[JAX] Collective Gemm test fixes

[jberchtold-nvidia]

Description

Please include a brief summary of the changes, relevant motivation and context.

Fixes # (issue)

Type of change

• Documentation change (change only to the documentation, either a fix or a new content)
• Bug fix (non-breaking change which fixes an issue)
• New feature (non-breaking change which adds functionality)
• Breaking change (fix or feature that would cause existing functionality to not work as expected)
• Infra/Build change
• Code refactoring

Changes

Please list the changes introduced in this PR:

• Change A
• Change B

Checklist:

• I have read and followed the contributing guidelines
• The functionality is complete
• I have commented my code, particularly in hard-to-understand areas
• I have made corresponding changes to the documentation
• My changes generate no new warnings
• I have added tests that prove my fix is effective or that my feature works
• New and existing unit tests pass locally with my changes

[greptile-apps]

Greptile Summary

This PR fixes two categories of issues in the JAX Collective GEMM test suite: in the shell harness, it adds robust fallback paths for TE_PATH and XML_LOG_DIR, introduces per-test NCCL ID file isolation with proper cleanup, and improves signal handling; in the Python test files, it removes the args.process_id == 0 guard so result assertions run on every worker process rather than only the rank-0 process.

• Shell harness (run_test_cgemm.sh): adds SCRIPT_DIR/REPO_ROOT detection for flexible path resolution, falls back to a mktemp temp dir when XML_LOG_DIR is not writable, creates a per-test NCCL ID subdirectory under a shared CGEMM_NCCL_FILE_BASE_DIR, cleans it up on exit, and splits the trap so INT/TERM correctly returns exit code 130.
• Python tests (test_gemm.py, test_dense_grad.py, test_layernorm_mlp_grad.py): drops the args.process_id == 0 condition from the enable_result_check block so that all worker processes run assert_allclose, providing broader validation coverage since each process holds a full replica of the gathered tensors via PartitionSpec(None).

Confidence Score: 5/5

All changes are targeted test fixes with no impact on production library code; the logic is correct and the test harness improvements make it more robust.

The Python changes correctly drop the process_id == 0 guard because gathered tensors are fully replicated via PartitionSpec(None) before assertions run, so every rank holds identical data. The shell script changes are purely defensive: fallback directories, per-test NCCL ID isolation, and proper signal handling. No production code paths are touched.

No files require special attention; all changes are confined to the collective GEMM test examples directory.

Important Files Changed

Filename	Overview
examples/jax/collective_gemm/run_test_cgemm.sh	Adds robust TE_PATH/XML_LOG_DIR fallbacks, per-test NCCL ID directory isolation with cleanup, and corrected INT/TERM signal handling with exit code 130.
examples/jax/collective_gemm/test_gemm.py	Removes args.process_id == 0 guard so all processes run result assertions; correct since gathered tensors are fully replicated via PartitionSpec(None).
examples/jax/collective_gemm/test_dense_grad.py	Same process_id guard removal as test_gemm.py; assertions now run on all worker ranks.
examples/jax/collective_gemm/test_layernorm_mlp_grad.py	Same process_id guard removal; all ranks now validate layernorm MLP gradient correctness.

Sequence Diagram

sequenceDiagram
    participant SH as run_test_cgemm.sh
    participant P0 as Process 0 (rank 0)
    participant PN as Process N (rank N)

    SH->>SH: mktemp CGEMM_NCCL_FILE_BASE_DIR
    loop For each TEST_CASE
        SH->>SH: mkdir CGEMM_NCCL_FILE_BASE_DIR/TEST_NAME
        SH->>SH: export NVTE_JAX_NCCL_FILE_PATH
        SH->>P0: "pytest --process-id=0"
        SH->>PN: "pytest --process-id=N"
        P0->>P0: compute + gather via PartitionSpec(None)
        PN->>PN: compute + gather via PartitionSpec(None)
        Note over P0,PN: All ranks now run assert_allclose
        P0->>P0: assert_allclose(ref, output)
        PN->>PN: assert_allclose(ref, output)
        SH->>SH: wait and check log for result
    end
    SH->>SH: cleanup and rm CGEMM_NCCL_FILE_BASE_DIR

Loading

_{Reviews (2): Last reviewed commit: "Fix other cgemm tests" | Re-trigger Greptile}

[greptile-apps]

mktemp -d is not guarded against failure. If /tmp is full or unavailable, CGEMM_NCCL_FILE_DIR would be empty/unset and NVTE_JAX_NCCL_FILE_PATH would export a garbage value, causing every subsequent test case to fail with an unintelligible NCCL ID file error.

Suggested change

	CGEMM_NCCL_FILE_DIR=$(mktemp -d /tmp/te_cgemm_nccl_ids.XXXXXX)
	export NVTE_JAX_NCCL_FILE_PATH="${NVTE_JAX_NCCL_FILE_PATH:-$CGEMM_NCCL_FILE_DIR}"
	CGEMM_NCCL_FILE_DIR=$(mktemp -d /tmp/te_cgemm_nccl_ids.XXXXXX) || { echo "ERROR: Failed to create temp dir for NCCL ID files"; exit 1; }
	export NVTE_JAX_NCCL_FILE_PATH="${NVTE_JAX_NCCL_FILE_PATH:-$CGEMM_NCCL_FILE_DIR}"