happyz
/
llama.cpp
mirror of https://github.com/ggerganov/llama.cpp.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

Increase importance boost for final pass

This commit is contained in:
Ed Addario 2025-11-29 10:31:39 +00:00
parent 5b557ca958
commit 229109f329
No known key found for this signature in database
GPG Key ID: E7875815A3230993
1 changed files with 1 additions and 1 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
src/llama-quant.cpp
View File

@ -1714,7 +1714,7 @@ static std::unordered_map<std::string, ggml_type> target_bpw_type(
if (err_gain < epsilon) { continue; } // no error improvement
double ratio = err_gain / (double)delta_bytes; // error reduction per byte
if (is_important(tensor_name)) { ratio *= 2.0; } // important tensors get 2x boost
if (is_important(tensor_name)) { ratio *= 5.0; } // important tensors get 5x boost
// For tie-breaking, prioritize the largest absolute error improvement.
if (ratio > best_ratio + epsilon || (std::abs(ratio - best_ratio) <= epsilon && err_gain > best_gain)) {