Commit 4e329bd
Add HIGGS per-layer quantization support to eval_ppl.py
- Add --higgs-assignment argument for dynamic bitwidth assignment
- Add load_higgs_assignment() to load JSON assignments
- Add compute_absmax_codebook() and compute_l2_codebook() helpers
- Apply per-layer quantization hooks based on assignment
Usage: python eval_ppl.py --model ... --higgs-assignment path/to/assignment.json
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>1 parent 6d81104 commit 4e329bd
1 file changed
+1504
-0
lines changed
0 commit comments