$ the-wire · showcase
Gloo deadlock fixed, Inductor keeps stream annotations
By RepoJournal · Filed · About PyTorch · Composed from the cited sources · methodology
Two correctness fixes in core PyTorch lead the day: a Gloo process group that could deadlock when it dropped the last reference to a finished work item, and an Inductor pass that ran reinserted scatter kernels on the default stream under a side-stream context, racing with the side stream.
[c10d] Fix Gloo runLoop deadlock: release finished works outside workMutex_ (#206050) (#206050) pytorch/pytorch
ProcessGroupGloo::runLoop re-took workMutex_ and then dropped its references to finished AsyncWork, so with the worker holding the last reference, ~AsyncWork ran under the mutex and could need the GIL to release a tensor. Finished works are now released outside the lock, which removes a deadlock path for anyone running Gloo collectives from Python threads.
[inductor] Keep user stream annotations on scatters rebuilt by reinplacing (#205925) pytorch/pytorch
Mutating a tensor view inside a torch.cuda.stream(s) block could make Inductor run the resulting slice_scatter kernels on the default stream with no event waits, producing stale or uninitialized values. Reinplacing now preserves the user's stream annotations on scatters it rebuilds; no BC break.
Run the vec_test_all_types tests in CI and fix fused multiply-add for Half and BFloat16 (#204074) pytorch/pytorch
at::vec::fmadd, fmsub, fnmadd and fnmsub on float16 and bfloat16 vectors now compute the exact product plus addend in float before rounding down, except float16 with FP16 vector arithmetic on aarch64. The vec_test_all_types tests now run in CI, so these numerics stay covered.
[qwen3_8_fn] Remove the deleted ModelConfigConverter from build_model_config (#6314) pytorch/torchtitan
main stopped importing Flash-Next after the ModelConfigConverter and the converters argument to build_model_config were deleted, because torchtitan/models/qwen3_8_flash_next/flavors.py still imported the removed names: the Flash-Next stack merged roughly 2.5 hours after that removal, against an older CI base. The stale imports are gone, restoring the import path.
[BugFix] Copy offline data in order in prefill_replay_buffer pytorch/rl
Around the long tail: prefill_replay_buffer sampled with replacement, so the documented copy duplicated and dropped transitions; buffers with one-dimensional local storage and a non-filtering sampler are now copied once each, oldest first after a round-robin wrap.