← Field notes
Field note / 01Kernels
What we turned off to make Helios faster on one H100
A Helios engineering note on disabling unstable FP8, keeping FlashAttention-3 and compiled inference, and verifying H100 workers before traffic.
FP8 looked like the headline win until minute two of every run, when accumulated quantization error collapsed the stream. It ships disabled.
What held instead: FlashAttention-3, a compiled hot path, and a playout ladder that stopped throwing generated frames away.
Then there is the box lottery. Nominally identical H100s do not behave alike, so an engine-verify gate now runs before any box takes traffic.
Windflow EngineeringRuntime team
Field notes are short on purpose: one decision, what it cost, what we kept. The long write-ups land here when they are ready.