serves pytorch-path chapters 01 · 06 · 08 (and the xla path's collectives chapter) · streaks check themselves off · all floors

Predict the machine, then look.

Three stations, every answer measured, never invented: forty-eight view chains run on real torch, eight compile scenarios verdicted by dynamo's own counters, and five collectives captured from a real four-process gloo run. Predict before the reveal; the streaks keep score in your browser only.

Active recall, one tap

x = torch.arange(32.).reshape(4, 8)
y = x.flatten()

def f(x):
    return torch.sin(x) + 1

cf = torch.compile(f)
cf(torch.randn(4))
cf(torch.randn(4))

Name the collective

Four ranks, their tensors before and after one collective, captured from a real four-process gloo run on this machine. The skill is reading what moved: who ended up equal, who kept a partial, who got a slice. It serves the pytorch distributed chapter and the xla path's collectives chapter alike, because the contract does not care which framework called it.