From 229109f329c498078f84da39b2c1ebb807e60646 Mon Sep 17 00:00:00 2001 From: Ed Addario Date: Sat, 29 Nov 2025 10:31:39 +0000 Subject: [PATCH] Increase importance boost for final pass --- src/llama-quant.cpp | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/src/llama-quant.cpp b/src/llama-quant.cpp index 2cb58d46bd..44f84ec949 100644 --- a/src/llama-quant.cpp +++ b/src/llama-quant.cpp @@ -1714,7 +1714,7 @@ static std::unordered_map 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)) {