From d08c7872d6ffe3f059f8647840a29aa390413e27 Mon Sep 17 00:00:00 2001 From: Georgi Gerganov Date: Tue, 1 Sep 2026 13:37:40 +0300 Subject: [PATCH] metal : add fa-vec tuning for M2 Max (#28015) Rows for M2 Max (30 GPU cores) collected with 'ggml-metal-tuning fa-vec --dtype f16,q8_0', pasted into fa_vec_tuned_table. ref: https://github.com/ggml-org/llama.cpp/discussions/27668#discussioncomment-18205786 Assisted-by: pi:llama.cpp/Qwen3.8-27B --- ggml/src/ggml-metal/ggml-metal-tuning.cpp | 77 +++++++++++++++++++++++ 1 file changed, 77 insertions(+) diff --git a/ggml/src/ggml-metal/ggml-metal-tuning.cpp b/ggml/src/ggml-metal/ggml-metal-tuning.cpp index 90fdd040d..bf44d5a7d 100644 --- a/ggml/src/ggml-metal/ggml-metal-tuning.cpp +++ b/ggml/src/ggml-metal/ggml-metal-tuning.cpp @@ -1029,6 +1029,83 @@ constexpr fa_vec_entry_t fa_vec_tuned_table[] = { { { GGML_METAL_DEVICE_M2, GGML_TYPE_Q8_0, 576, 512, -1, 0 }, { 1, 4 } }, { { GGML_METAL_DEVICE_M2, GGML_TYPE_Q8_0, 576, 512, -1, 1 }, { 1, 4 } }, + { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_F16, 32, 32, 1, 3 }, { 2, 4 } }, + { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_F16, 32, 32, 2, 3 }, { 2, 4 } }, + { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_F16, 64, 64, 1, 0 }, { 1, 4 } }, + { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_F16, 64, 64, 2, 0 }, { 1, 4 } }, + { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_F16, 64, 64, -1, 1 }, { 1, 4 } }, + { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_F16, 64, 64, 1, 2 }, { 1, 2 } }, + { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_F16, 64, 64, 2, 1 }, { 2, 4 } }, + { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_F16, 64, 64, 3, 2 }, { 1, 2 } }, + { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_F16, 64, 64, 3, 3 }, { 2, 4 } }, + { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_F16, 64, 64, 3, 4 }, { 2, 4 } }, + { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_F16, 96, 96, 2, 1 }, { 2, 4 } }, + { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_F16, 128, 128, 3, 2 }, { 1, 4 } }, + { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_F16, 128, 128, 3, 3 }, { 2, 2 } }, + { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_F16, 192, 192, 3, 0 }, { 1, 4 } }, + { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_F16, 192, 128, 1, 0 }, { 1, 4 } }, + { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_F16, 192, 128, 2, 0 }, { 1, 4 } }, + { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_F16, 192, 128, 1, 1 }, { 1, 4 } }, + { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_F16, 192, 128, 1, 2 }, { 1, 4 } }, + { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_F16, 192, 128, 2, 1 }, { 1, 4 } }, + { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_F16, 192, 128, 2, 2 }, { 1, 4 } }, + { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_F16, 192, 128, 3, 1 }, { 2, 2 } }, + { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_F16, 320, 256, 2, 0 }, { 1, 4 } }, + { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_F16, 320, 256, 3, 0 }, { 1, 4 } }, + { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_F16, 320, 256, 2, 3 }, { 1, 4 } }, + { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_F16, 320, 256, 2, 4 }, { 1, 4 } }, + { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_F16, 512, 512, 2, 0 }, { 4, 4 } }, + { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_F16, 512, 512, 2, 1 }, { 4, 1 } }, + { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_F16, 512, 512, 2, 2 }, { 4, 4 } }, + { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q8_0, 32, 32, -1, 1 }, { 2, 4 } }, + { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q8_0, 32, 32, 1, 2 }, { 1, 4 } }, + { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q8_0, 32, 32, 2, 2 }, { 4, 4 } }, + { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q8_0, 32, 32, 3, 2 }, { 4, 4 } }, + { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q8_0, 32, 32, 3, 3 }, { 4, 4 } }, + { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q8_0, 64, 64, -1, 0 }, { 1, 