CUDA: fix race condition in MMQ stream-k fixup (#13299)

JohannesGaessler · web-flow · commit 93c4e2390598 · 2025-05-04T14:16:39.000+02:00
diff --git a/ggml/src/ggml-cuda/mmq.cuh b/ggml/src/ggml-cuda/mmq.cuh
@@ -2958,6 +2958,7 @@ static __global__ void mul_mat_q_stream_k_fixup(
     for (int j = threadIdx.y*WARP_SIZE + threadIdx.x; j < mmq_x; j += nwarps*WARP_SIZE) {
         ids_dst_shared[j] = ids_dst[col_low + j];
     }
+    __syncthreads();
 
     const int offset_dst = it*mmq_y;
     dst += offset_dst;

Original file line number	Diff line number	Diff line change
`@@ -2958,6 +2958,7 @@ static __global__ void mul_mat_q_stream_k_fixup(`
`2958`	`2958`	`for (int j = threadIdx.yWARP_SIZE + threadIdx.x; j < mmq_x; j += nwarpsWARP_SIZE) {`
`2959`	`2959`	`ids_dst_shared[j] = ids_dst[col_low + j];`
`2960`	`2960`	`}`
	`2961`	`+ __syncthreads();`
`2961`	`2962`
`2962`	`2963`	`const int offset_dst = it*mmq_y;`
`2963`	`2964`	`dst += offset_dst;`