vLLM Splits Prefill and Decode, TileRT Takes Over Decode
vllm_project · x · 2026-07-15
The vLLM team shared a specific use case: after decoupling vLLM's prefill and TileRT's decode, the decode side can be hot-swapped based on the workload.
Key Points
- The TileRT team integrated this via the vLLM V1 connector interface, requiring no changes to vLLM code.
- It defaults to vLLM's decode to ensure throughput; for scenarios prioritizing low latency, like agents or real-time assistants, it can switch to TileRT decode.
- Both sides maintain the same OpenAI-compatible interface and prefix caching; switching is essentially just a routing change.
Performance Data
- TileRT reported that on GLM-5.1-FP8, using 8× B200 + MTP, single-user decode reached around 618 tok/s.
- This is roughly 2 times faster than its baseline without MTP.
- At peak acceptance rates, speed approaches 800 tok/s.
The authors thanked TileRT and inferact for their collaboration.
Related event: vLLM and TileRT Introduce Decoupled Inference Stack(3 posts)→
More from Infra
- SF Compute founder: buying compute is 'an absolutely awful experience' right now — IgorCarron · 2026-09-11
- SmolVM open-sources persistent computer infrastructure for agents that outlive chat sessions — aniketmaurya · 2026-09-11
- PyTorch Day Korea 2026 launches first offline conf, CFP closes Sept 13 — PyTorch · 2026-09-11
- Local LLM server dilemma: 4x CMP-170HX (price up 53% in 20 days) vs Mac Studio M5 Ultra — rumboll · 2026-09-11
- llama.cpp lands Flash Attention tuning for RDNA4, big prefill gains on AMD — pmttyji · 2026-09-11
- Your p99 latency benchmark may be lying: a deep dive into coordinated omission — Franc0Fernand0 · 2026-09-11