graph : remove redundant scale_w parameter (#20235)
This commit is contained in:
@@ -119,8 +119,7 @@ llm_build_hunyuan_moe::llm_build_hunyuan_moe(const llama_model & model, const ll
|
||||
n_expert, n_expert_used,
|
||||
LLM_FFN_SILU,
|
||||
true, // norm_topk_prob
|
||||
false,
|
||||
0.0,
|
||||
hparams.expert_weights_scale,
|
||||
LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX,
|
||||
il);
|
||||
cb(cur_moe, "ffn_moe_out", il);
|
||||
|
||||
Reference in New Issue
Block a user