What problem does it solve? Deploying a neural network on LNN chips (venus, mars, arcs, venusa) often fails late in the cycle because of unsupported operators, precision mismatches, or memory limits in the Linger/Thinker toolchain. This Skill acts as an architecture gate that audits a model against real toolchain capabilities before training or deployment, and triages failures into structure, split-policy, threshold, or toolchain-defect categories. ## Core Features & Use Cases - Capability Discovery: Scans a local linger/ and thinker/ checkout to extract per-platform operators, precision support, SRAM/PSRAM limits, and tpacker defaults instead of relying on memory. - Static ONNX Audit: Inspects ONNX graphs for op support, dynamic axes, parameter bytes, and oversized tensors, mapping Linger quantized ops to Thinker equivalents. - End-to-End Gate: Runs Linger export, tpacker packaging, and tvalidator/runtime validation, then triages blockers and auto-generates minimal author issue bundles when the toolchain itself appears defective. - Use Case: A customer proposes a Conv2d+GRU model for the venusa platform with dynamic sequence length. The Skill audits the ONNX, runs tpacker with the declared dynamic shape, identifies a weight-chunk overflow, recommends a channel_out split, and verifies packaging passes. ## Quick Start Ask the assistant to evaluate your model architecture or ONNX file against a specific LNN target platform and input shape, for example by requesting a compatibility audit of model.onnx for the venusa platform.
