Skip to content

fix(quantizer): use rounded int8 values for SQ8 metadata to fix recall drop #329

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
richyreachy merged 2 commits into from
Apr 16, 2026
Merged

fix(quantizer): use rounded int8 values for SQ8 metadata to fix recall drop #329

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
a74ab62
fix(quantizer): use rounded int8 values for SQ8 metadata to fix recal…
JoeJRW Apr 15, 2026
b7b49b8
Merge branch 'main' into fix/sq8-quantizer-rounding-consistency
richyreachy Apr 15, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
5 changes: 2 additions & 3 deletions src/core/quantizer/record_quantizer.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -44,11 +44,10 @@ class RecordQuantizer {
scale = 254 / std::max(max - min, epsilon);
bias = -min * scale - 127;
for (size_t i = 0; i < dim; ++i) {
float v = vec[i] * scale + bias;
float v = std::round(vec[i] * scale + bias);
squared_sum += v * v;
sum += v;
(reinterpret_cast<int8_t *>(out))[i] =
static_cast<int8_t>(std::round(v));
(reinterpret_cast<int8_t *>(out))[i] = static_cast<int8_t>(v);
int8_sum += (reinterpret_cast<int8_t *>(out))[i];
}
extras = reinterpret_cast<float *>(static_cast<int8_t *>(out) + dim);
Expand Down
4 changes: 2 additions & 2 deletions tests/core/algorithm/hnsw/hnsw_streamer_test.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2772,7 +2772,7 @@ TEST_F(HnswStreamerTest, TestFetchVectorCosineInt8Converter) {
for (size_t i = 0; i < cnt; i++) {
float add_on = i * 10;
for (size_t j = 0; j < dim; ++j) {
if (j < dim / 4)
if (j < 3 * dim / 4)
vec[j] = fixed_value;
else
vec[j] = fixed_value + add_on;
Expand Down Expand Up @@ -2812,7 +2812,7 @@ TEST_F(HnswStreamerTest, TestFetchVectorCosineInt8Converter) {
for (size_t i = 0; i < query_cnt; i++) {
float add_on = i * 10;
for (size_t j = 0; j < dim; ++j) {
if (j < dim / 4)
if (j < 3 * dim / 4)
vec[j] = fixed_value;
else
vec[j] = fixed_value + add_on;
Expand Down
Loading