CUDA: fix bug in topk-moe softmax (#16711)

2025-10-22 12:33:08 +08:00 · 2025-10-22 12:33:08 +08:00 · 9285325ce0
parent 03792ad936
commit 9285325ce0
1 changed files with 1 additions and 1 deletions
--- 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;
        }
    }