Limite 1B Violetto drafter

Draft model for speculative decoding of Limite 1B Violetto, built for the silicon-violetto Metal engine.

Usage

Setup silicon-violetto then:

uv run hf download Belluxx/limite-1b-violetto-drafter drafter.vlt --local-dir data/spec
./target/release/violetto serve --weights models/limite-int8.vlt --drafter data/spec/drafter.vlt

Speed

Decode tok/s on an M4 Max (40 GPU cores). Setup: int8 weights, temperature 0.6.

Every context and batch size is in the engine's performance.md.

Output should match plain decoding, except where rounding flips near ties.

Model

• Architecture: 2 layers on the target's layers 35 and 47, with convolutions and context injection, a 1024-token attention window and a 16,384-token head
• Training: KL to the target on its own traces (103M positions, 400M anchors)
• Decoding: the engine verifies a tree of the drafts in one step
• Acceptance: tokens per step on held-out traces:

For more infos and reproduction steps, check here.

Files