model: apply .scale tensors automatically + fix correctness (#32669)

* model: apply .scale tensors automatically

* rm build_moe_ffn "_s" overload

* use llama_prec_policy

* fix ci

* build_lora_mm_id: rm w_s
This commit is contained in:
Xuan-Son Nguyen authored and GitHub committed 2026-11-21 16:30:01 +01:00
1 parent e50c254a22
commit 19bc41dbb4
39 files changed
+164 -290

No files matched your search

+2 -8
View File
@@ -561,10 +561,7 @@ ggml_tensor * llama_model_qwen3next::graph::build_layer_ffn(ggml_tensor * cur, c
LLM_FFN_SILU, true,
hparams.expert_weights_scale,
LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX, il,
nullptr, model.layers[il].ffn_gate_up_exps,
model.layers[il].ffn_up_exps_s,
model.layers[il].ffn_gate_exps_s,
model.layers[il].ffn_down_exps_s);
nullptr, model.layers[il].ffn_gate_up_exps);
cb(moe_out, "ffn_moe_out", il);
// Add shared experts if present - following Qwen3Next reference implementation
@@ -748,10 +745,7 @@ llama_model_qwen3next::graph_mtp::graph_mtp(const llm_graph_params & params)
LLM_FFN_SILU, true,
hparams.expert_weights_scale,
LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX, il,
nullptr, layer.ffn_gate_up_exps,
layer.ffn_up_exps_s,
layer.ffn_gate_exps_s,
layer.ffn_down_exps_s);
nullptr, layer.ffn_gate_up_exps);
cb(moe_out, "mtp_ffn_moe_out", il);
if (layer.ffn_up_shexp != nullptr) {