← Back to Quantization

mixed int4/int6/int8

Quantization
Used in
3 PRs
Best BPB
1.0587
Avg BPB
1.1732

Submissions

PR #1731by Victory963
1.0785
PR #1821by anjing00monyet-arch
1.3825
PR #1919by dev-pratap-singh
1.0587

Hyperparameters Across PRs

pr_numberbitsscope
1731—embeddings, attention, MLP, residuals
1821—embeddings
1919—embeddings and block weights