fmha_multirow: add fence.sc.gpu after QK GEMM for TMEM visibility
This commit is contained in:
@@ -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
|
||||
// ================================================================
|
||||
|
||||
Reference in New Issue
Block a user