vLLM Natively Supports Kimi Acceleration Scheme
vllm_project · x · 2026-07-13
vLLM and novitalabs are pushing the speculative decoding ecosystem for Kimi models.
- novitalabs trained and open-sourced DSpark speculators for Kimi-K2.6 and Kimi-K2.7-Code.
- DSpark is a spec decoding method for DeepSeek: it drafts an entire token block at once, which is then verified by the main model, accelerating decoding speed.
- vLLM natively supports this scheme starting in v0.25.0, meaning it can be directly plugged in to accelerate inference.
- novita's batch size=1 benchmarks show:
- Kimi-K2.6 average throughput increased by 2.55x (+155%)
- Kimi-K2.7-Code average throughput increased by 2.36x (+136%)
More from Infra
- Encrypted GPT-2 speedrun challenges FHE inference: best is 3.031 s/token, goal sub-1s — tensorqt · 2026-10-27
- Open-source Latch compiles PyTorch into training megakernels, 2.75x faster on H100 — sguada · 2026-10-27
- NVIDIA Claims HGX Rubin NVL8 Delivers 10x More Tokens per Megawatt Than B200 — thursdai_pod · 2026-10-27
- Together lets multiple LoRA post-training experiments share the same GPU capacity — togethercompute · 2026-10-27
- Commoditization dilemma: either human labor or GPUs become the commodity — _AustinCalvert_ · 2026-10-27
- DeepSpeed shows AutoEP/AutoTP/AutoSP parallelism gains, hosts office hours — PyTorch · 2026-10-27