From a435c7725b17fffe0045c2d4b7e38fe9f42d4070 Mon Sep 17 00:00:00 2001
From: Beinsezii <beinsezii@gmail.com>
Date: Thu, 1 Jan 2026 21:50:51 -0800
Subject: [PATCH] mmq.cu: tune mmq/wmma switching for RDNA

---
 ggml/src/ggml-cuda/mmq.cu | 22 +++++++++++++++-------
 1 file changed, 15 insertions(+), 7 deletions(-)

diff --git a/ggml/src/ggml-cuda/mmq.cu b/ggml/src/ggml-cuda/mmq.cu
index 0a0d440a18..ccb9ebed57 100644
--- a/ggml/src/ggml-cuda/mmq.cu
+++ b/ggml/src/ggml-cuda/mmq.cu
@@ -333,16 +333,24 @@ bool ggml_cuda_should_use_mmq(enum ggml_type type, int cc, int64_t ne11, int64_t
     }
 
     if (amd_wmma_available(cc)) {
-        if (n_experts > 64 || ne11 <= 128) {
+        // High expert counts almost always better on MMQ
+        // due to a large amount of graph splits
+        // https://github.com/ggml-org/llama.cpp/pull/18202
+        if (n_experts >= 64) {
             return true;
         }
-        if (type == GGML_TYPE_Q4_0 || type == GGML_TYPE_Q4_1 || type == GGML_TYPE_Q5_0 || type == GGML_TYPE_Q5_1) {
-            return true;
+
+        switch (type) {
+            // These quants are really bad on MMQ
+            case GGML_TYPE_Q2_K:
+            case GGML_TYPE_Q6_K:
+            // These quants are usually worse but not always
+            case GGML_TYPE_IQ2_XS:
+            case GGML_TYPE_IQ2_S:
+                return ne11 <= 128;
+            default:
+                return true;
         }
-        if (ne11 <= 256 && (type == GGML_TYPE_Q4_K || type == GGML_TYPE_Q5_K)) {
-            return true;
-        }
-        return false;
     }
 
     return (!GGML_CUDA_CC_IS_CDNA(cc)) || ne11 < MMQ_DP4A_MAX_BATCH_SIZE;