wmantly/llama.cpp-xrip
1
0 Fork 0
Code Issues Pull Requests Actions 24 Packages Projects Releases Wiki Activity

mimo : support dflash (convert + feature extraction) (#29650)

Browse Source 
* convert : update to support dflash

* cont : fix

Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@huggingface.co>

---------

Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@huggingface.co>
This commit is contained in:
Georgi GerganovandSigbjørn Skjæret authored and GitHub committed 2026-09-30 17:04:35 +03:00
1 parent ca2e2037b6
commit 22bdcc4cdd
4 files changed
+79 -2
Download Patch File Download Diff File Expand all files Collapse all files

No files matched your search

src/models/dflash.cpp
+6
View File
@@ -8,6 +8,7 @@ void llama_model_dflash::load_arch_hparams(llama_model_loader & ml) {
ml.get_key(LLM_KV_EMBEDDING_SCALE, hparams.f_embedding_scale, false);
ml.get_key(LLM_KV_ATTENTION_SCALE, hparams.f_attention_scale, false);
ml.get_key(LLM_KV_ATTENTION_VALUE_SCALE, hparams.f_attn_value_scale, false);
hparams.llm_ffn_op = LLM_FFN_SILU;
std::string hidden_act;
@@ -738,6 +739,11 @@ llama_model_dflash::graph<false>::graph(const llama_model & model, const llm_gra
? build_attn(inp_attn_iswa, layer.wo, NULL, layer.wo_s, Qcur, Kcur, Vcur, nullptr, layer.attn_sinks, nullptr, kq_scale, il)
: build_attn(inp_attn, layer.wo, NULL, layer.wo_s, Qcur, Kcur, Vcur, nullptr, layer.attn_sinks, nullptr, kq_scale, il);
if (hparams.f_attn_value_scale != 0.0f) {
cur = ggml_scale(ctx0, cur, hparams.f_attn_value_scale);
cb(cur, "attn_out_scaled", il);
}
if (attn_dynamic) {
cur = build_dflash2_conv(*this, cur, attn_dynamic, layer.dflash_attn_conv_base, 1);
cb(cur, "attn_conv_out", il);