wan-ascend-adaptation
Pass
Audited by Gen Agent Trust Hub on Oct 18, 2026
Risk Level: SAFE
Full Analysis
- [SAFE]: The skill provides comprehensive and accurate guidance for a specific engineering task (CUDA to Ascend NPU porting). All technical recommendations, including the use of HCCL, fused operators, and MindIE acceleration libraries, align with established best practices for the Huawei Ascend ecosystem.
- [COMMAND_EXECUTION]: The documentation includes standard command-line examples for model quantization using
msmodelslimand environment variable configuration (e.g.,ALGO,PRECISION). These are benign and necessary for the described workflow. - [DYNAMIC_EXECUTION]: The skill describes advanced monkey-patching techniques to replace standard PyTorch functional operators (like
F.conv3dandF.conv2d) with distributed, patch-parallel versions. This is a legitimate architectural pattern used to optimize memory-intensive tasks like VAE decoding on specialized hardware. - [EXTERNAL_DOWNLOADS]: The skill references several specialized libraries (e.g.,
torch_npu,mindiesd,msmodelslim,yunchang,torch_atb) which are standard dependencies for Ascend NPU development.
Audit Metadata