Google ZERO-TIM Achieves Bitwise-Identical Logprobs on 64 TPU v5p Chips

·2026.11.16 09:00

Key point

The method uses consistent arithmetic and standardized reduction orders across TPU meshes to ensure true on-policy stability.

Details

  • Google's ZERO-TIM implementation achieves bitwise-identical logprobs for sampler and trainer across 64 TPU v5p chips.
  • The solution addresses floating-point addition errors inherent in separate serving and training paths.
  • Stability is ensured by using consistent arithmetic operations and standardized reduction orders across TPU meshes.
  • The approach guarantees true on-policy stability in RL training by eliminating numerical discrepancies.

This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.

Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.