Source: vLLM docs: V1 User Guide
Official guide to the vLLM V1 engine's goals, removed V0 features and behavior differences.
Contents
- What: official documentation (latest, read 2026-11-14)
- Author: vLLM project
- Date: 2026-11-14 (accessed)
- Original: vLLM docs: V1 User Guide
Summary
States V1’s goals (hackable code, near-zero CPU overhead, optimizations on by default) and lists what changed for users relative to V0: features dropped, different logprob semantics, and per-model-family support status.
Key claims
- V1 keeps V0’s models, kernels and utilities but re-architects scheduler, KV cache manager, worker, sampler and API server; V0 is fully deprecated.
- Removed:
best_of, per-request logits processors (replaced by global ones set at startup), GPU-to-CPU KV swapping for preemption, request-level structured-output backend choice. - Logprobs default to raw model output, before temperature or penalties;
--logprobs-modeselects raw or processed logprobs/logits. - Requests asking for prompt logprobs bypass the prefix cache and recompute the full prefill.
- Chunked prefill is on by default where possible; CUDA graph capture uses more memory than in V0; FCFS and priority scheduling via
--scheduling-policy. - Hardware: NVIDIA, AMD, Intel GPU, TPU and CPU functional; other accelerators through plugins (e.g. vllm-ascend, vllm-gaudi).
- Mamba and hybrid models work but without prefix caching; most encoder-decoder models go through plugins (Whisper is native); pooling models fully supported.