trunk/7da85ca47b7aec4f5dc133d856bb8ec5fc5b36c1: Add CUDA stream transport operations (#198918)
- PyTorch: 1780 events in the last 90 days
- PyTorch: 1758th Release in the last 90 days
- Previous: earlier the same day · trunk/90769ae442226c087d2faf307a9be834c5c645e0: Add Auto PR Triage in shadow mode (#195778)
What happened
Add Transport.read_stream / write_stream , which order a transfer on a CUDA stream (default: current) and return once enqueued. Transfers can be captured with torch.cuda.graph . Ordinary read/write are unchanged. The default implementation works for any host-driven backend; backends with native stream support can override it. A host callback schedules write_async / read_async on a shared asyncio loop thread, and cuS…
Summary assembled by rule from the sources below