char-fork/llama.cpp

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

Nick Farrellsycl: make --fit respect --fit-target better (#27629)cc83d7b48

char/aliceai-t5-moe
brancheschar/aliceai-t5-moeall refs →
905 B18 linesraw
1#ifndef GGML_SYCL_FATTN_ONEDNN_HPP
2#define GGML_SYCL_FATTN_ONEDNN_HPP
3
4#include "common.hpp"
5
6// Static-only check: fused-XMX oneDNN Graph SDPA path==flash-attn op
7// (f16 KV, no softcap/ALiBi, single stream, tuned head_dim, prefill-sized q.)
8bool ggml_sycl_flash_attn_ext_onednn_supported(const ggml_tensor * dst, bool use_shape_limit = true);
9
10// True when the oneDNN path binds an F16 KV cache in place instead of staging a dense copy of
11// it. Depends only on the types and strides of K and V, so the answer holds for every call.
12bool ggml_sycl_fattn_onednn_binds_kv(const ggml_tensor * K, const ggml_tensor * V);
13
14// Run flash attention through oneDNN's fused xmx SDPA
15// execute the cached SDPA partition, write the f32 dst. Falls back to the TILE kernel on any failure.
16void ggml_sycl_flash_attn_ext_onednn(ggml_backend_sycl_context & ctx, ggml_tensor * dst);
17
18#endif // GGML_SYCL_FATTN_ONEDNN_HPP