metal : add fa-vec tunings for M1 (#28078)

This commit is contained in:
Niklas Wenzel
2026-08-31 13:58:55 +02:00
committed by GitHub
parent 41ef91f7c8
commit 5d4a3be26d
+190
View File
@@ -68,6 +68,196 @@ fa_vec_cfg_t fa_vec_baseline_cfg(int dk, int dv) {
// sweep and paste its output. See ggml-metal-tuning.h for the row/lookup semantics.
// ref: https://github.com/ggml-org/llama.cpp/pull/27824
constexpr fa_vec_entry_t fa_vec_tuned_table[] = {
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_F16, 32, 32, 2, 3 }, { 2, 4 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_F16, 64, 64, -1, 0 }, { 1, 4 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_F16, 64, 64, -1, 1 }, { 1, 4 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_F16, 128, 128, 2, 0 }, { 1, 4 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_F16, 128, 128, -1, 1 }, { 1, 4 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_F16, 128, 128, 1, 1 }, { 1, 1 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_F16, 128, 128, 1, 2 }, { 1, 1 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_F16, 128, 128, 1, 3 }, { 1, 1 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_F16, 128, 128, 1, 4 }, { 1, 1 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_F16, 192, 128, 1, 0 }, { 1, 4 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_F16, 192, 128, 1, 1 }, { 1, 4 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_F16, 192, 128, 1, 2 }, { 1, 4 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_F16, 192, 128, 1, 3 }, { 1, 4 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_F16, 192, 128, 1, 4 }, { 1, 4 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_F16, 320, 256, -1, 1 }, { 1, 4 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_F16, 320, 256, 1, 1 }, { 1, 2 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_F16, 320, 256, 1, 2 }, { 1, 2 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_F16, 320, 256, 1, 3 }, { 1, 2 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_F16, 320, 256, 1, 4 }, { 1, 2 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_Q4_0, 32, 32, -1, 1 }, { 4, 4 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_Q4_0, 32, 32, 1, 1 }, { 2, 4 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_Q4_0, 32, 32, 1, 4 }, { 2, 4 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_Q4_0, 32, 32, 2, 1 }, { 2, 4 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_Q4_0, 32, 32, 2, 4 }, { 2, 4 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_Q4_0, 32, 32, 3, 1 }, { 2, 4 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_Q4_0, 32, 32, 3, 4 }, { 2, 4 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_Q4_0, 64, 64, -1, 0 }, { 1, 4 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_Q4_0, 64, 64, -1, 1 }, { 2, 4 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_Q4_0, 64, 64, 1, 2 }, { 1, 4 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_Q4_0, 64, 64, 2, 2 }, { 1, 4 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_Q4_0, 96, 96, -1, 1 }, { 2, 4 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_Q4_0, 96, 96, 1, 2 }, { 1, 4 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_Q4_0, 96, 96, 1, 4 }, { 1, 4 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_Q4_0, 96, 96, 2, 2 }, { 1, 4 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_Q4_0, 96, 96, 2, 4 }, { 1, 4 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_Q4_0, 96, 96, 3, 2 }, { 1, 4 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_Q4_0, 128, 128, -1, 0 }, { 1, 4 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_Q4_0, 128, 128, -1, 1 }, { 1, 4 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_Q4_0, 192, 192, -1, 0 }, { 1, 4 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_Q4_0, 192, 192, -1, 1 }, { 1, 4 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_Q4_0, 192, 128, -1, 0 }, { 1, 4 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_Q4_0, 192, 128, -1, 1 }, { 1, 4 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_Q4_0, 256, 256, -1, 0 }, { 1, 4 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_Q4_0, 256, 256, -1, 1 }, { 1, 4 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_Q4_0, 320, 256, -1, 0 }, { 1, 4 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_Q4_0, 320, 256, -1, 1 }, { 1, 4 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_Q4_0, 512, 512, -1, 0 }, { 1, 4 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_Q4_0, 512, 512, -1, 1 }, { 1, 4 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_Q4_0, 576, 512, -1, 0 }, { 1, 4 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_Q4_0, 576, 512, -1, 1 }, { 1, 4 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_Q4_1, 32, 32, -1, 1 }, { 2, 4 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_Q4_1, 32, 32, 1, 2 }, { 1, 4 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_Q4_1, 32, 32, 2, 2 }, { 1, 4 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_Q4_1, 32, 32, 2, 3 }, { 4, 4 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_Q4_1, 32, 32, 3, 2 }, { 4, 4 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_Q4_1, 32, 32, 3, 3 }, { 4, 4 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_Q4_1, 64, 64, -1, 0 }, { 1, 4 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_Q4_1, 64, 64, -1, 1 }, { 2, 4 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_Q4_1, 64, 64, 1, 2 }, { 1, 4 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_Q4_1, 96, 96, 1, 3 }, { 2, 4 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_Q4_1, 