← Back to Quantization

int4

Quantization
Used in
15 PRs
Best BPB
1.0599
Avg BPB
1.1670

Hyperparameters Across PRs

pr_numberbitsscope
3054MLP and attention weights
3754all
4774bigram logit table
4824bigram logit table
4854bigram logit table
17324MLP FC2 and residuals
18444skip_gates and skip_weights
18554LQER correction
18744LQER asymmetric per-group correction
20264LQER correction
20314LQER correction
20384model weights
20404model export
20444exported model
20564all