← Back to Quantization

mixed int4/int5

Quantization
Used in
3 PRs
Best BPB
1.1257
Avg BPB
1.1564

Hyperparameters Across PRs

pr_numberbitsscope
375MLP and attention
1221all
1238Q/K int5, V/O and MLP int4 in middle layers; boundary layers int5