4 } }, + { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q8_0, 64, 64, -1, 1 }, { 2, 4 } }, + { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q8_0, 64, 64, 1, 2 }, { 1, 4 } }, + { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q8_0, 64, 64, 1, 3 }, { 4, 4 } }, + { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q8_0, 64, 64, 2, 2 }, { 4, 4 } }, + { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q8_0, 64, 64, 2, 3 }, { 4, 4 } }, + { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q8_0, 64, 64, 3, 2 }, { 4, 4 } }, + { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q8_0, 64, 64, 3, 3 }, { 4, 4 } }, + { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q8_0, 96, 96, -1, 1 }, { 2, 4 } }, + { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q8_0, 96, 96, 1, 2 }, { 1, 4 } }, + { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q8_0, 96, 96, 1, 3 }, { 4, 4 } }, + { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q8_0, 96, 96, 2, 2 }, { 1, 4 } }, + { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q8_0, 96, 96, 2, 3 }, { 4, 4 } }, + { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q8_0, 96, 96, 2, 4 }, { 1, 4 } }, + { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q8_0, 128, 128, -1, 0 }, { 1, 4 } }, + { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q8_0, 128, 128, -1, 1 }, { 1, 4 } }, + { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q8_0, 128, 128, 2, 3 }, { 4, 4 } }, + { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q8_0, 128, 128, 2, 4 }, { 2, 4 } }, + { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q8_0, 128, 128, 3, 3 }, { 2, 4 } }, + { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q8_0, 192, 192, -1, 0 }, { 1, 4 } }, + { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q8_0, 192, 192, -1, 1 }, { 1, 4 } }, + { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q8_0, 192, 128, 1, 0 }, { 1, 4 } }, + { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q8_0, 192, 128, 3, 0 }, { 1, 4 } }, + { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q8_0, 192, 128, -1, 1 }, { 1, 4 } }, + { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q8_0, 192, 128, 3, 3 }, { 1, 2 } }, + { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q8_0, 192, 128, 3, 4 }, { 1, 2 } }, + { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q8_0, 256, 256, -1, 0 }, { 1, 2 } }, + { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q8_0, 256, 256, -1, 1 }, { 2, 4 } }, + { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q8_0, 256, 256, 2, 2 }, { 4, 2 } }, + { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q8_0, 256, 256, 2, 3 }, { 4, 2 } }, + { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q8_0, 320, 256, 1, 1 }, { 2, 4 } }, + { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q8_0, 320, 256, 1, 3 }, { 4, 2 } }, + { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q8_0, 320, 256, 2, 1 }, { 2, 4 } }, + { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q8_0, 320, 256, 2, 2 }, { 4, 2 } }, + { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q8_0, 320, 256, 2, 3 }, { 4, 2 } }, + { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q8_0, 320, 256, 2, 4 }, { 2, 4 } }, + { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q8_0, 320, 256, 3, 1 }, { 2, 4 } }, + { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q8_0, 320, 256, 3, 3 }, { 4, 2 } }, + { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q8_0, 512, 512, -1, 0 }, { 1, 4 } }, + { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q8_0, 512, 512, -1, 1 }, { 1, 4 } }, + { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q8_0, 512, 512, 3, 1 }, { 1, 1 } }, + { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q8_0, 576, 512, -1, 0 }, { 1, 4 } }, + { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q8_0, 576, 512, -1, 1 }, { 1, 4 } }, + { { GGML_METAL_DEVICE_M2_ULTRA, GGML_TYPE_F16, 64, 64, -1, 0 }, { 1, 4 } }, { { GGML_METAL_DEVICE_M2_ULTRA, GGML_TYPE_F16, 64, 64, -1, 1 }, { 1, 4 } }, { { GGML_METAL_DEVICE_M2_ULTRA, GGML_TYPE_F16, 128, 128, 2, 1 }, { 1, 4 } },