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

graph : remove redundant scale_w parameter (#20235)

Browse Source
This commit is contained in:
Sigbjørn Skjæret
2026-03-08 18:58:28 +01:00
committed by GitHub
parent 451ef08432
commit 35bee031e1
41 changed files with 85 additions and 86 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/models/step35-iswa.cpp
+2 -5
View File
@@ -119,9 +119,6 @@ llm_build_step35_iswa::llm_build_step35_iswa(const llama_model & model, const ll
cb(cur, "ffn_out", il);
} else {
// MoE routed experts
const bool norm_w = hparams.expert_weights_norm;
const float w_scale = hparams.expert_weights_scale;
const bool scale_w = w_scale != 0.0f;
ggml_tensor * moe_out = build_moe_ffn(cur,
model.layers[il].ffn_gate_inp,
model.layers[il].ffn_up_exps,
@@ -129,8 +126,8 @@ llm_build_step35_iswa::llm_build_step35_iswa(const llama_model & model, const ll
model.layers[il].ffn_down_exps,
model.layers[il].ffn_exp_probs_b,
n_expert, n_expert_used,
LLM_FFN_SILU,
norm_w, scale_w, w_scale,
LLM_FFN_SILU, hparams.expert_weights_norm,
hparams.expert_weights_scale,
(llama_expert_gating_func_type) hparams.expert_gating_func,
il);
cb(moe_out, "ffn_moe_out", il);