[Feature] Simplify `tree_speculative_sampling_target_only` · sgl-project/sglang#12562

2025-11-03T15:02:49.000Z

### Checklist - [x] 1. If the issue you raised is not a feature but a question, please raise a discussion at https://github.com/sgl-project/sglang/discussions/new/choose Otherwise, it will be closed. - [x] 2. Please use English, otherwise it will be closed. ### Motivation Some TODOs for newcomers to SGLang - Remove passing argument `draft_probs` - Add a more efficient kernel (torch/triton/cuda) for topk=1 special case. ### Related resources _No response_

(2 commenti) (0 reazioni) (0 assegnatari)Python (28.442 star) (6216 fork)auto 404

good first issue

Descrizione

Checklist

1. If the issue you raised is not a feature but a question, please raise a discussion at https://github.com/sgl-project/sglang/discussions/new/choose Otherwise, it will be closed.
2. Please use English, otherwise it will be closed.

Motivation

Some TODOs for newcomers to SGLang

Remove passing argument draft_probs
Add a more efficient kernel (torch/triton/cuda) for topk=1 special case.

Related resources

No response

Guida contributor

Tech stack: pythonpytorch
Dominio: backendperformance
Tipo issue: feature
Difficoltà: 3
Tempo stimato: 1-2 days
Stato attività: fresh
Chiarezza: mostly clear
Prerequisiti: PythonPyTorch
Adatta ai principianti: 50
Direzione di ricerca: Individua la funzione tree speculative sampling target only nel codice, rimuovi il parametro draft probs dalla sua firma e da tutte le chiamate, quindi implementa un kernel personalizzato efficiente (usando torch, triton o CUDA) per il caso speciale topk=1 per migliorare le prestazioni.