Skill UI

This guide details the use of the GGUF format and quantization techniques for running large language models (LLMs) efficiently. GGUF enables optimal inference on consumer hardware, Apple Silicon, and systems requiring CPU-only operation. By compressing models using various K-quant methods (like Q4_K_M), developers can significantly reduce memory footprint and hardware requirements while maintaining high performance.

Mamba is a selective state-space model architecture delivering O(n) inference for million-token sequences, offering hardware-aware Mamba-1 and Mamba-2 variants that beat Transformers on speed without KV cache overhead.

Compress large language models quickly using Wanda, SparseGPT, and N:M pruning to enforce hardware-friendly sparsity, enabling faster inference and deployment without retraining.

A comprehensive guide demonstrating the full lifecycle of GPU resource utilization on Vast.ai. This includes searching for available hardware using CLI and REST APIs, provisioning a dedicated GPU instance (e.g., using PyTorch images), connecting via SSH, running deep learning workloads, and finally, securely destroying the instance to stop billing. Ideal for testing cloud integration or managing ML compute pipelines.

Hardware-agnostic quantum ML framework enabling gradient-based training of quantum circuits, hybrid quantum-classical models, and seamless integration with PyTorch/JAX/TensorFlow alongside device portability across IBM/Google/Rigetti/IonQ.

An expert guide for deploying and optimizing Large Language Models (LLMs) entirely on local hardware. Mastering the entire local AI ecosystem, including inference engines like Ollama, vLLM, and llama.cpp. Provides deep knowledge on quantization formats (GGUF, EXL2, AWQ), VRAM calculation, and ensuring privacy-first, offline deployment for state-of-the-art models.

This skill is designed for running comprehensive evaluations of Hugging Face Hub models on local hardware. It supports multiple evaluation frameworks, including `inspect-ai` and `lighteval`, and provides flexibility in backend selection (e.g., `vllm`, Transformers, or `accelerate`) to optimize for performance and compatibility. Use this for local GPU benchmarking, smoke testing, and backend selection, but note that it does not handle remote job orchestration or result publishing.

Hardware-agnostic quantum machine learning framework with automatic differentiation for training quantum circuits, building hybrid quantum-classical models, and deploying across IBM, Google, Rigetti, IonQ, and simulators.

This comprehensive skill manages the entire computer vision pipeline, from custom dataset annotation to model deployment. Users can fine-tune YOLO models on proprietary COCO-formatted datasets using an agent-driven workflow. Key features include hardware-aware training (supporting CUDA, MPS, ROCm), automatic export to optimized formats (TensorRT, CoreML, OpenVINO, ONNX), and seamless, one-click deployment as an active detection skill.

A state-of-the-art solution for real-time object detection using the advanced YOLO 2026 model. This skill detects 80+ COCO object classes (people, vehicles, animals) and outputs bounding boxes with high precision. It features automatic hardware acceleration, optimizing the model for various platforms (CUDA, MPS, OpenVINO, ROCm) to ensure maximum speed and efficiency, whether running on CPU, GPU, or Neural Engine.

This skill provides real-time object detection capabilities by leveraging the Google Coral Edge TPU accelerator. It operates natively within Windows Subsystem for Linux (WSL), allowing high-speed, low-latency inference on live camera frames. By utilizing hardware acceleration, it achieves impressive performance (e.g., ~4ms inference at 320x320), making it ideal for edge computing projects requiring rapid visual analysis and resource efficiency.

This skill provides high-performance, real-time object detection using the Intel OpenVINO runtime. It operates via a Docker container, ensuring cross-platform compatibility. Supports detection across various Intel accelerators (NCS2, iGPU, Arc GPU) and CPUs. Users can customize parameters like confidence threshold, target classes, processing FPS, and model precision to optimize performance for specific hardware and accuracy needs.