Fix glm4moe bug (#15088)
This commit is contained in:
@@ -13800,10 +13800,6 @@ struct llm_build_glm4_moe : public llm_graph_context {
|
||||
LLM_FFN_SILU, LLM_FFN_PAR, il);
|
||||
cb(cur, "ffn_out", il);
|
||||
} else {
|
||||
// MoE layer with shared experts
|
||||
const int64_t n_expert = hparams.n_expert;
|
||||
const int64_t n_expert_used = hparams.n_expert_used;
|
||||
|
||||
// Process routed experts using existing MoE infrastructure
|
||||
ggml_tensor * routed_out = build_moe_ffn(cur,
|
||||
model.layers[il].ffn_gate_inp,
|
||||
|
||||
Reference in New Issue
Block a user