when we call benchmarker.benchmark(fn, (), {}) it attempts to infer the device from the args and kwargs, which are both empty. in this case the default behavior is to assume CPU, since `is_cpu_device` is implemented as `all([x.device == "cpu" for x in ... if x is Tensor])`, and `all([]) == True`. I've added a PR that makes this raise an error, but we should just fix this one callsite first
Pull Request resolved: https://github.com/pytorch/pytorch/pull/133290
Approved by: https://github.com/eellison
move benchmarking out of `torch._inductor.runtime.runtime_utils` and into `torch._inductor.runtime.benchmarking`, and prefer this path over directly accessing Triton's benchmarking
Fixes #ISSUE_NUMBER
Pull Request resolved: https://github.com/pytorch/pytorch/pull/132827
Approved by: https://github.com/eellison
The test fail internally [T195592444](https://www.internalfb.com/intern/tasks/?t=195592444) (This is meta internal link). But we don't see the failure in OSS.
It turns out that there are 2 issues:
1. `run_test('cuda')` is improperly handled since it tries to import a module named 'cuda' if cuda is available. Since the import fails, all tests in the file are skipped. This hides the failure in OSS. The failure is exposed in internal tests since the main block which runs `run_test('cuda')` is skipped sometimes.
2. fix the real issue that incompatible inputs are provided to `do_bench`.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/131402
Approved by: https://github.com/eellison
