From 3326fa2387e58b4650df288dd0bf6fee12e01205 Mon Sep 17 00:00:00 2001 From: Beinsezii Date: Fri, 2 Jan 2026 15:05:30 -0800 Subject: [PATCH] mmq.cu: move amd wmma mmq/wmma switching behind IS_RDNA3 --- ggml/src/ggml-cuda/mmq.cu | 38 ++++++++++++++++++++++---------------- 1 file changed, 22 insertions(+), 16 deletions(-) diff --git a/ggml/src/ggml-cuda/mmq.cu b/ggml/src/ggml-cuda/mmq.cu index ccb9ebed57..99d1a7b833 100644 --- a/ggml/src/ggml-cuda/mmq.cu +++ b/ggml/src/ggml-cuda/mmq.cu @@ -333,23 +333,29 @@ bool ggml_cuda_should_use_mmq(enum ggml_type type, int cc, int64_t ne11, int64_t } if (amd_wmma_available(cc)) { - // 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; - } - - 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: + // RDNA 4 is consistently worse on rocblas + // https://github.com/ggml-org/llama.cpp/pull/18537#issuecomment-3706422301 + if (GGML_CUDA_CC_IS_RDNA3(cc)) { + // 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; + } + + 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; + } + } else { + return true; } }