Crafted with ❤️ by Devs Do Code (Sree)
GGUF Technical Specifications
Delve into the intricacies of GGUF, a meticulously crafted format that builds upon the robust foundation of the GGJT model. Tailored for heightened extensibility and user-centric functionality, GGUF introduces a suite of indispensable features:
Single-file Deployment: Streamline distribution and loading effortlessly. GGUF models have been meticulously architected for seamless deployment, necessitating no external files for supplementary information.
Extensibility: Safeguard the future of your models. GGUF seamlessly accommodates the integration of new features into GGML-based executors, ensuring compatibility with existing models.
mmap Compatibility: Prioritize efficiency. GGUF models are purposefully engineered to support mmap, facilitating rapid loading and saving, thus optimizing your workflow.
User-Friendly: Simplify your coding endeavors. Load and save models effortlessly, irrespective of the programming language used, obviating the dependency on external libraries.
Full Information: A comprehensive repository in a single file. GGUF models encapsulate all requisite information for loading, eliminating the need for users to furnish additional data.
The differentiator between GGJT and GGUF lies in the deliberate adoption of a key-value structure for hyperparameters (now termed metadata). Bid farewell to untyped lists, and embrace a structured approach that seamlessly accommodates new metadata without compromising compatibility with existing models. Augment your model with supplementary information for enhanced inference and model identification.
QUANTIZATION_METHODS:
| Method | Quantization | Advantages | Trade-offs |
|---|---|---|---|
| q2_k | 2-bit integers | Significant model size reduction | Minimal impact on accuracy |
| q3_k_l | 3-bit integers | Balance between model size reduction and accuracy preservation | Moderate impact on accuracy |
| q3_k_m | 3-bit integers | Enhanced accuracy with mixed precision | Increased computational complexity |
| q3_k_s | 3-bit integers | Improved model efficiency with structured pruning | Reduced accuracy |
| q4_0 | 4-bit integers | Significant model size reduction | Moderate impact on accuracy |
| q4_1 | 4-bit integers | Enhanced accuracy with mixed precision | Increased computational complexity |
| q4_k_m | 4-bit integers | Optimized model size and accuracy with mixed precision and structured pruning | Reduced accuracy |
| q4_k_s | 4-bit integers | Improved model efficiency with structured pruning | Reduced accuracy |
| q5_0 | 5-bit integers | Balance between model size reduction and accuracy preservation | Moderate impact on accuracy |
| q5_1 | 5-bit integers | Enhanced accuracy with mixed precision | Increased computational complexity |
| q5_k_m | 5-bit integers | Optimized model size and accuracy with mixed precision and structured pruning | Reduced accuracy |
| q5_k_s | 5-bit integers | Improved model efficiency with structured pruning | Reduced accuracy |
| q6_k | 6-bit integers | Balance between model size reduction and accuracy preservation | Moderate impact on accuracy |
| q8_0 | 8-bit integers | Significant model size reduction | Minimal impact on accuracy |
- Downloads last month
- 9