graph : remove redundant scale_w parameter (#20235)
This commit is contained in:
@@ -1,5 +1,4 @@
|
||||
#include "models.h"
|
||||
#include "ggml.h"
|
||||
|
||||
#include "llama-memory-recurrent.h"
|
||||
|
||||
@@ -341,7 +340,7 @@ llm_build_kimi_linear::llm_build_kimi_linear(const llama_model & model, const ll
|
||||
hparams.n_expert,
|
||||
hparams.n_expert_used,
|
||||
LLM_FFN_SILU, true,
|
||||
hparams.expert_weights_scale, hparams.expert_weights_scale,
|
||||
hparams.expert_weights_scale,
|
||||
(llama_expert_gating_func_type) hparams.expert_gating_func,
|
||||
il);
|
||||
cb(moe_out, "ffn_moe_out", il);
|
||||
|
||||
Reference in New Issue
Block a user