Repository navigation

Add imatrix calibration for the GGML formats and use it for fused MoE experts - #2735

Open

cjluo-nv wants to merge 9 commits into

mainfrom

chenjiel/iq-imatrix-calibration

Open

Add imatrix calibration for the GGML formats and use it for fused MoE experts#2735
cjluo-nv wants to merge 9 commits into
mainfrom
chenjiel/iq-imatrix-calibration

Commits on Oct 13, 2026
- Add imatrix calibration for the GGML formats and use it for fused MoE experts
  
  cjluo-nv
  and
  claude
  committed
  d12cd57
- Add an IQ2_XXS experts-only recipe calibrated with imatrix
  
  cjluo-nv
  and
  claude
  committed
  8013d53
- Test that layerwise imatrix sees the earlier layers' quantization
  
  cjluo-nv
  and
  claude
  committed
  a2d5a16
- Collect imatrix statistics in a single pass in the IQ2 recipes
  
  cjluo-nv
  and
  claude
  committed
  5f6799a
- Add the imatrix step to the IQ1 recipes
  
  cjluo-nv
  and
  claude
  committed
  8ebb377
- Add experts-only recipes for every IQ format
  
  cjluo-nv
  and
  claude
  committed
  7fdbba6
- Check imatrix targets before calibrating and bound the hook's memory
  
  cjluo-nv
  and
  claude
  committed
  690c5a1
- Check imatrix targets before max calibration runs its forward
  
  cjluo-nv
  and
  claude
  committed
  ad8fe16
- Move the imatrix changelog entry to 0.49.0
  
  cjluo-nv
  and
  claude
  committed
  9e33d08