graph : remove redundant scale_w parameter (#20235)
This commit is contained in:
@@ -124,7 +124,7 @@ ggml_tensor * llm_build_nemotron_h::build_ffn_layer(ggml_tensor * cur, const lla
|
||||
model.layers[il].ffn_exp_probs_b,
|
||||
n_expert, n_expert_used,
|
||||
LLM_FFN_RELU_SQR, hparams.expert_weights_norm,
|
||||
hparams.expert_weights_scale, hparams.expert_weights_scale,
|
||||
hparams.expert_weights_scale,
|
||||
LLAMA_EXPERT_GATING_FUNC_TYPE_SIGMOID,
|
||||
il);
|
||||
cb(moe_out, "ffn_moe_out", il);
|
||||
|
||||
Reference in New Issue
Block a user