EXL3 drafters for Qwen3.8-27B in TandemLLM
TandemLLM uses these files for speculative decoding with the EXL3 quants of Qwen3.8-27B. Every EXL3 size uses the same files.
| file | what it is | size |
|---|---|---|
exl3-drafters-k3/ft-b8-v2.safetensors |
our block-8 drafter, DFlash2-b8, in EXL3 at 3 bits | 0.65 GB |
exl3-drafters-k3/ft-b16.safetensors |
our block-16 drafter, DFlash2-b16, in EXL3 at 3 bits | 0.65 GB |
exl3-draft-head-k4/model.safetensors |
the head the drafters use to pick tokens, in EXL3 at 4 bits | 0.64 GB |
A drafter only guesses tokens, and the target model checks every guess, so these smaller files change the speed and never the text.
How to use them
TandemLLM 0.8.1 and later download them for you: ops/use-profile.sh exl3-<size> fetches the quant, these files, our BF16 drafters and the base checkpoint, and checks every file. They land in ~/.qwen38-spark-engine/weights/, with the two folders as they are here. To get them by hand:
hf download 0xBakeer/TandemLLM-Qwen3.8-27B-EXL3-drafters --local-dir ~/.qwen38-spark-engine/weights
Other EXL3 loaders don't need these files; they only run the quant itself.