← Back to events
ActiveAIRelease

viable/strict/1790175101: [Pipeline] Support outer gradient accumulation (#196641)

Photo: PyTorch Releases

What happened

Add a finalize_gradients step option so callers can keep FSDP accumulation open across complete pipeline schedules. Use the public FSDP finalizer. FSDP finalization reshards parameters, but the pipeline runtime kept the stage marked as unsharded. Track stages retained across calls separately and update the runtime state when reduction reshards a stage. This preserves the duplicate-unshard checks without forcing anot…

Summary assembled by rule from the sources below

Why it's spreading

Sources