From: ravel7524 <redacted>
Date: Wed, 20 May 2026 01:52:21 +0000 (+0200)
Subject: ggml-cuda: tune RDNA3 Q6_K MMVQ nwarps (#23349)
X-Git-Tag: upstream/0.0.10438~1193
X-Git-Url: https://git.djapps.eu/?a=commitdiff_plain;h=b39a7bf1b038e8d1c3d58a313228dfc4f3993f5c;p=pkg%2Fggml%2Fsources%2Fllama.cpp

ggml-cuda: tune RDNA3 Q6_K MMVQ nwarps (#23349)
---

diff --git a/ggml/src/ggml-cuda/mmvq.cu b/ggml/src/ggml-cuda/mmvq.cu
index da48f313a..73a0991e2 100644
--- a/ggml/src/ggml-cuda/mmvq.cu
+++ b/ggml/src/ggml-cuda/mmvq.cu
@@ -359,7 +359,9 @@ static constexpr __host__ __device__ int calc_nwarps(ggml_type type, int ncols_d
                 case GGML_TYPE_Q5_1:
                 case GGML_TYPE_Q8_0:
                 case GGML_TYPE_Q4_K:
+                    return 8;
                 case GGML_TYPE_Q6_K:
+                    return 2;
                 case GGML_TYPE_IQ4_NL:
                     return 8;
                 default: