fix(ascend): synchronize PyTorch operators with NPU streams

[baominghelly]

Summary

• Add an Ascend C10 stream adapter in src/torch/ascend/c10.h using c10_npu::NPUStreamGuard and c10_npu::getStreamFromExternal.
• Enable generated PyTorch slot 8 operators to execute on the caller-provided Ascend stream, with an explicit default-stream fallback when the runtime stream is null.
• Discover and link the installed torch_npu C++ headers and library for Ascend PyTorch builds, and add regression coverage for code generation and NPU stream ordering.

Motivation

Ascend was not included in the generated PyTorch backend's C10 stream-guard path. Consequently, slot 8 operators could execute on a different NPU stream from the InfiniOps runtime stream, which risks incorrect ordering and results when surrounding work is asynchronous.

This change gives Ascend the same stream-guard behavior as the other PyTorch-backed accelerator platforms. It also fails at configure time with a clear diagnostic when the installed torch_npu does not provide the required external-stream API.

Type of Change

• feat — new feature / new operator / new platform
• fix — bug fix
• perf — performance improvement (no behavioral change)
• refactor — code restructuring without behavior change
• test — adding or fixing tests only
• docs — documentation only
• build / ci — build system or CI configuration
• chore — tooling, formatting, or other non-code changes
• Breaking change (requires a ! in the Conventional Commits prefix or a BREAKING CHANGE: footer)

Platforms Affected

• CPU (WITH_CPU)
• NVIDIA (WITH_NVIDIA)
• Iluvatar (WITH_ILUVATAR)
• MetaX (WITH_METAX)
• Cambricon (WITH_CAMBRICON)
• Moore (WITH_MOORE)
• Ascend (WITH_ASCEND)
• PyTorch C++ bindings (WITH_TORCH)
• Build system / CMake / CI
• Python bindings / user-facing API

Smoke Test Result

Ascend 910C, CANN 9.0.0, Python 3.11, PyTorch 2.10.0, and torch_npu 2.10.0.post4:

python -m pytest tests/test_generate_torch_ops.py -q
21 passed in 1.13s

cmake -S /workspace/InfiniOps-ascend-torch-stream-20260902 \
  -B /workspace/ascend-torch-stream-test-20260902/build-6b256e0 \
  -DCMAKE_INSTALL_PREFIX=/workspace/ascend-torch-stream-test-20260902/python3/infini \
  -DPython_EXECUTABLE=/workspace/torch-npu-wheel-audit-20260902/venv-post4/bin/python
cmake --build /workspace/ascend-torch-stream-test-20260902/build-6b256e0 -j16
[100%] Built target ops

python -m pytest tests/test_abs.py --devices ascend -q
20 passed in 1.14s

python -m pytest tests/test_torch_ops.py -k 'tril or triu' --devices ascend -q
18 passed, 9 deselected in 1.08s

The CMake cache used WITH_ASCEND=ON, WITH_CPU=ON, WITH_TORCH=ON, INFINI_RT_ROOT=/workspace/InfiniRT-6b256e0-prefix, and generated abs, tril, and triu PyTorch operators.

A CPU-only regression build using WITH_CPU=ON, WITH_ASCEND=OFF, and WITH_TORCH=ON also completed successfully:

cmake -S /workspace/InfiniOps-ascend-torch-stream-20260902 \
  -B /workspace/ascend-torch-stream-test-20260902/build-cpu-regression-v2
cmake --build /workspace/ascend-torch-stream-test-20260902/build-cpu-regression-v2 -j16
[100%] Built target ops

Test Results on Supported Platforms

Platform	Affected	Build / Smoke Result	Full Result / Notes
NVIDIA	No	N/A - not affected	N/A - generated include remains guarded by WITH_NVIDIA
Iluvatar	No	N/A - not affected	N/A
MetaX	No	N/A - not affected	N/A
Cambricon	No	N/A - not affected	N/A
Moore	No	N/A - not affected	N/A
Ascend	Yes	Build passed; targeted smoke passed	Generator: 21 passed; abs: 20 passed; tril/triu: 18 passed

Additional integration results

Llama-3.2-3B-Instruct: inference passed
ChatGLM3-6B: inference passed
GLM-4-9B-0414: inference passed

Trace output confirmed:
"operator_name": "Triu", "device_type": "ascend", "implementation": 8
"operator_name": "Tril", "device_type": "ascend", "implementation": 8

Benchmark / Performance Impact

N/A. This is a correctness and stream-ordering fix; no performance claim is made.

Notes for Reviewers

• The torch_npu integration is enabled only when both WITH_ASCEND and generated PyTorch sources are present. Non-Ascend builds do not discover or link torch_npu.
• torch_npu 2.10.0.post4 was validated. An older 2.10.0.post2 wheel does not expose c10_npu::getStreamFromExternal; CMake now reports this incompatibility explicitly.
• The model-level tests were performed in a separate integration checkout that also contained the existing Ascend rotary-embedding fix. The slot trace verifies that tril and triu used implementation 8.
• No native Ascend slot 0 implementation for tril or triu is included in this PR.

[voltjia]

这个感觉不是必要的，按理说应该都是相当于转发到 C10 的 getStreamFromExternal 就行，我看别的平台都没加 nullptr 的处理，可以看看昇腾是不是这样。

[baominghelly]

确认过了，昇腾这里和其他平台的行为不同。当前验证的 torch_npu 2.10.0.post4 对应源码提交 5dd8ef3f9b375b5ae4a83538d5785754148c3302，其 c10_npu::getStreamFromExternal 会显式检查 stream != nullptr，否则直接报 External NPU stream does not support nullptr.：

https://github.com/Ascend/pytorch/blob/5dd8ef3f9b375b5ae4a83538d5785754148c3302/torch_npu/csrc/core/npu/NPUStream.cpp

InfiniOps 的 stream=0 会传入空指针来表示默认流，因此这里不能无条件转发；非空流仍然直接调用 getStreamFromExternal。我补充了英文注释说明这个平台差异，并重新验证了 Ascend 构建和 tests/test_abs.py --devices ascend，结果为 20 passed。