matt/ollama37
1
0
Fork 0
mirror of https://github.com/dogkeeper886/ollama37.git synced 2025-12-16 10:47:01 +00:00
Code Issues Packages Projects Releases Wiki Activity

gemma3 quantization (#9776)

Browse Source
This commit is contained in:
Patrick Devine
2025-03-14 17:41:07 -07:00
committed by GitHub
parent 2d2247e59e
commit ef378ad673
5 changed files with 149 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files

9
llama/llama.cpp/src/llama-quant.cpp vendored
View File

@@ -737,6 +737,15 @@ static void llama_model_quantize_impl(const std::string & fname_inp, const std::
// This used to be a regex, but <regex> has an extreme cost to compile times.
bool quantize = name.rfind("weight") == name.size() - 6; // ends with 'weight'?
// don't quantize vision stuff
quantize &= name.find("v.blk.") == std::string::npos;
quantize &= name.find("mm.mm_input_projection.weight") == std::string::npos;
quantize &= name.find("mm.mm_soft_emb_norm.weight") == std::string::npos;
quantize &= name.find("v.patch_embedding.weight") == std::string::npos;
quantize &= name.find("v.position_embedding.weight") == std::string::npos;
quantize &= name.find("v.post_layernorm.weight") == std::string::npos;
// quantize only 2D and 3D tensors (experts)
quantize &= (ggml_n_dims(tensor) >= 2);