What problem does it solve? Deploying quantized ONNX models on Thinker hardware (Venus, ARCS, VenusA) requires iterative packing with tpacker, tuning memory and split thresholds to reduce hardware operator splitting, and verifying that Thinker inference matches the Linger reference. Doing this manually is error-prone and slow. ## Core Features & Use Cases - Automated Packing & Tuning: Builds a packing baseline with tpacker, analyzes operator splits between pre/post-split graphs, and iteratively optimizes threshold1-4, DMA prefetch, and tensor placement within real memory limits. - Consistency Validation: Runs tvalidator with fixed inputs to confirm Linger and Thinker produce matching outputs, including multi-resource DAG end-to-end checks. - Failure Reporting: Generates a minimal reproducible failure report with environment info, exact commands, and the first actionable error when packing or validation fails. - Use Case: You have a quantized ONNX model exported from Linger targeting the ARCS platform. The skill packs it within the 393216-byte shared memory limit, minimizes Conv/Linear splitting, and verifies bit-level consistency with the Linger reference. ## Quick Start Pack my quantized model model.onnx for the venus platform with tpacker, minimize operator splitting within the real memory limits, and validate consistency with tvalidator.
