Support for guided decoding in vLLM
Open
external
r0.6.0
vllm
x-futurehouse
- Dominant language
- Python
- Stars
- 2k
- Forks
- 561
- Avg merge
- 4d 5h
- Merged PRs (30d)
- 145
Description
Feature request:
NeMo-RL's vLLM integration currently doesn't support guided decoding parameters (guided_json, guided_choice, etc.) available in vLLM.
This would enable reliable tool calling and structured JSON output, improving performance with models like Llama-3-Instruct.
Implementation looks straightforward - just need to pass guided decoding params through VllmGenerationWorker._build_sampling_params to vLLM's SamplingParams.
Contributor guide
Assessment
This issue has not been assessed yet.