4 ms·
Non-determinism is configurable at the level of the mathematical model, but current production systems do not support deterministic evaluation of LLMs.
by danpalmer 6mo ago
Non-determinism is configurable at the level of the mathematical model, but current production systems do not support deterministic evaluation of LLMs.
- orbital-decay 6mo agoThey do, though. Providers don't because batching makes it cheaper. Among the providers, DeepSeek seems to support it for v4 (and have actually optimized their kernels for batching), and Gemini Flash is "almost deterministic".
- danpalmer 6mo agoI'm pretty sure that the determinism issue is at the floating point math level, or even the hardware level. Just disabling batching and reducing the temperature to 0 does not result in truly deterministic answers.
- orbital-decay 6mo agoFP math itself is deterministic on real hardware, if the order of operations stays the same. Output reproducibility is much less of a problem than it seems, see for example https://docs.vllm.ai/en/latest/usage/reproducibility/ https://docs.vllm.ai/en/latest/usage/reproducibility/
- nnevatie 6mo agoThe FP math is deterministic. However, the environments in which inference is run and specifically batching make current LLM services practically non-deterministic.