From 6d917540ae2f492156646f3f7dd3c205ecf074a2 Mon Sep 17 00:00:00 2001
From: Aman Gupta <amangupta052@gmail.com>
Date: Wed, 22 Oct 2025 10:42:43 +0800
Subject: [PATCH] CUDA: fix bug in topk-moe softmax

---
 ggml/src/ggml-cuda/topk-moe.cu | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/ggml/src/ggml-cuda/topk-moe.cu b/ggml/src/ggml-cuda/topk-moe.cu
index d782ad948d254..e28c810ac5df7 100644
--- a/ggml/src/ggml-cuda/topk-moe.cu
+++ b/ggml/src/ggml-cuda/topk-moe.cu
@@ -141,7 +141,7 @@ __launch_bounds__(4 * WARP_SIZE, 1) __global__ void topk_moe_cuda(const float *
         wt_sum              = warp_reduce_sum(wt_sum);
         const float inv_sum = 1.0f / wt_sum;
 
-        for (int i = threadIdx.x; i < n_expert_used; i += WARP_SIZE) {
+        for (int i = 0; i < experts_per_thread; i++) {
             output_weights[i] *= inv_sum;
         }
     }