Can a Single GPU Fake Distributed Communications?

ezyang · x · 2026-07-17

ezyang proposes a thought experiment for distributed training: take the training configuration of any large LLM, isolate the perspective of a single rank, keep the exact same local compute kernels, but assume you only have one GPU and must "fake" all the inputs originally received via communication.

The question is: can you run 10 steps under these conditions without the loss becoming NaN? This essentially tests the coupling between communication and local computation in distributed training, as well as the numerical stability requirements when reproducing a specific rank on a single GPU.

Related event: Opus 4.8 Fails Single-GPU Distributed Communication Challenge(2 posts)→

Original post →

More from Infra

Infra channel →