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

Refactor pareto pruning and convexification

This commit is contained in:
Ed Addario 2025-09-21 13:42:54 +01:00
parent 6b8cedf3bc
commit c466c53808
No known key found for this signature in database
GPG Key ID: E7875815A3230993
1 changed files with 43 additions and 50 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

93
src/llama-quant.cpp
View File

@ -1146,8 +1146,7 @@ static std::unordered_map<std::string, ggml_type> target_bpw_type(
for (size_t i = 0; i < base_sz; ++i) { for (size_t i = 0; i < base_sz; ++i) {
ggml_type ts_type = base_arr[i]; ggml_type ts_type = base_arr[i];
if (is_iq(ts_type) && !has_valid_imatrix) { if (is_iq(ts_type) && !has_valid_imatrix) {
LLAMA_LOG_WARN("%s: skipping %s quantization for %s, no or mismatched imatrix provided\n", LLAMA_LOG_WARN("%s: skipping %s quantization for %s, no or mismatched imatrix provided\n", __func__, ggml_type_name(ts_type), name.c_str());
__func__, ggml_type_name(ts_type), name.c_str());
continue; continue;
} }
@ -1214,60 +1213,54 @@ static std::unordered_map<std::string, ggml_type> target_bpw_type(
info.candidate.push_back(candidate_types{ tensor->type, bpw, ggml_nbytes(tensor), 0.0 }); info.candidate.push_back(candidate_types{ tensor->type, bpw, ggml_nbytes(tensor), 0.0 });
} }
// Keep only the pareto‑optimal candidates: if A has >= bytes and >= error than B, drop A. // Keep only the pareto‑optimal candidates and enforce convexity in (bytes, error) curve
{ {
std::vector<candidate_types> pruned; auto & candidates = info.candidate;
pruned.reserve(info.candidate.size()); if (!candidates.empty()) {
std::sort(candidates.begin(), candidates.end(), [](const candidate_types & a, const candidate_types & b) {
if (a.bytes != b.bytes) { return a.bytes < b.bytes; }
// Sort by bytes ascending, error ascending return a.error < b.error;
std::sort(info.candidate.begin(), info.candidate.end(), [](const candidate_types & a, const candidate_types & b) { });
if (a.bytes != b.bytes) { return a.bytes < b.bytes; }
return a.error < b.error;
});
double best_err = infinity; std::vector<candidate_types> pareto;
size_t last_bytes = std::numeric_limits<size_t>::max(); pareto.reserve(candidates.size());
for (const auto & c : info.candidate) { double best_err = infinity;
// Only keep the best error seen so far at strictly larger byte sizes size_t last_bytes = std::numeric_limits<size_t>::max();
if (c.bytes != last_bytes) { for (const auto & c : candidates) {
// first time we see this byte size if (c.bytes != last_bytes) {
last_bytes = c.bytes; last_bytes = c.bytes;
if (c.error < best_err) { if (c.error < best_err) {
pruned.push_back(c); best_err = c.error;
best_err = c.error; pareto.push_back(c);
}
} }
} else {
// same bytes: we already sorted by error; skip
}
}
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); candidates.swap(pareto);
if (candidates.size() >= 3) {
std::vector<candidate_types> hull;
hull.reserve(candidates.size());
auto slope = [](const candidate_types & a, const candidate_types & b) {
const double dx = b.bytes - a.bytes;
return dx <= 0.0 ? infinity : (b.error - a.error) / dx;
};
for (const auto & p : candidates) {
while (hull.size() >= 2) {
double s1 = slope(hull[hull.size() - 2], hull[hull.size() - 1]);
double s2 = slope(hull[hull.size() - 1], p);
if (s2 + epsilon < s1) { hull.pop_back(); }
else { break; }
}
hull.push_back(p);
}
candidates.swap(hull);
}
} }
} }