← Back to Quantization

mixed int8/fp16

Quantization
Used in
1 PRs
Best BPB
1.1884
Avg BPB
1.1884

Submissions

PR #74by takhir-iota
1.1884

Hyperparameters Across PRs

pr_numberbitsscope
748all weights except tok_emb.weight kept in fp16; blocks.5. selectively coarsened