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

"Convexify" candidate list

This commit is contained in:
Ed Addario 2025-09-13 09:41:52 +01:00
parent 12e816b511
commit 2b516068e2
No known key found for this signature in database
GPG Key ID: E7875815A3230993
1 changed files with 26 additions and 0 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

26
src/llama-quant.cpp
View File

@ -1257,6 +1257,32 @@ static std::unordered_map<std::string, ggml_type> target_bpw_type(
info.candidate.swap(pruned);
}
// Enforce convexity in (bytes, error) curve
{
const auto & c = info.candidate;
if (c.size() >= 3) {
std::vector<candidate_types> convex;
convex.reserve(c.size());
auto slope = [](const candidate_types & a, const candidate_types & b) -> double {
const double dx = (double)b.bytes - (double)a.bytes;
if (dx <= 0.0) { return infinity; }
return ((double)b.error - (double)a.error) / dx;
};
for (const auto & p : c) {
while (convex.size() >= 2) {
double s1 = slope(convex[convex.size() - 2], convex[convex.size() - 1]);
double s2 = slope(convex[convex.size() - 1], p);
if (s2 + epsilon < s1) { convex.pop_back(); }
else { break; }
}
convex.push_back(p);
}
info.candidate.swap(convex);
}
}
// Initialize choice at the smallest bpw candidate
info.choice = 0;
info.min_bpw = info.candidate.front().bpw;