The GitHub Actions job "CI" on tvm-ffi.git/ascend-addon has failed.
Run started by GitHub user voidLitchi (triggered by tqchen).

Head commit for run:
5d065ef48fec68def2a65f3cc389a9dca5dceee6 / Yuzhi Zhang 
<[email protected]>
[FEAT] Add Ascend stream support to the Torch C-DLPack addon

## Motivation

PyTorch exposes Ascend tensors through its PrivateUse1 backend and exports them 
as kDLExtDev through DLPack. The existing Torch C-DLPack addon can convert 
these tensors, but its current_work_stream callback only handles CUDA and ROCm. 
TVM-FFI therefore receives a null stream for Ascend and may launch a kernel on 
the default NPU stream instead of the caller's current torch.npu stream.

## Changes

- Detect Ascend after CUDA and ROCm on Linux, and use an Ascend-specific addon 
even when PyTorch already exposes a DLPack C exchange API.
- Preserve the upstream PyTorch API preference for CUDA and ROCm.
- Add a mutually exclusive --build-with-ascend option and reject it on Windows 
and macOS.
- Discover torch_npu headers and libraries from the installed package, define 
BUILD_WITH_ASCEND, and link libtorch_npu.
- Return c10_npu::getCurrentNPUStream(device_id).stream() for kDLExtDev from 
the addon's current_work_stream callback.
- Reuse the upstream ABI-aware addon cache key and Torch-version-based C++ 
standard selection.

## Testing

- Add unit coverage for backend selection, platform fallback, existing API 
preference, and mutually exclusive build options.
- Add an Ascend runtime test that compares current_work_stream against a 
non-default torch.npu stream.
- Verify the modified Python files compile and run standalone selection and 
argument-parsing checks on Windows. The Ascend runtime test requires a Linux 
host with torch_npu and Ascend hardware.

Report URL: https://github.com/apache/tvm-ffi/actions/runs/30275239262

With regards,
GitHub Actions via GitBox


---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]

Reply via email to