Source: vLLM docs: Reproducibility

vLLM's statement of what it does and does not guarantee about reproducible outputs: default seed 0, offline multiprocessing switch, batch invariance for online serving.

Contents

Summary

A short page stating that vLLM trades reproducibility for performance by default, and listing the switches that restore it.

Key claims

  • vLLM does not guarantee reproducible results by default.
  • Offline: set VLLM_ENABLE_V1_MULTIPROCESSING=0 to make scheduling deterministic, or enable batch invariance so outputs ignore scheduling.
  • Online serving: only batch invariance is available.
  • In the V1 engine the seed defaults to 0 and cannot be unset, so runs match even at temperature > 0; workers need consistent random state for features such as speculative decoding.
  • Reproducibility holds only on the same hardware and the same vLLM version.

Feeds

LLM non-determinism

Suggest changes →