CUDA: fix MMQ stream-k for --split-mode row (llama/8167)

diff --git a/src/ggml-cuda/mmq.cuh b/src/ggml-cuda/mmq.cuh
index 31fcbf1397b6bb15475af811299efc8064006dd3..1396e7a753ac34175c93faac19459fc38868c44b 100644 (file)
--- a/src/ggml-cuda/mmq.cuh
+++ b/src/ggml-cuda/mmq.cuh
@@ -2475,7 +2475,7 @@ static void launch_mul_mat_q(ggml_backend_cuda_context & ctx, const mmq_args & a
 
     const dim3 block_nums_mmq(nsm, 1, 1);
 
-    ggml_cuda_pool & pool = ctx.pool();
+    ggml_cuda_pool & pool = ctx.pool(id);
     ggml_cuda_pool_alloc<float> tmp_fixup(pool, block_nums_mmq.x * mmq_x*mmq_y);
 
     if (args.ne01 % mmq_y == 0) {