Gate MX/NVFP4 safetensors test configs on sm100

[alanhuangyoo]

test/prototype/safetensors/test_safetensors_support.py builds _ALL_TEST_CONFIGS and adds the MX / NVFP4 entries based on the torch version only:

    # MX and NVFP4 configs require torch >= 2.11
    if torch_version_at_least("2.11.0.dev"):
        _ALL_TEST_CONFIGS += [
            (MXDynamicActivationMXWeightConfig(), False),
            (NVFP4DynamicActivationNVFP4WeightConfig(), False),
        ]

Both kernels need sm100+, but the class only gates on sm89:

@unittest.skipIf(
    torch.cuda.is_available() and not is_sm_at_least_89(), "Need sm89+ for CUDA"
)

So on sm89/sm90 those parametrized cases run and fail instead of being skipped:

E  AssertionError: NVFP4 DYNAMIC mode is only supported on sm100+ machines
   torchao/prototype/mx_formats/inference_workflow.py:301

E  RuntimeError: CUDA error: CUBLAS_STATUS_NOT_SUPPORTED when calling
   `cublasLtMatmulAlgoGetHeuristic(...)`
   torchao/prototype/mx_formats/mx_tensor.py:803

The Int4WeightOnlyConfig block right above already gates on is_sm_at_least_100() for the same reason, and the helper is imported in this file, so this just applies the same guard to the MX/NVFP4 block.

The CPU path is unaffected — test_safetensors and test_safetensors_sharded both start with if device == "cpu": self.skipTest("Need GPU available"), so these configs never executed on CPU.

Before, on sm90:

7 failed, 13 passed, 20 skipped

After:

4 failed, 12 passed, 16 skipped

The 3 fixed cases are the MX/NVFP4 ones. The 4 that remain are ImportError: Requires mslk >= 1.0.0 from Int4WeightOnlyConfig and are unrelated to this change (missing optional dependency in my environment).

ruff check and ruff format --check are clean on the file.

This is not covered by CI: regression_test.yml runs its CUDA job on linux.g5.12xlarge.nvidia.gpu (A10G, sm86), where the class-level is_sm_at_least_89 skip hides the whole file.

Env: torch 2.13.0+cu130, H20 (sm90), single node.

[pytorch-bot]

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/ao/4821

• 📄 Preview Python docs built from this PR

Note: Links to docs will display an error until the docs builds have been completed.

❌ 1 New Failure

As of commit 09d120d with merge base 88d8e73 ():

NEW FAILURE - The following job has failed:

• Run PT2E GPU Tests / test-rocm (ROCM Nightly, linux.rocm.gpu.ecosystem.mi350.1, --pre torch torchvision --index-url ht... / linux-job (gh)
  Could not load credentials from any providers

This comment was automatically generated by Dr. CI and updates every 15 minutes.

[alanhuangyoo]

Label check ran before the module label was added, so it should go green on a re-run.

test-rocm died at setup (29s, no tests ran) and #4820 has the same failure, so it doesn't look related to this change.