96, 96, 2, 1 }, { 2, 4 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_Q4_1, 96, 96, 2, 3 }, { 4, 4 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_Q4_1, 96, 96, 3, 1 }, { 2, 4 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_Q4_1, 96, 96, 3, 3 }, { 4, 4 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_Q4_1, 128, 128, -1, 0 }, { 1, 4 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_Q4_1, 128, 128, -1, 1 }, { 2, 4 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_Q4_1, 128, 128, 1, 2 }, { 1, 4 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_Q4_1, 128, 128, 2, 2 }, { 1, 4 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_Q4_1, 192, 192, -1, 0 }, { 1, 4 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_Q4_1, 192, 192, -1, 1 }, { 1, 4 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_Q4_1, 192, 128, -1, 0 }, { 1, 4 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_Q4_1, 192, 128, -1, 1 }, { 1, 4 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_Q4_1, 256, 256, -1, 0 }, { 1, 4 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_Q4_1, 256, 256, -1, 1 }, { 1, 4 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_Q4_1, 320, 256, -1, 0 }, { 1, 4 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_Q4_1, 320, 256, -1, 1 }, { 1, 4 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_Q4_1, 512, 512, -1, 0 }, { 1, 4 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_Q4_1, 512, 512, -1, 1 }, { 1, 4 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_Q4_1, 576, 512, -1, 0 }, { 1, 4 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_Q4_1, 576, 512, -1, 1 }, { 1, 4 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_Q5_0, 32, 32, -1, 1 }, { 4, 4 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_Q5_0, 32, 32, 1, 1 }, { 2, 4 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_Q5_0, 32, 32, 2, 1 }, { 2, 4 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_Q5_0, 32, 32, 3, 1 }, { 2, 4 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_Q5_0, 64, 64, -1, 0 }, { 1, 4 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_Q5_0, 64, 64, -1, 1 }, { 4, 4 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_Q5_0, 64, 64, 1, 1 }, { 2, 4 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_Q5_0, 64, 64, 2, 1 }, { 2, 4 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_Q5_0, 64, 64, 3, 1 }, { 2, 4 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_Q5_0, 96, 96, -1, 1 }, { 4, 4 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_Q5_0, 96, 96, 1, 1 }, { 2, 4 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_Q5_0, 96, 96, 2, 1 }, { 2, 4 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_Q5_0, 96, 96, 2, 4 }, { 2, 4 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_Q5_0, 96, 96, 3, 1 }, { 2, 4 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_Q5_0, 96, 96, 3, 4 }, { 2, 4 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_Q5_0, 128, 128, -1, 0 }, { 1, 4 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_Q5_0, 128, 128, -1, 1 }, { 2, 4 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_Q5_0, 128, 128, 1, 2 }, { 1, 4 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_Q5_0, 192, 192, -1, 0 }, { 1, 4 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_Q5_0, 192, 192, -1, 1 }, { 2, 4 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_Q5_0, 192, 192, 1, 2 }, { 1, 4 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_Q5_0, 192, 192, 2, 2 }, { 1, 4 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_Q5_0, 192, 128, -1, 0 }, { 1, 4 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_Q5_0, 192, 128, -1, 1 }, { 1, 4 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_Q5_0, 192, 128, 1, 3 }, { 4, 2 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_Q5_0, 192, 128, 2, 3 }, { 4, 2 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_Q5_0, 192, 128, 2, 4 }, { 2, 4 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_Q5_0, 192, 128, 3, 1 }, { 2, 4 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_Q5_0, 192, 128, 3, 3 }, { 4, 2 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_Q5_0, 192, 128, 3, 4 }, { 2, 4 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_Q5_0, 256, 256, -1, 0 }, { 1, 4 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_Q5_0, 256, 256, -1, 1 }, { 1, 4 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_Q5_0, 320, 256, -1, 0 }, { 1, 4 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_Q5_0, 320, 256, -1, 1 }, { 1, 4 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_Q5_0, 512, 512, -1, 0 }, { 1, 4 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_Q5_0, 512, 512, -1, 1 }, { 1, 4 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_Q5_0, 576, 512, -1, 0 }, { 1, 4 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_Q5_0, 576, 512, -1, 1 }, { 1, 4 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_Q5_1, 32, 32, -1, 1 }, { 4, 4 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_Q5_1, 32, 32, 1, 1 }, { 2, 4 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_Q5_1, 32, 32, 2, 1 }, { 2, 4 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_Q5_1, 32, 32, 3, 1 }, { 2, 4 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_Q5_1, 64, 64, -1, 0 }, { 1, 4 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_Q5_1, 64, 64, -1, 1 }, { 4, 4 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_Q5_1, 64, 64, 1, 1 }, { 2, 4 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_Q5_1, 64, 64, 2, 1 }, { 2, 4 