聆思智能
Official@LISTENAI · China
LISTENAI provides skills for quantization-aware training, ONNX export, and Thinker NPU deployment of neural networks via the Linger toolchain.
Agent Skills by 聆思智能
Showing 3 vetted skills indexed across 1 GitHub repositories.
LISTENAIlnn-quant-training
Builds, diagnoses, and validates Linger quantization-aware training and ONNX export workflows.
LISTENAIthinker-model-pack-validation
Packages quantized ONNX models with tpacker and validates Linger-Thinker inference consistency.
LISTENAIlnn-model-design
Audit and adapt neural network architectures for Linger quantization and Thinker NPU deployment.
Frequently Asked Questions About 聆思智能
FAQPage SchemaWhat tasks can I accomplish with LISTENAI's Linger and Thinker skills?▼
You can build and debug quantization-aware training for Linger 3.x, configure mixed precision and clamp overrides in YAML, calibrate MOM/TQT quantizers, export quantized ONNX models, package them with tpacker for Thinker hardware, and validate inference consistency with tvalidator.
Who are these skills designed for?▼
They target embedded ML engineers and model developers deploying neural networks on LISTENAI's Thinker NPU platforms, especially those converting PyTorch models through the Linger quantization toolchain and diagnosing accuracy or operator-compatibility issues.
What environment is required for Thinker model pack validation?▼
You need Linux, a Python environment with linger, onnx, and pythinker installed, the Thinker source code, tpacker, tvalidator, CMake, and the target platform simulation library to package and validate quantized ONNX models.
How does the skill handle missing deployment requirements?▼
When a request lacks the target platform or model input size and range, the model-design skill stops and asks the user to confirm those details before giving any hardware compatibility conclusion, preventing incorrect deployment guidance.
Can I diagnose accuracy loss after quantization?▼
Yes. The QAT skill supports inspecting checkpoints, comparing float-versus-quant outputs, reviewing YAML quantization settings, and applying local clamp or const_module overrides to recover accuracy when the quantized model misses its target.