gistfile1.txt

---
  CUDA Streams + torch.compile: GitHub Issues

  Feature Requests / Support Gaps

  #: #118204 (https://github.com/pytorch/pytorch/issues/118204)
  Title: streams x torch.compile: stream is treated as None sometimes
  Filed: Jan 2024
  By: zou3519
  Status: Closed
  Now Supported?: Yes -- current_stream() and stream variables now survive graph breaks
  ────────────────────────────────────────
  #: #176154 (https://github.com/pytorch/pytorch/issues/176154)
  Title: StreamContextVariable's create_proxy logic removed in Dynamo post-2.9,
  breaking
    custom backends relying on set_stream nodes
  Filed: Mar 2026
  By: BobBob-Fu
  Status: Closed
  Now Supported?: Yes -- regression from your PR #166472
    (https://github.com/pytorch/pytorch/pull/166472); stream context now uses FX
    annotations instead of set_stream proxy nodes
  ────────────────────────────────────────
  #: #178132 (https://github.com/pytorch/pytorch/issues/178132)
  Title: torch.compile streams support in AOTI
  Filed: Mar 2026
  By: mlazos (you)
  Status: Open
  Now Supported?: Not yet -- AOTI needs C++ equivalent of Python object lookup table
  for
    streams/events
  ────────────────────────────────────────
  #: #179858 (https://github.com/pytorch/pytorch/issues/179858)
  Title: torch.compile doesn't use async all_to_all
  Filed: Apr 2026
  By: embg
  Status: Open
  Now Supported?: Not yet -- all_to_all isn't converted to async by dynamo, unlike
    all_reduce/all_gather

  Bugs Filed Against Stream Support (your PRs fixed)

  #: #177691 (https://github.com/pytorch/pytorch/issues/177691)
  Title: torch.compile not tracing streams properly
  Filed: Mar 2026
  By: benchislett
  Status: Closed (assigned mlazos)
  ────────────────────────────────────────
  #: #180179 (https://github.com/pytorch/pytorch/issues/180179)
  Title: torch.cuda.Stream traced incorrectly by torch.compile
  Filed: Apr 2026
  By: borisfom
  Status: Closed (assigned mlazos) -- non-default stream stream_id passed instead of
    cuda_stream pointer, causing segfault
  ────────────────────────────────────────
  #: #180396 (https://github.com/pytorch/pytorch/issues/180396)
  Title: [inductor] Stream handling in torch.compile with reduce-overhead mode is
  broken
  Filed: Apr 2026
  By: tirthasheshpatel
  Status: Closed (assigned mlazos, pt2: ubn) -- stream fork during CUDA graph capture
    caused cudaErrorStreamCaptureUnsupported
  ────────────────────────────────────────
  #: #182084 (https://github.com/pytorch/pytorch/issues/182084)
  Title: Torch.compile crashes with multi-stream and triton kernels
  Filed: May 2026
  By: benchislett
  Status: Closed (assigned mlazos) -- codegen naming collision between stream1 (Stream
    object) and stream1 (raw int handle) on device >= 1
  ────────────────────────────────────────
  #: #185546 (https://github.com/pytorch/pytorch/issues/185546)
  Title: torch.compile(mode='reduce-overhead') fails when user model uses wait_stream
  Filed: May 2026
  By: tfogal
  Status: Closed (assigned mlazos) -- cudaErrorStreamCaptureUnjoined when side stream
    work not properly joined back

  Still Open Issues

  #: #188405 (https://github.com/pytorch/pytorch/issues/188405)
  Title: torch.compile reconstructs graph-created torch.Stream/torch.Event subclass as
    base class across graph break         
  Filed: Jun 2026
  By: Stevennnn                           
  Status: Open -- affects out-of-tree backends using device-agnostic Stream/Event
    subclasses
  ────────────────────────────────────────
  #: #188536 (https://github.com/pytorch/pytorch/issues/188536)
  Title: Graph-break reconstruction loses identity of graph-created torch.Event
    subclasses
  Filed: Jun 2026
  By: Stevennnn
  Status: Open -- record() and wait() end up on different event objects after graph
    break
  ────────────────────────────────────────
  #: #155679 (https://github.com/pytorch/pytorch/issues/155679)
  Title: Torch compile CUDA graphs leads to a large number of CUDA streams
  Filed: Jun 2025
  By: prajwal1210
  Status: Open -- 113 streams created when combining reduce-overhead with distributed
    all-reduce
  ────────────────────────────────────────
  #: #151199 (https://github.com/pytorch/pytorch/issues/151199)
  Title: [CUDA Graph tree] Cannot capture buffer allocation on side CUDA Streams
  Filed: Apr 2025
  By: lirundong
  Status: Closed -- graph tree assumed single-stream memory model, side-stream
    allocations failed validation

  Pre-existing Requests (older)

  #: #25540 (https://github.com/pytorch/pytorch/issues/25540)
  Title: Enabling multi-stream per-thread default stream
  Filed: Sep 2019
  Status: Closed
  ────────────────────────────────────────
  #: #59692 (https://github.com/pytorch/pytorch/issues/59692)
  Title: CUDA streams run sequentially, expected to run parallel
  Filed: Jun 2021
  Status: Open