Repository navigation

Collect imatrix statistics with one fused float32 reduction - #3044

Merged

cjluo-nv merged 5 commits into

mainfrom

chenjiel/imatrix-fused-reduction

Oct 31, 2026

Merged

Collect imatrix statistics with one fused float32 reduction#3044
cjluo-nv merged 5 commits into
mainfrom
chenjiel/imatrix-fused-reduction

Commits on Oct 30, 2026
- Collect imatrix statistics with one fused float32 reduction
  
  cjluo-nv
  and
  claude
  committed
  aad9691
- Keep imatrix's bounded float32 path off the GPU and for float64
  
  cjluo-nv
  and
  claude
  committed
  7251213
- Reduce imatrix GPU inputs in 512-row groups to bound scratch memory
  
  cjluo-nv
  and
  claude
  committed
  bbe3b08
- Bound imatrix's GPU group sums to slices of 512 groups
  
  cjluo-nv
  and
  claude
  committed
  ee4941f
- Reduce imatrix GPU inputs of up to 1,536 rows in one call
  
  cjluo-nv
  and
  claude
  committed
  d5b4249