vrr/koboldcpp
2
0
Fork 0
mirror of https://github.com/LostRuins/koboldcpp.git synced 2025-09-13 10:29:43 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions 1

llama : merge build_moe_ffn_from_probs function into build_moe_ffn (#14968)

Browse source
This commit is contained in:
Dongliang Wei 2025-07-31 20:12:20 +08:00 • committed by GitHub
parent a9f77a8be3
commit c1dacaa99b
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
3 changed files with 32 additions and 114 deletions
Download patch file Download diff file Expand all files Collapse all files

16
src/llama-model.cpp
View file

@ -17320,10 +17320,18 @@ struct llm_build_smallthinker : public llm_graph_context{
cur = build_norm(ffn_inp, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il);
cb(cur, "ffn_norm", il);
ggml_tensor * ffn_out = build_moe_ffn_from_probs(cur, probs, model.layers[il].ffn_up_exps,
model.layers[il].ffn_gate_exps, model.layers[il].ffn_down_exps,
nullptr, n_expert, n_expert_used,
static_cast<llama_expert_gating_func_type>(hparams.expert_gating_func), il);
ggml_tensor * ffn_out =
build_moe_ffn(cur,
nullptr,
model.layers[il].ffn_up_exps,
model.layers[il].ffn_gate_exps,
model.layers[il].ffn_down_exps,
nullptr,
n_expert, n_expert_used,
LLM_FFN_RELU, true,
false, 0.0,
static_cast<llama_expert_gating_func_type>(hparams.expert_gating_func),
il, probs);
cb(ffn_out, "ffn_out", il);
cur = ffn_out;