diff --git a/csrc/libtorch_stable/quantization/w8a8/cutlass/c3x/scaled_mm_blockwise_sm120_fp8_dispatch.cuh b/csrc/libtorch_stable/quantization/w8a8/cutlass/c3x/scaled_mm_blockwise_sm120_fp8_dispatch.cuh index a9008ce442..8795326250 100644 --- a/csrc/libtorch_stable/quantization/w8a8/cutlass/c3x/scaled_mm_blockwise_sm120_fp8_dispatch.cuh +++ b/csrc/libtorch_stable/quantization/w8a8/cutlass/c3x/scaled_mm_blockwise_sm120_fp8_dispatch.cuh @@ -250,7 +250,7 @@ void cutlass_gemm_blockwise_sm120_fp8_dispatch(torch::stable::Tensor& out, torch::stable::Tensor const& b_scales) { int M = a.size(0); // more heuristic tuning can be done here by checking N/K dimensions as well - bool swap_ab = (M <= 64) || (M % 4 != 0); + bool swap_ab = (M <= 64); if (!swap_ab) { if (M <= 256) {