Repository navigation

MOD-19169: Use native FP16 arithmetic for SQ8–FP16 inner product - #1074

Open

dor-forer wants to merge 8 commits into

mainfrom

codex/mod19169-native-fp16-first

Open

MOD-19169: Use native FP16 arithmetic for SQ8–FP16 inner product#1074
dor-forer wants to merge 8 commits into
mainfrom
codex/mod19169-native-fp16-first

Commits on Oct 4, 2026
- MOD-19169 Add native FP16 arithmetic for SQ8 FP16 inner product
  dor-forer
  committed
  ac4d892
- Use four SQ8 FP16 accumulators with FP32 final reduction
  dor-forer
  committed
  ef7311a
Commits on Oct 5, 2026
- Name SQ8 FP16 dispatch thresholds and reduction policy
  dor-forer
  committed
  2596821
- Rename SQ8 FP16 helper to LoadQuantizedValues
  dor-forer
  committed
  48f05a3
- Split native SQ8 FP16 inner product into focused helpers
  dor-forer
  committed
  490a384
- Rename SQ8 byte pointers to quantized_values
  dor-forer
  committed
  7897ea8
- Make SQ8 FP16 tail block counts explicit
  dor-forer
  committed
  742193c
Commits on Oct 21, 2026
- Test minimum FP16 subnormals without distance cancellation
  dor-forer
  committed
  e56df68