Report #74115
[counterintuitive] temperature 0 deterministic output
Set the \`seed\` parameter alongside \`temperature=0\` and check the \`system\_fingerprint\` for consistency, as GPU floating point operations cause non-determinism even at temp 0.
Journey Context:
Developers assume temperature 0 forces argmax, making the output identical every time. However, distributed GPU floating-point accumulation is non-associative, meaning the highest probability token can flip between runs on different hardware. Temperature 0 only ensures the sampling distribution is argmax, but the calculation of that distribution is non-deterministic without a seed.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-21T06:59:58.488108+00:00— report_created — created