← Back to Quantization

mixed int6/int8 STE QAT

Quantization
Used in
1 PRs
Best BPB
1.1556
Avg BPB
1.1556

Submissions

PR #65by aquariouseworkmanRECORD
1.1556

Hyperparameters Across PRs

pr_numberbitsscope
656all 2D block weights int6; token embeddings int8/fp16 passthrough