← Back to Quantization

fp16

Quantization
Used in
26 PRs
Best BPB
1.1361
Avg BPB
1.1885

Hyperparameters Across PRs

pr_numberbitsscope
4216tied embeddings / output head
6016tied embeddings
6316tied embeddings passthrough
8916tied embeddings / logit head
9516embeddings
10716tied embeddings
11416tied embedding and last 2 layers' key projections
15116embeddings
15516tied embeddings
16316embeddings
16616tied embeddings
18616tied embedding and late-K layers
19116tied embeddings and last two c_k weights
25116embeddings
26716tied embeddings and last-layer key projections
27116embeddings
28416tied embeddings
35116tied embeddings and small tensors
35516tok_emb.weight
37216token embedding and last layer c_k
38116embeddings
43416tied embeddings
45216embeddings
51516tied embeddings passthrough
102716tied embeddings
104816selected embedding rows