← Back to Quantization

mixed int6/int5

Quantization
Used in
5 PRs
Best BPB
1.0924
Avg BPB
1.2534

Submissions

PR #504by fleeb83
1.5275
PR #1123by sisegod
1.1986
PR #1170by Christopher-Lee-McClendon
1.1199
PR #1279by dexhunter
1.0924
PR #1543by PavelPaha
1.3286

Hyperparameters Across PRs

pr_numberbitsscope
504—mlp, attn, bigram, trigram
1123—Q/K, V/O, MLP, embeddings
1170—MLP layers
1279—all
1543—attention weights and MLP weights