extension [ghstack-poisoned]
test/test_cpp_extensions_aot.py
Outdated
| self.assertEqual(cpu_t, torch.ones_like(t, device="cpu")) | ||
|
|
||
| def _make_cuda_tensors(prior_mem): | ||
| cuda_t = libtorch_agnostic.ops.my_ones_like(t, "cuda") |
There was a problem hiding this comment.
Why not? Device is not a supported type for the stable dispatcher registration?
| auto inner_type = arg_type->castRaw<at::OptionalType>()->getElementType(); | ||
|
|
||
| // our contract is that IValue None = StableIValue nullptr | ||
| if (to<std::nullptr_t>(stable_ivalue) == nullptr) { |
There was a problem hiding this comment.
I don't think that's ok, you are definitely going to confuse a value 0 with empty optional here.
I think the simplest would be to do some tagging here: have a special value that means "nullopt" and forbid the content of the optional being that value (with a runtime check).
My guess here is that you want something that is unlikely to be a used int64 or pointer or etc.
There was a problem hiding this comment.
to respect the branch cut, i got rid of supporting optional for now
This PR still needs testing through some cpp extension [ghstack-poisoned]
This PR still needs testing through some cpp extension [ghstack-poisoned]
This PR still needs testing through some cpp extension [ghstack-poisoned]
This PR still needs testing through some cpp extension [ghstack-poisoned]
| // calls the op overload defined by a given opName, overloadName, and a | ||
| // stack of StableIValues. This call will populate any return values of the | ||
| // op into the stack in their StableIValue form, with ret0 at index 0, ret1 | ||
| // at index 1, and so on. |
There was a problem hiding this comment.
todo after this: we should document our StableIValue stack definition and expose to/from helpers for it.
There was a problem hiding this comment.
I actually did that, just in library.h https://github.com/pytorch/pytorch/blob/main/torch/csrc/stable/library.h#L69
|
@pytorchbot merge |
Merge startedYour change will be merged once all checks pass (ETA 0-4 Hours). Learn more about merging in the wiki. Questions? Feedback? Please reach out to the PyTorch DevX Team |
Merge failedReason: 1 jobs have failed, first few of them are: inductor / cuda12.6-py3.10-gcc9-sm86 / test (inductor_timm, 1, 2, linux.g5.4xlarge.nvidia.gpu) Details for Dev Infra teamRaised by workflow job |
|
@pytorchbot merge -f "Failures are preexisting on trunk" |
Merge startedYour change will be merged immediately since you used the force (-f) flag, bypassing any CI checks (ETA: 1-5 minutes). Please use Learn more about merging in the wiki. Questions? Feedback? Please reach out to the PyTorch DevX Team |
This PR still needs testing through some cpp extension
Stack from ghstack (oldest at bottom):