Fix group scale gemm when K==128 (#2275)

Co-authored-by: Jiazhen Han <jiazhenh@nvidia.com>
2025-05-02 12:41:18 -07:00
parent f535c33634
commit 89f6bf2739
1 changed files with 1 additions and 1 deletions
--- a/include/cutlass/gemm/collective/sm90_mma_tma_gmma_ss_warpspecialized_fp8_blockwise_scaling.hpp
+++ b/include/cutlass/gemm/collective/sm90_mma_tma_gmma_ss_warpspecialized_fp8_blockwise_scaling.hpp
@@ -1001,7 +1001,7 @@ struct CollectiveMma<
      // Advance smem_pipe_read and smem_pipe_release
      ++smem_pipe_release;
    }
-    {
+    if (k_tile_count == 1) {
      pipeline.consumer_wait(smem_pipe_read, barrier_token);

      //