From c590c9179205bc8e9a4440099579a696e6df321c Mon Sep 17 00:00:00 2001
From: Mike Lothian <mike@fireburn.co.uk>
Date: Thu, 1 Oct 2026 09:59:47 +0100
Subject: [PATCH] vulkan : use the float mat-vec for single tokens with q4_K,
 q5_K and iq4_xs on RDNA4

Assisted-by: Claude Code (Claude Opus 5.5)
---
 ggml/src/ggml-vulkan/ggml-vulkan.cpp | 5 +++++
 1 file changed, 5 insertions(+)

diff --git a/ggml/src/ggml-vulkan/ggml-vulkan.cpp b/ggml/src/ggml-vulkan/ggml-vulkan.cpp
index 8a89d320d..4bdec5114 100644
--- a/ggml/src/ggml-vulkan/ggml-vulkan.cpp
+++ b/ggml/src/ggml-vulkan/ggml-vulkan.cpp
@@ -6581,6 +6581,11 @@ static bool ggml_vk_should_use_mmvq(const vk_device& device, uint32_t m, uint32_
         switch (src0_type) {
         case GGML_TYPE_Q8_0:
             return device->architecture == vk_device_architecture::AMD_GCN;
+        // single tokens are faster with the float path on RDNA4
+        case GGML_TYPE_Q4_K:
+        case GGML_TYPE_Q5_K:
+        case GGML_TYPE_IQ4_XS:
+            return device->architecture != vk_device_architecture::AMD_RDNA4;
         default:
             return true;
         }
-- 
2.56.0

