fmha_multirow: add fence.sc.gpu after QK GEMM for TMEM visibility

This commit is contained in:
2026-05-28 23:13:31 +00:00
parent d840fbbf85
commit 43ba672e15

View File

@@ -164,6 +164,10 @@ fmha_6warp_multirow_kernel(FmhaMultiRowParams params) {
__syncthreads();
}
// Ensure TMEM visibility across all warps after QK GEMM
asm volatile("fence.sc.gpu;" ::: "memory");
__syncthreads();
// ================================================================
// SOFTMAX PASS 1: compute row_max for each row
// ================================================================