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_Q5_1, 64, 64, 3, 1 }, { 2, 4 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_Q5_1, 96, 96, -1, 1 }, { 4, 4 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_Q5_1, 96, 96, 1, 1 }, { 2, 4 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_Q5_1, 96, 96, 2, 1 }, { 2, 4 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_Q5_1, 96, 96, 3, 1 }, { 2, 4 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_Q5_1, 128, 128, -1, 0 }, { 1, 4 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_Q5_1, 128, 128, -1, 1 }, { 2, 4 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_Q5_1, 128, 128, 1, 2 }, { 1, 4 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_Q5_1, 192, 192, -1, 0 }, { 1, 4 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_Q5_1, 192, 192, -1, 1 }, { 1, 4 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_Q5_1, 192, 192, 1, 3 }, { 2, 4 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_Q5_1, 192, 192, 3, 1 }, { 2, 4 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_Q5_1, 192, 128, -1, 0 }, { 1, 4 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_Q5_1, 192, 128, -1, 1 }, { 1, 4 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_Q5_1, 192, 128, 1, 3 }, { 4, 2 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_Q5_1, 192, 128, 2, 3 }, { 4, 2 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_Q5_1, 192, 128, 3, 3 }, { 4, 2 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_Q5_1, 256, 256, -1, 0 }, { 1, 4 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_Q5_1, 256, 256, -1, 1 }, { 1, 4 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_Q5_1, 320, 256, -1, 0 }, { 1, 4 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_Q5_1, 320, 256, -1, 1 }, { 1, 4 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_Q5_1, 512, 512, -1, 0 }, { 1, 4 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_Q5_1, 512, 512, -1, 1 }, { 1, 4 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_Q5_1, 512, 512, 3, 3 }, { 1, 1 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_Q5_1, 576, 512, -1, 0 }, { 1, 4 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_Q5_1, 576, 512, -1, 1 }, { 1, 4 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_Q8_0, 32, 32, -1, 1 }, { 2, 4 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_Q8_0, 32, 32, 1, 2 }, { 1, 4 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_Q8_0, 32, 32, 2, 2 }, { 1, 4 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_Q8_0, 32, 32, 2, 3 }, { 4, 4 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_Q8_0, 32, 32, 2, 4 }, { 1, 4 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_Q8_0, 32, 32, 3, 2 }, { 1, 4 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_Q8_0, 32, 32, 3, 3 }, { 4, 4 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_Q8_0, 64, 64, -1, 0 }, { 1, 4 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_Q8_0, 64, 64, -1, 1 }, { 2, 4 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_Q8_0, 64, 64, 1, 2 }, { 1, 4 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_Q8_0, 64, 64, 2, 2 }, { 1, 4 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_Q8_0, 64, 64, 3, 2 }, { 1, 4 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_Q8_0, 96, 96, 1, 3 }, { 4, 4 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_Q8_0, 96, 96, 2, 1 }, { 2, 4 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_Q8_0, 96, 96, 2, 3 }, { 4, 4 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_Q8_0, 96, 96, 3, 1 }, { 2, 4 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_Q8_0, 96, 96, 3, 3 }, { 2, 4 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_Q8_0, 128, 128, -1, 0 }, { 1, 4 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_Q8_0, 128, 128, -1, 1 }, { 1, 4 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_Q8_0, 192, 192, -1, 0 }, { 1, 4 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_Q8_0, 192, 192, -1, 1 }, { 1, 4 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_Q8_0, 192, 128, -1, 0 }, { 1, 4 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_Q8_0, 192, 128, -1, 1 }, { 1, 4 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_Q8_0, 256, 256, -1, 0 }, { 1, 2 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_Q8_0, 256, 256, -1, 1 }, { 2, 4 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_Q8_0, 320, 256, 1, 1 }, { 2, 4 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_Q8_0, 320, 256, 1, 3 }, { 4, 2 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_Q8_0, 320, 256, 2, 1 }, { 2, 4 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_Q8_0, 320, 256, 2, 3 }, { 2, 4 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_Q8_0, 320, 256, 3, 1 }, { 2, 4 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_Q8_0, 320, 256, 3, 3 }, { 2, 4 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_Q8_0, 512, 512, -1, 0 }, { 1, 4 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_Q8_0, 512, 512, -1, 1 }, { 1, 4 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_Q8_0, 512, 512, 2, 2 }, { 1, 1 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_Q8_0, 512, 512, 3, 2 }, { 1, 1 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_Q8_0, 576, 512, -1, 0 }, { 1, 4 } },
{ { GGML_METAL_DEVICE_M1, GGML_TYPE_Q8_0, 576, 512, -1, 1 }, { 1, 4 } },
{ { GGML_METAL_DEVICE_M1_PRO, GGML_TYPE_F16, 32, 32, 3, 3 }, { 2, 4 } },
{ { GGML_METAL_DEVICE_M1_PRO, GGML_TYPE_F16, 64, 64, -1, 0 }, { 1, 4 } },
{ { GGML_METAL_DEVICE_M1_PRO, GGML_TYPE_F16, 64, 64, -1, 1 }, { 1, 4 } },