[Gxj230958]

fastvideo.utils.current_stream() returns None on non-CUDA platforms, so PyHcclCommunicator cannot obtain the NPU stream needed for warmup or default collective calls. Use torch.npu.current_stream(self.device) for communicator warmup, all_reduce, and broadcast, while preserving explicitly supplied streams.

The CPU regression tests live in fastvideo/tests/distributed/test_pyhccl_stream.py and are explicitly collected by .buildkite/scripts/unit_test.sh. Five cases cover communicator-device selection, explicit streams, and initialization warmup synchronization. NPU and HCCL calls are mocked in these tests, so they run in the existing GPU CI environment without Ascend hardware.

Validation

• Buildkite Fastcheck #2277: all six lanes and fastcheck-passed passed at f68a261. Remote documentation and pre-commit checks also passed.
• Stream regression tests: 5 passed.
• CI collection and merge-plan contract tests: 59 passed.
• Real two-rank Ascend 910B2C verification on devices 8 and 9: communicator initialization and warmup, all-reduce and broadcast on a non-default current stream, and both operations with an explicit stream all passed with exact expected values. Runtime: PyTorch 2.9.0 and torch-npu 2.9.0.post1.
• Two-rank Wan2.1-1.3B training completed communicator initialization and three training updates.
• Applicable repository pre-commit hooks passed, including mypy.
• Documentation type checking and formatting passed: npm run check --prefix docs-site and npm run format:check --prefix docs-site.

python -m pytest --confcutdir=fastvideo/tests/distributed \
  fastvideo/tests/distributed/test_pyhccl_stream.py -q
python -m pytest --confcutdir=fastvideo/tests/contract \
  fastvideo/tests/contract/test_ci_test_collection.py \
  fastvideo/tests/contract/test_merge_ci_plan.py -q

CI configuration approval

Maintainer sign-off is required for the change to .buildkite/scripts/unit_test.sh. The change adds fastvideo/tests/distributed/test_pyhccl_stream.py to the existing unit lane, alongside the existing test_sp_warmup_device.py entry.

Before merge, the maintainer needs to explicitly approve this CI configuration change, provide an approving review, and trigger the required merge gate for full-suite-passed.

[mergify]

Merge Protections

🔴 1 of 1 protections blocking · waiting on 👀 reviews and 🤖 CI

	Protection	Waiting on
🔴	PR merge requirements	👀 reviews and 🤖 CI

🔴 PR merge requirements

Waiting for

• #approved-reviews-by>=1
• check-success=full-suite-passed

This rule is failing.

• #approved-reviews-by>=1
• check-success=full-suite-passed
• check-success=fastcheck-passed
• check-success~=pre-commit
• title~=(?i)^\[(feat|feature|bugfix|fix|refactor|perf|ci|doc|docs|misc|chore|kernel|new.?model|skill|skills|infra)\]

[greptile-apps]

Confidence Score: 5/5

[Medium impact] The PR appears safe to merge; no actionable issues were found.

Summary

PyHcclCommunicator now gets the current NPU stream for its bound device.

• PyHcclCommunicator uses its device's current NPU stream for HCCL calls.

No actionable issues found. Tests were not run during this review.

Acknowledged by Gxj230958: remote CI reports an unrelated formatting failure in docs-site/islands/Cookbook.tsx and a microscope test failure whose cause is unknown because the logs require organization access.

_{Reviews (1) · Last reviewed commit: "[bugfix] Use NPU current streams in PyHc..." · Reviewed by Greptile}

[fv-superagent]

fv-superagent score: 4/5

The unit lane is red only on a known flake, not on this change.

Build #2268 ends 1 failed, 3590 passed; the failure is test_cudnn_sdpa_dense.py::test_cudnn_route_is_bit_identical_to_cudnn_dispatched_sdpa, a bf16 bit-identity comparison, and that same test passed in build #2265 the same day. This PR touches only pyhccl.py and a new local test, so it cannot affect cuDNN SDPA. /test unit reruns the lane.

On the change: replacing the CUDA-only fastvideo.utils.current_stream() with torch.npu.current_stream(self.device) in PyHcclCommunicator, including the warmup path, is right, and binding the stream to self.device is what the NPU needs.

One note: tests/local_tests/test_pyhccl_stream.py is not collected by any CI lane (tests/local_tests/ is local-only per tests/README.md), so the two regression cases run only by hand. The two-rank Ascend run is the evidence that matters and cannot be reproduced here.

The build check shows a 2026-11-02 failure in the docs-site type step; this PR does not touch docs-site/, so treat that as stale.

— fv-superagent · cudnn flakes, pyhccl gets the blame · run by aryan5v

[fv-superagent]

fv-superagent score: 4/5 (was 4/5)

Re-reviewed head f68a261. You moved the regression test to fastvideo/tests/distributed/test_pyhccl_stream.py and added just that file to .buildkite/scripts/unit_test.sh — the same pattern as the existing test_sp_warmup_device.py entry — and the unit lane passed in build #2277 (fastcheck-passed green). The flake I flagged is gone, and the fix itself is unchanged and right.

One note for whoever merges: .buildkite/scripts/unit_test.sh is CI configuration, so adding the file to the lane needs a maintainer's sign-off, not just a code review.

— fv-superagent · the flake has left the building · run by aryan5v