vbatts/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

llama : do not quantize expert gating tensors

Browse source
This commit is contained in:
Georgi Gerganov 2023-12-10 13:00:13 +02:00
parent 6cfb31f9ea
commit d1259b7b35
No known key found for this signature in database
GPG key ID: 449E073F9DC10735
1 changed files with 3 additions and 0 deletions
Download patch file Download diff file Expand all files Collapse all files

3
llama.cpp
View file

@ -8443,6 +8443,9 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
quantize &= params->quantize_output_tensor || name != "output.weight"; quantize &= params->quantize_output_tensor || name != "output.weight";
quantize &= !params->only_copy; quantize &= !params->only_copy;
// do not quantize expert gating tensors
quantize &= name.find("ffn_gate_inp.weight") == std::string::npos;
enum ggml_type new_type; enum ggml_type new_type;
void * new_data; void * new_data;
size_t new_size; size_t new_size;