What problem does it solve? Verifying whether a new model architecture can run on Ascend NPU normally requires downloading tens of gigabytes of real weights, which is slow and wasteful when you only need to confirm architectural compatibility, operator integration feasibility, or memory and latency estimates. ## Core Features & Use Cases - Random-Weight Construction: Build diffusers/transformers components (transformer, text encoder, VAE, tokenizer) from config files with random weights, assemble a pipeline, and run inference on NPU. - Footprint Reduction: Shrink models via num_layers trimming, disabling CFG, skipping VAE decode, and config caching to fit NPU memory and speed up verification. - Operator & Fusion Pre-Check: Test candidate NPU operators and compile-time fusion patterns on a small dummy graph before committing to real-weight integration, with mandatory compile-vs-eager dual reports. - Use Case: Before integrating a new fusion operator for MiniMax-H3, run a 2-layer dummy model to confirm the pattern hits the actual compiled graph, then only proceed to real-weight kernel diff and wall-clock verification if it does. ## Quick Start Ask the assistant to verify whether a specific model architecture can run on NPU using a dummy run with random weights and reduced layers.
