[API Compatibility] Add torch-compatible NVTX range context manager

[ShigureNyako]

PR Category

User Experience

PR Types

Improvements

Description

This PR adds a torch-compatible NVTX range context manager to Paddle's existing NVTX namespace:

• paddle.cuda.nvtx.range(msg, *args, **kwargs)
• paddle.device.nvtx.range(msg, *args, **kwargs)

The wrapper follows torch.cuda.nvtx.range behavior: it formats msg, calls the existing range_push, yields as a context manager/decorator, and always calls range_pop in finally. It does not introduce a torch dependency and preserves Paddle's current CUDA/NVTX availability behavior by reusing the existing range_push / range_pop APIs.

Background: in NVIDIA/cudnn-frontend's Paddle adaptation, with torch.cuda.nvtx.range(...) call sites had to be replaced by a local context manager wrapping paddle.cuda.nvtx.range_push/pop. Adding this compatibility wrapper lets such external projects keep the torch-style call shape when using import paddle as torch.

Validation:

• pre-commit run --files python/paddle/cuda/__init__.py python/paddle/device/__init__.py test/legacy_test/test_cuda_unittest.py
• git diff --check
• python3 - <<'PY' ... compile edited files ... PY
• Targeted TestNvtx runtime was not runnable in this source checkout because there is no built libpaddle; CUDA/NVTX runtime behavior remains covered by the existing guarded CUDA test path plus the new monkeypatched no-CUDA behavior test.

是否引起精度变化

否

[ShigureNyako]

Self-review:

• Diff is limited to paddle.cuda.nvtx.range, mirrored paddle.device.nvtx.range, and focused TestNvtx coverage.
• The wrapper matches torch.cuda.nvtx.range call shape (msg.format(*args, **kwargs)) and guarantees range_pop() in finally for normal and exceptional exits.
• No torch dependency is introduced; no stream/current_stream code is touched.

@SigureMo 请帮忙 review，谢谢。

[SigureMo]

Suggested change

	def range(msg: str, *args, **kwargs):
	def range(msg: str, *args: Any, **kwargs: Any) -> generator 的返回值咋标来着太久了我忘了，你自己查去:

[ShigureNyako]

Done，已在 bedbde59a5daedb2e00463161c8fc9e44db9bb5b 中补上类型标注：

• *args: Any
• **kwargs: Any
• -> Generator[None, None, None]

Generator 按 Paddle 现有 contextmanager 写法放在 TYPE_CHECKING 下从 collections.abc 引入，并同步更新了 paddle.cuda.nvtx.range / paddle.device.nvtx.range。

验证：pre-commit run --files python/paddle/cuda/__init__.py python/paddle/device/__init__.py test/legacy_test/test_cuda_unittest.py 通过。

[SigureMo]

LGTMeow

[ShigureNyako]

/re-run all-failed

[ShigureNyako]

/re-run all-failed

[ShigureNyako]

/re-run all-failed

[codecov-commenter]

Codecov Report

✅ All modified and coverable lines are covered by tests.
⚠️ Please upload report for BASE (develop@8df3d8a). Learn more about missing BASE report.

Additional details and impacted files

@@             Coverage Diff             @@
##             develop    #79251   +/-   ##
===========================================
  Coverage           ?   100.00%           
===========================================
  Files              ?         2           
  Lines              ?        18           
  Branches           ?         0           
===========================================
  Hits               ?        18           
  Misses             ?         0           
  Partials           ?         0           

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[ShigureNyako]

/re-run all-failed

[ShigureNyako]

/re-run all-failed

[ShigureNyako]

/re-run all-failed

[SigureMo]

停停停，先别 rerun 了，得等 #79240 修好了我给你 rerun

[ShigureNyako]

收到，先不继续 rerun 了，等 #79240 修好后再处理。

[ShigureNyako]

CI 已全部通过，PR 现在是 approved / clean。我这边没有 merge 权限（MergePullRequest 权限不足），麻烦 @SigureMo 合并一下，谢谢。