skills.rest
ntd237
ntd237COMMUNITY
Explore all skills by this author →
GitHub

convert-model

Plans and ranks ML model conversion paths across ONNX, TensorRT, OpenVINO, TFLite, Core ML, RKNN, Hailo, and ncnn runtimes.

1|Updated Apr 20, 2026
Search Tags:#quantization#edge deployment#onnx#tensorrt#model conversion#openvino#inference runtime
One-click install
npx skills add https://github.com/ntd237/draft_toolkits_20042026 --skill convert-model-ntd237
Or copy as Structured Prompt for Agent▼
Please help me install this Agent Skill.
Skill: convert-model
Source: https://github.com/ntd237/draft_toolkits_20042026/tree/main/.agents/skills/convert-model
Command: npx skills add https://github.com/ntd237/draft_toolkits_20042026 --skill convert-model-ntd237

SYSTEM DOCUMENTATION & REQUIREMENTS

💡 This Skill includes references (resource) components.

What problem does it solve? Choosing a model export target without checking hardware, toolchain, and graph compatibility leads to failed conversions, broken deployments, and wasted effort. This Skill separates the requested file format from the real runtime target, validates feasibility, and recommends ranked conversion paths that actually fit the deployment environment. ## Core Features & Use Cases - Feasibility Verification: Checks hardware fit (NVIDIA GPU for TensorRT, Intel for OpenVINO, Rockchip NPU for RKNN), library availability, and model-graph compatibility before recommending any target, delivering a clear verdict: feasible, feasible with setup, or not a good fit. - Ranked Conversion Paths: Generates ranked alternatives (default 3, or exactly N when requested) covering the best fit, a conservative fallback, and a portable fallback like ONNX or TorchScript, each with conversion chain, prerequisites, and trade-offs. - Direct Conversion Mode: When both source and target formats are specified, keeps the requested route as primary and only broadens to alternatives if the route is blocked. - Use Case: You have a PyTorch .pt model and want TensorRT output, but your machine is CPU-only. The Skill flags TensorRT as hardware-bound, then ranks ONNX, OpenVINO IR, and TorchScript as viable paths with full conversion chains and configuration guidance. ## Quick Start Ask the assistant to recommend the best conversion paths for your model, for example: "I have a PyTorch .pt model and need the best 3 conversion options for deployment on a Rockchip NPU board."

Frequently Asked Questions about convert-model

High-intent search queries and answers about installing and using this skill.

FAQPage Schema
How do I convert a PyTorch model to ONNX?▼

PyTorch to ONNX conversion uses the standard PyTorch export stack with ONNX tooling, and a generic CPU is sufficient for export. Key configuration points include choosing a conservative opset version, defining input shapes, and enabling dynamic axes only when truly needed.

What is the best conversion path for edge deployment?▼

The best edge conversion path depends on the target hardware family: TFLite or ncnn for mobile and embedded CPUs, RKNN for Rockchip NPUs, and Hailo for Hailo accelerators. When hardware is undecided, export to ONNX first to preserve optionality before branching into vendor-specific compilers.

Can I use TensorRT without an NVIDIA GPU?▼

TensorRT requires NVIDIA GPU hardware along with matching CUDA, cuDNN, and TensorRT versions, so it is not a good fit for CPU-only machines. On CPU-only systems, ONNX, OpenVINO IR, or TorchScript are more viable alternatives depending on the deployment target.

ONNX vs TensorRT vs OpenVINO, which runtime should I choose?▼

Choose TensorRT for NVIDIA GPU deployment, OpenVINO for Intel CPU or iGPU targets, and ONNX as a portable interchange format when the final runtime is undecided. The decision should follow the actual production hardware, not just the requested file extension.

Why does my converted model produce wrong predictions?▼

Wrong predictions after successful export usually come from preprocessing mismatches such as NCHW versus NHWC layout, RGB versus BGR channel order, normalization constants, or post-processing differences. Compare preprocessing and post-processing behavior between the original and converted model before blaming the runtime.

When should I use INT8 quantization for model conversion?▼

INT8 quantization is appropriate only when the target runtime and hardware benefit from it and a representative calibration dataset reflecting real deployment inputs exists. Without calibration data, FP16 or FP32 should rank higher since poor calibration can severely damage accuracy.

Related Skills

Explore More Software Engineering Skills→
gracefullightgracefullight

tensorflow-neural-networks

Build and train neural networks using TensorFlow Keras APIs and custom layer implementations.

Community
Intermediate
agentic-inagentic-in
577

Inference

Select inference paths, serving stacks, and benchmarking plans for ML models.

Official
Intermediate
mattmirelesmattmireles
57

coreml-validate

Validate CoreML model numerical accuracy against PyTorch reference models.

Community
Advanced
doanchienthangdevdoanchienthangdev
4

model-deployment

Deploy machine learning models to production with Docker and Kubernetes.

Community
Advanced