Back

pytorch #174314

[ROCM] GroupedGemm 2d2d has potential uninitalized data

by alex-minooka · Feb 06, 2026 at 19:22 UTC · scan-f13fc997a07264de

Medium Risk (25%)

Get this automatically on every PR

Install the Axiomo GitHub App to get Signals as check runs and PR comments on every pull request.

Install App

Risk Assessment

Risk level: Medium (25%)

Risk Drivers

  • new_contributor: First contribution from alex-minooka

Intent

2/3 criteria met

Fix uninitialized data issue in ROCm grouped_gemm for small token counts.

Acceptance Criteria

  • Initialize 2d2d shapes to zeros.

    Added initialization to zeros in GroupedMMUtils.h.

  • Skip experts with no tokens.

    Added condition to skip zero-sized groups in ck_group_gemm.hip.

  • ?
    Prevent gradient norm NaNs in torchtitan upstream.

    No direct evidence in diff; requires runtime verification.

Confidence: 90.0% Source: pr description AI: openai

Contributors

alex-minooka PR Author 1 commit ? New Contributor
Account Age: 843 days
Prior PRs: 1

First-time contributor to this repository. unfamiliar with 2 files.

Evidence

Evidence Completeness: 43.0%
ci_passing Pending
tests_passing Failing
build_successful Passing
Missing: lint_passing, security_scan_clean, coverage_maintained

Supply Chain

None Risk
Modifies dependencies
Modifies lockfile
Modifies CI config
Modifies build scripts

Focus Files

Review 2 file(s)

aten/src/ATen/native/GroupedMMUtils.h +7

Source code

medium
aten/src/ATen/native/hip/ck_group_gemm.hip +15

Standard file

low

Triage

9

minutes to review

low

effort level

none

staleness risk

Standard review process

Recommendation

COMMENT 50.0% readiness

Some concerns to address before approval

Next Steps

Question

Why is lint_passing missing? Consider adding this check.

Question

Why is security_scan_clean missing? Consider adding this check.

Nitpick

First contribution - consider welcoming and providing extra context