Lumpiasty/llama.cpp
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

spec : add DFlash support (#22105)

Browse Source 
* spec: add DFlash v2 support

* dflash: support sliding window attention per layer_types

* docs: add dflash section

---------

Co-authored-by: Kashif Rasul <kashif.rasul@gmail.com>
This commit is contained in:
Ruixiang Wang
2026-06-28 16:01:34 +03:00
committed by GitHub
co-authored by Kashif Rasul
parent c1a1c8ee94
commit d1b34251bc
14 changed files with 712 additions and 9 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/llama-graph.cpp
+6 -1
View File
@@ -486,7 +486,11 @@ void llm_graph_input_attn_kv::set_input(const llama_ubatch * ubatch) {
mctx->set_input_k_idxs(self_k_idxs, ubatch);
mctx->set_input_v_idxs(self_v_idxs, ubatch);
mctx->set_input_kq_mask(self_kq_mask, ubatch, cparams.causal_attn);
// the mask is left unallocated when the graph only stores K/V without attending
// (e.g. DFlash's KV-injection pass)
if (self_kq_mask && self_kq_mask->buffer) {
mctx->set_input_kq_mask(self_kq_mask, ubatch, cparams.causal_attn);
}
if (self_k_rot) {
mctx->set_input_k_rot(self_k_rot);
@@ -904,6 +908,7 @@ void llm_graph_result::reset() {
t_logits = nullptr;
t_embd = nullptr;
t_embd_pooled = nullptr;
t_h_nextn = nullptr;
t_layer_inp.resize(LLAMA_MAX_LAYERS);
std::fill(t_layer_inp.begin(), t_layer_inp.end(), nullptr);