sleepy/llama.cpp
1
0
Fork 0
Code Issues 16 Pull Requests Actions 35 Packages Projects Releases Wiki Activity

llama: Add option to merge gate and exp weights (#19139)

Browse Source 
* llama: Add option to merge gate and exp weights

* Update convert_hf_to_gguf.py

Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>

* Update convert_hf_to_gguf.py

Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>

* update constants.py

* add gate_up for the all MoE models

* convert: simplify merge tensor condition

* update constants.py

* reduce number of models, add create_tensor_gate_up helper

---------

Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>
This commit is contained in:
Aman Gupta
2026-02-26 21:01:08 +08:00
committed by GitHub
parent ffaafde16f
commit b68d75165a
12 changed files with 135 additions and 44 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/models/deepseek2.cpp
+3 -1
View File
@@ -218,7 +218,9 @@ llm_build_deepseek2::llm_build_deepseek2(const llama_model & model, const llm_gr
LLM_FFN_SILU, hparams.expert_weights_norm,
hparams.expert_weights_scale, hparams.expert_weights_scale,
(llama_expert_gating_func_type) hparams.expert_gating_func,
il);
il,
nullptr,
model.layers[il].ffn_gate_up_exps);
cb(moe_out, "ffn_moe_out", il);
// FFN shared expert
src/models/qwen35moe.cpp
+2 -1
View File
@@ -380,7 +380,8 @@ ggml_tensor * llm_build_qwen35moe ::build_layer_ffn(ggml_tensor * cur, const int
model.layers[il].ffn_gate_exps, model.layers[il].ffn_down_exps,
nullptr,
n_expert, n_expert_used, LLM_FFN_SILU,
true, false, 0.0, LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX, il);
true, false, 0.0, LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX, il,
nullptr, model.layers[il].ffn_gate_up_exps);
cb(moe_out, "ffn_moe_out", il);
// Add shared experts if present - following Qwen3Next reference implementation
src/models/qwen3next.cpp
+2 -1
View File
@@ -479,7 +479,8 @@ ggml_tensor * llm_build_qwen3next::build_layer_ffn(ggml_tensor * cur, const int
model.layers[il].ffn_gate_exps, model.layers[il].ffn_down_exps,
nullptr,
n_expert, n_expert_used, LLM_FFN_SILU,
true, false, 0.0, LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX, il);
true, false, 0.0, LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX, il,
nullptr, model.layers[il].ffn_gate_up_exps);
cb(moe_out, "ffn_moe_out", il);
// Add shared experts if present - following Qwen3Next reference implementation