← Back to Quantization

mixed int5/int6 with QAT

Quantization
Used in
1 PRs
Best BPB
1.4222
Avg BPB
1.4222

Submissions

PR #619by zeal175
1.4222

Hyperparameters Across PRs

pr_numberbitsscope
619—int5 for MLP weights, int6 for attention/bigram-sensitive weights