Add devlog: Cross-stream memory management: wait_event instead of record_stream

[weifengpy]

New post under content/eager/: Cross-stream memory management: wait_event instead of record_stream.

What it covers

• Why producer/consumer pipelining across CUDA streams (e.g. FSDP2's all-gather → forward) needs explicit lifetime management: the caching allocator only tracks the stream a block was allocated on.
• Why Tensor.record_stream is correct but hands block reuse to the allocator's event polling. Peak memory then depends on how far the CPU runs ahead, with links into CUDACachingAllocator.cpp.
• Two fixes: stall the producer with wait_event, and the recommended one, keep a Python reference and wait_event on the allocation stream before del.
• How FSDP2 applies this in forward (AllGatherState) and backward (ReduceScatterState, set_reduce_scatter_max_input_buffers).

Files

• content/eager/2026-09-29-cross-stream-memory-without-record-stream.md
• static/images/eager/cross-stream-*.svg: 9 swimlane diagrams with light and dark variants via prefers-color-scheme

Verification

• All PyTorch source links are pinned to 1a0b56b8 on main, and each claim was checked against the code there.
• The demo script (in the post's appendix) was run on an H100:
  • no lifetime management corrupts 50/50 trials;
  • record_stream, stall and keepalive all pass;
  • they use at most 4, 1 and 2 param blocks per run, respectively.
• hugo --minify builds cleanly, and the page was checked in headless Chromium in light and dark mode.