char-fork/llama.cpp

git clone https://git.t4t.associates/char-fork/llama.cpp

YiChen Lvmetal : per-device tuned (Q, NE) for flash-attn vec (#26570)f280b2698

char/aliceai-t5-moe
brancheschar/aliceai-t5-moeall refs →
703 B18 linesraw
1#pragma once
2
3#include "ggml-backend.h"
4
5struct tuner_opts {
6    const char * dtype_filter   = nullptr;  // comma-separated, e.g. "f16,q4_0"; null = all
7    const char * dk_filter      = nullptr;  // comma-separated dk values, e.g. "128,192"; null = all
8    int          reps           = 7;
9    unsigned     seed           = 1234;
10    bool         cooldown       = true;
11    double       cool_drift     = 0.10;
12    double       cool_eps       = 0.03;
13    int          cool_max_wait  = 120;
14    int          cool_max_retry = 2;
15};
16
17// Returns false only when the required Metal proc bridges are unavailable.
18bool tuner_fa_vec_run(ggml_backend_t backend, ggml_backend_dev_t dev, const tuner_opts & opts);