LLM inference in C/C++
C++ ggml
HIP: broaden MoE ncols_opt tile heuristic on RDNA3.5 architecture (#28935)
WenqiangJia2026fccf7166fb4c
diff --git a/ggml/src/ggml-cuda/mmq.cu b/ggml/src/ggml-cuda/mmq.cuindex 9b6038adff..b13b34ee9c 100644--- a/ggml/src/ggml-cuda/mmq.cu+++ b/ggml/src/ggml-cuda/mmq.cu@@ -247,7 +247,7 @@ void ggml_cuda_mul_mat_q(// Each expert only sees ne12*n_expert_used/ne02 tokens on average.// On RDNA3 and RDNA4 it is faster to pick the tile size against this value instead of ne12.int64_t ncols_opt = ne12;- if (GGML_CUDA_CC_IS_RDNA3_0(cc) || GGML_CUDA_CC_IS_RDNA4(cc)) {+ if (GGML_CUDA_CC_IS_RDNA3(cc) || GGML_CUDA_CC_IS_RDNA4(cc)) {ncols_opt = (ne12*n_expert_used + ne02 - 1) / ne02;}