← Back to Quantization

GPTQ with early QAT

Quantization
Used in
1 PRs
Best BPB
1.1215
Avg BPB
1.1215

Submissions

PR #578by newjordan
1.1215

Hyperparameters Across PRs

pr_numberbitsscope
5786all weights (per-row int6 quantization with Hessian-aware error compensation)