Skip to content
Hot eventWatching

Google open-sources Zero-TIM for bit-exact on-policy RL on TPU

1 reports1 reporting sourcesUpdated 1 d ago

Start here

Report summary

Google implemented Zero-TIM in Tunix, enabling the sampler decoding, server-side prefill, and trainer forward pass to produce bitwise-identical logprobs for the same action token under identical weights, and validated it on a 64-chip TPU v5p (DP8×TP8) multi-host mesh.

From Google Developers Blog

Report timeline

Follow the reports to see the event from different sides.

November 16
  1. Google Developers Blog
    Google Achieves Bitwise Zero-Error On-Policy RL for Training and Inference with Zero-TIM on TPUs

    Google implemented Zero-TIM in Tunix, enabling the sampler decoding, server-side prefill, and trainer forward pass to produce bitwise-identical logprobs for the same action token under identical weights, and validated it on a 64-chip TPU v5p (DP8×TP8) multi-host mesh.

Heat for this event

Not enough continuous observations yet to draw a trend.