id: TC-INFERENCE-004
name: CUBLAS Fallback Verification
suite: inference
priority: 4
timeout: 120000

dependencies:
  - TC-INFERENCE-002

steps:
  - name: Check for CUBLAS errors in logs
    command: cd docker && docker compose logs 2>&1 | grep -i "CUBLAS_STATUS" | grep -v "SUCCESS" | head -10 || echo "No CUBLAS errors"

  - name: Check compute capability detection
    command: cd docker && docker compose logs 2>&1 | grep -iE "compute|capability|cc.*3" | head -10 || echo "No compute capability logs"

  - name: Verify no GPU errors
    command: cd docker && docker compose logs 2>&1 | grep -iE "error|fail" | grep -i gpu | head -10 || echo "No GPU errors"

criteria: |
  CUBLAS should work correctly on Tesla K80 using legacy fallback.

  Expected:
  - No CUBLAS_STATUS_ARCH_MISMATCH errors
  - No CUBLAS_STATUS_NOT_SUPPORTED errors
  - Compute capability 3.7 may be mentioned in debug logs
  - No fatal GPU-related errors

  The K80 uses legacy CUBLAS functions (cublasSgemmBatched)
  instead of modern Ex variants. This should work transparently.

  Accept warnings. Only fail on actual CUBLAS errors.