← Back to Quantization

mixed int6/int7

Quantization
Used in
14 PRs
Best BPB
1.0135
Avg BPB
1.0585

Hyperparameters Across PRs

pr_numberbitsscope
1707—all
1934—embeddings
1953—weights + embeddings
1958—model weights
19926attention/MLP matrices
19927token embeddings
1995—matrix weights + embeddings
20676weights and embeddings
21206weights and embeddings
2123—embeddings and block weights
2124—embeddings and block weights
2128—attn/MLP int6, embeddings int7
2130—model weights
2133—weights + embeddings