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 -2
View File
@@ -203,6 +203,8 @@ void llama_model_deepseek41::load_arch_tensors(llama_model_loader & ml) {
}
layer.ffn_gate_inp = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP, "weight", i), {n_embd, n_expert}, 0);
// the reference computes the router in F32: https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash-1017/blob/1bae0e1088993b9b82de08f59de5be5a34e42170/inference/model.py#L813
prec_policy.set_acc(layer.ffn_gate_inp, GGML_PREC_F32);
layer.ffn_exp_probs_b = create_tensor(tn(LLM_TENSOR_FFN_EXP_PROBS_B, "bias", i), {n_expert}, 0);
layer.ffn_exp_probs_b_vl = create_tensor(tn(LLM_TENSOR_FFN_EXP_PROBS_B_VL, "bias", i), {n_expert}, TENSOR_NOT_REQUIRED);
layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
@@ -874,9 +876,7 @@ llama_model_deepseek41::graph::graph(const llm_graph_params & params) :
cur = build_norm(cur, layer.ffn_norm, nullptr, LLM_NORM_RMS, il);
cb(cur, "ffn_norm", il);
// the reference computes the router in F32: https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash-1017/blob/1bae0e1088993b9b82de08f59de5be5a34e42170/inference/model.py#L813
ggml_tensor * logits = build_lora_mm(layer.ffn_gate_inp, cur);
ggml_prec_set_acc(logits, GGML_PREC_F32);
cb(logits, "ffn_moe_logits", il);
// image (embedding) tokens use the vision router bias, as DeepSeek-V4