Commit 05389780 for whisper.cpp

commit 0538978040072fea7a4c871b28b027f70c623e1b
Author: Yonghye Kwon <developer.0hye@gmail.com>
Date:   Wed Oct 28 17:13:01 2026 +0900

    cuda : add top-k MoE fusion for 384 experts (llama/31288)

    Assisted-by: Codex

diff --git a/ggml/src/ggml-cuda/topk-moe.cu b/ggml/src/ggml-cuda/topk-moe.cu
index 158b736f..a93d3a97 100644
--- a/ggml/src/ggml-cuda/topk-moe.cu
+++ b/ggml/src/ggml-cuda/topk-moe.cu
@@ -402,6 +402,10 @@ static void launch_topk_moe_cuda(ggml_backend_cuda_context & ctx,
             ggml_cuda_kernel_launch(topk_moe_cuda<288, has_bias>, launch_params,
                 logits, weights, ids, bias, n_rows, n_expert_used, clamp_val, scale_val, config);
             break;
+        case 384: // DeepSeek V4.1
+            ggml_cuda_kernel_launch(topk_moe_cuda<384, has_bias>, launch_params,
+                logits, weights, ids, bias, n_rows, n_expert_used, clamp_val, scale_val, config);
+            break;
         case 512:
             ggml_cuda_kernel_launch(topk_moe_cuda<512, has_bias>, launch_params,
                 logits, weights, ids, bias, n_rows, n_expert_used, clamp_val, scale_val, config);
@@ -472,7 +476,7 @@ bool ggml_cuda_should_use_topk_moe(const ggml_tensor * gating_op,
     // must match an instantiation of launch_topk_moe_cuda: a power of 2 up to 512,
     // or one of the non-power-of-2 expert counts of supported models
     const int n_expert = ids->nb[1] / ids->nb[0];
-    if (((n_expert & (n_expert - 1)) != 0 || n_expert > 512) && n_expert != 288 && n_expert != 576) {
+    if (((n_expert & (n_expert - 1)) != 0 || n_expert > 512) && n_expert != 288 && n_expert != 384 && n_expert != 576) {
         return false;
     }