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

diff --git a/ggml/src/ggml-vulkan/ggml-vulkan.cpp b/ggml/src/ggml-vulkan/ggml-vulkan.cpp
index 97b0fd742e0c071c11cc7cdb468a01fc83f4e695..c8f0449a94aa4d5a291750f10500edaeba5e6646 100644 (file)
--- a/ggml/src/ggml-vulkan/ggml-vulkan.cpp
+++ b/ggml/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;
         }