Google releases Zero-TIM for Tunix on TPU v5p
Google researchers have developed Zero-TIM, a modular implementation within Tunix that eliminates train–inference numerical mismatch in large language model reinforcement learning on Cloud TPU. By aligning within-shard kernel accumulation, ordering cross-shard reductions, standardizing logprob scoring, and wrapping serving kernels with custom differentiation rules, the system achieves bitwise-identical logprobs across sampler, prefill, and trainer paths. Verified on a 64 TPU v5p mesh, the technique improves on-policy training stability and solve rates. Open article →