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

Update comments

This commit is contained in:
Ed Addario 2025-08-22 09:02:56 +01:00
parent 35c1504441
commit bb0d912c1f
No known key found for this signature in database
GPG Key ID: E7875815A3230993
1 changed files with 3 additions and 2 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

5
src/llama-quant.cpp
View File

@ -703,6 +703,7 @@ static std::unordered_map<std::string, ggml_type> target_bpw_type(
auto name_tn = LLM_TN(model.arch); auto name_tn = LLM_TN(model.arch);
auto can_quantize = [&](const ggml_tensor * t) -> bool { auto can_quantize = [&](const ggml_tensor * t) -> bool {
// This list should be kept in sync with llama_tensor_quantize_impl()
const std::string name = ggml_get_name(t); const std::string name = ggml_get_name(t);
bool q = name.rfind("weight") == name.size() - 6; bool q = name.rfind("weight") == name.size() - 6;
q &= ggml_n_dims(t) >= 2; q &= ggml_n_dims(t) >= 2;
@ -902,7 +903,7 @@ static std::unordered_map<std::string, ggml_type> target_bpw_type(
constexpr float bias_lambda = 1.0; constexpr float bias_lambda = 1.0;
//bias_lambda defines the weight of the bias term in the weigthed MSE error function //bias_lambda defines the weight of the bias term in the weigthed MSE error function
// 0.0 means no bias (standard MSE) 1.0 means equal weight for bias and error, // 0.0 means no bias (standard MSE) 1.0 means equal weight for bias and error,
// 2.0 means twice as much weight for bias, etc // 2.0 means twice as much weight for bias, etc. Default is 1.0.
if (activations && bias_lambda != 0.0) { if (activations && bias_lambda != 0.0) {
const double proj = bias_numerator * bias_numerator / (bias_denominator + epsilon); const double proj = bias_numerator * bias_numerator / (bias_denominator + epsilon);
err_numerator += bias_lambda * proj; err_numerator += bias_lambda * proj;
@ -1192,7 +1193,7 @@ static std::unordered_map<std::string, ggml_type> target_bpw_type(
double bpw_now = current_bpw(); double bpw_now = current_bpw();
float target_bpw = params->target_bpw; float target_bpw = params->target_bpw;
// If minimal bpw is already above the target, we're constrained by geometry; return closest (min bpw) // If minimal bpw is already above the target, we're constrained by the tensor's shape; return closest (min bpw)
if (bpw_now >= target_bpw) { if (bpw_now >= target_bpw) {
std::unordered_map<std::string, ggml_type> overrides; std::unordered_map<std::string, ggml_type> overrides;
for (const auto & ti : all) { for (const auto & ti : all) {