From: Johannes Gäßler <redacted>
Date: Sat, 8 Nov 2025 07:26:18 +0000 (+0100)
Subject: CUDA: fix MMQ stream-k fixup ne1 indices (llama/17089)
X-Git-Tag: upstream/0.9.4.185~13
X-Git-Url: https://git.djapps.eu/?a=commitdiff_plain;h=30a0c4c981d18cab19410c32aef1b50cefc33e74;p=pkg%2Fggml%2Fsources%2Fggml

CUDA: fix MMQ stream-k fixup ne1 indices (llama/17089)
---

diff --git a/src/ggml-cuda/mmq.cuh b/src/ggml-cuda/mmq.cuh
index c9a07e82..2e133b6b 100644
--- a/src/ggml-cuda/mmq.cuh
+++ b/src/ggml-cuda/mmq.cuh
@@ -3494,7 +3494,7 @@ static __global__ void mul_mat_q_stream_k_fixup(
     const int col_diff = col_high - col_low;
 
     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];
+        ids_dst_shared[j] = ids_dst[col_low + jt*mmq_x + j];
     }
     __syncthreads();