jdelony/ik_llama.cpp
1
0
Fork 0
mirror of https://github.com/ikawrakow/ik_llama.cpp.git synced 2026-06-28 04:30:15 -05:00
Code Issues Packages Projects Releases Wiki Activity

Revert DFlash SWA optimization (#2039)

Browse Source
This commit is contained in:
Kawrakow 2026-06-26 11:00:09 +02:00 committed by GitHub
parent 1255b1e479
commit f96eaddba8
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
1 changed files with 5 additions and 3 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

8
src/graphs/build_dflash.cpp
View File

@ -315,9 +315,11 @@ ggml_cgraph * llm_build_context::build_dflash() {
hparams.attn_soft_cap ? hparams.f_attn_logit_softcapping : 0.0f);
cb(cur, "flash_attn", il);
ggml_build_forward_expand(gf, cur);
if (use_swa) {
cur->op_params[4] = hparams.n_swa;
}
// Somethiong goes wrong with thisi optimization.
// I guess, the cross context does not mingle well with it.
//if (use_swa) {
// cur->op_params[4] = hparams.n_swa;
//}
cur = ggml_reshape_2d(ctx0, cur, model.layers[il].wo->ne[0], n_tokens);
cb(cur, "flash_attn_reshaped", il);