← Back to Quantization

STE QAT int6

Quantization
Used in
5 PRs
Best BPB
1.1502
Avg BPB
1.1935

Submissions

PR #66by arjun-krishna1
1.1632
PR #81by polarizedfortnite-cpu
1.1670
PR #86by aruniyerRECORD
1.1502
PR #122by mtybadger
1.1603
PR #488by pkim02
1.3267

Hyperparameters Across PRs

pr_numberbitsscope
666CastedLinear weights / MLP and attention weights
816all weights except tied embeddings
866all block weights
1226row-wise weights; embeddings kept in fp16
4886all weights