The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill references the installation of standard, well-known Python packages such as diffusers, transformers, accelerate, and peft from official PyPI registries. Documentation and weight references point to trusted organizations including HuggingFace and GitHub.
[COMMAND_EXECUTION]: Local execution is facilitated through scripts like run_pipeline.py and benchmark_pipeline.py. Multi-card distributed inference utilizes torchrun and the hccl backend, which are standard utilities for the Ascend NPU ecosystem. Input arguments are used for configuration within these local processes.
[PROMPT_INJECTION]: The skill ingests user-provided text prompts to drive image and video generation pipelines. This constitutes an indirect prompt injection surface as described in Category 8:
Ingestion points: The args.prompt parameter in run_pipeline.py, benchmark_pipeline.py, and run_context_parallel.py.
Boundary markers: No specific delimiters or safety warnings are applied to the user-supplied prompt before it is passed to the diffusion models.
Capability inventory: The skill possesses the capability to execute model inference on NPU/CPU and write generated files (PNG, MP4, JSON) to the local filesystem.
Sanitization: No explicit sanitization or filtering of the prompt text is performed before processing by the model.
[DYNAMIC_EXECUTION]: In run_pipeline.py and benchmark_pipeline.py, the skill dynamically loads pipeline classes from the diffusers library using getattr based on the --pipeline-class argument. This is a common and legitimate design pattern for supporting diverse model architectures within the Diffusers framework.

diffusers-ascend-pipeline