From: Jeff Bolz <redacted>
Date: Fri, 5 Dec 2025 20:21:57 +0000 (-0600)
Subject: vulkan: enable mmvq for q2_k on NVIDIA (llama/17675)
X-Git-Tag: upstream/0.9.4.395~60
X-Git-Url: https://git.djapps.eu/?a=commitdiff_plain;h=04deeb820f3d1981b1c0cb5a6436ebc6479c104a;p=pkg%2Fggml%2Fsources%2Fggml

vulkan: enable mmvq for q2_k on NVIDIA (llama/17675)
---

diff --git a/src/ggml-vulkan/ggml-vulkan.cpp b/src/ggml-vulkan/ggml-vulkan.cpp
index 97b0fd74..c8f0449a 100644
--- a/src/ggml-vulkan/ggml-vulkan.cpp
+++ b/src/ggml-vulkan/ggml-vulkan.cpp
@@ -6948,6 +6948,10 @@ static bool ggml_vk_should_use_mmvq(const vk_device& device, uint32_t m, uint32_
     // Quantization overhead is not worth it for small k
     switch (device->vendor_id) {
     case VK_VENDOR_ID_NVIDIA:
+        if (src0_type == GGML_TYPE_Q2_K) {
+            return true;
+        }
+
         if (k <= 4096) {
             return false;
         }