skip test nvfp4_gemm on unsupported arches

[liqiangxl]

Same as #5810
err msg Exception raised from runGemm at /opt/pytorch/nvfuser/cutlass/nvfp4_scaled_mm.cu:255

[github-actions]

Auto-merge Status

✅ Internal CI is finished
✅ No failed checks
✅ PR is mergeable
ℹ️ PR mergeable_state: clean

Description

• Replace manual compute capability checking with dedicated microarchitecture_is utility function

• Update skip condition to only test nvfp4_gemm on compute capability 10.0

• Improve skip reason message for better clarity about architecture requirements

Changes walkthrough

	Relevant files
Bug fix	test_cutlass_nvfp4_gemm.py Update nvfp4_gemm test skip logic with utility function    tests/python/direct/test_cutlass_nvfp4_gemm.py Import microarchitecture_is utility function from python.direct_utils Replace manual compute capability check with microarchitecture_is(10, 0) condition Update skip reason to specify compute capability 10.0 requirement +3/-3

PR Reviewer Guide

Here are some key observations to aid the review process:

🧪 PR contains tests

⚡ Recommended focus areas for review

Architecture Restriction Change The change modifies the compute capability check from a range-based approach (>= 10.0 and < 12.0) to an exact match (== 10.0). This significantly narrows the supported architectures and may exclude valid compute capabilities like 10.1, 10.2, 11.0, 11.1 that were previously supported. The rationale for this restriction change should be validated, especially given the referenced error from nvfp4_scaled_mm.cu:255. if not microarchitecture_is(10, 0): pytest.skip( reason="Nvfp4 Requires compute capability 10.0, other arches are not tested.", allow_module_level=True, ) Error Root Cause Analysis The PR description mentions an exception from runGemm at nvfp4_scaled_mm.cu:255, but doesn't provide details about the root cause. The change to restrict to only compute capability 10.0 may be a workaround rather than a proper fix. The underlying issue causing failures on other architectures should be investigated to determine if this restriction is appropriate or if a more comprehensive solution is needed. if not microarchitecture_is(10, 0): pytest.skip( reason="Nvfp4 Requires compute capability 10.0, other arches are not tested.", allow_module_level=True, )

[liqiangxl]

!build

[greptile-apps]

Greptile Summary

This PR restricts nvfp4 GEMM tests to run only on compute capability 10.0 devices, preventing runtime errors on unsupported architectures. The change follows the same pattern as PR #5810 which fixed similar issues for mxfp8 GEMM tests.

Changes:

• Replaced manual compute capability check with microarchitecture_is(10, 0) helper function
• Updated skip message to clarify that only 10.0 is tested
• Made the check more restrictive: old logic allowed 10.0-11.x, new logic allows only 10.0

Note: This is more restrictive than the previous implementation which allowed compute capabilities 10.x up to (but not including) 12.0. The commit message confirms this is intentional ("only test nvfp4_gemm on 10.0 device").

Confidence Score: 5/5

• This PR is safe to merge with minimal risk
• The change is a straightforward test infrastructure update that prevents test failures on unsupported hardware. It follows an established pattern from PR skip test cutlass mxfp8_gemm on unsupported arches #5810, uses an existing utility function correctly, and only affects test execution (not production code). The more restrictive logic is intentional as confirmed by the commit message.
• No files require special attention

Important Files Changed

Filename	Overview
tests/python/direct/test_cutlass_nvfp4_gemm.py	Updated architecture check to restrict testing to compute capability 10.0 only, consistent with PR #5810 pattern

Sequence Diagram

sequenceDiagram
    participant Test as test_cutlass_nvfp4_gemm.py
    participant Utils as direct_utils.microarchitecture_is
    participant CUDA as torch.cuda
    participant Pytest as pytest.skip

    Note over Test: Module import phase
    Test->>Utils: microarchitecture_is(10, 0)
    Utils->>CUDA: get_device_properties(current_device())
    CUDA-->>Utils: device properties (major, minor)
    Utils->>Utils: Check: major == 10 and minor == 0
    alt Not compute capability 10.0
        Utils-->>Test: False
        Test->>Pytest: skip(reason="...", allow_module_level=True)
        Note over Pytest: Tests skipped for this module
    else Compute capability 10.0
        Utils-->>Test: True
        Note over Test: Proceed with test execution
        Test->>Test: test_nvfp4_gemm()
        Test->>Test: test_nvfp4_gemm_epilogue()
        Test->>Test: test_nvfp4_grouped_mm()
    end

Loading