• wenjh's avatar
    [Workaround] Force NVTE_FORCE_ROCM_GEMM=1 · 6dfe66e9
    wenjh authored
    
    
    The acc problem in test_grouped_linear_accuracy and test_grouped_gemm is
    because calc test out and ref out using diff kernel.
    Make NVTE_FORCE_ROCM_GEMM=1 can force these tests to call rocm gemm using
    same kernel.
    Signed-off-by: wenjh's avatarwenjh <wenjh@sugon.com>
    6dfe66e9
test_numerics.py 76.6 KB