huggingface/safetensors
Safetensors
Safetensors is a secure tensor serialization format and library designed for storing and distributing model weights. Its primary purpose is to provide a safe file format for machine learning tensors that prevents the execution of arbitrary or malicious code during the deserialization process.
The project is distinguished by its use of zero-copy memory mapping, which reads data from disk directly into memory to minimize overhead. It enables cross-framework compatibility, allowing tensor data to be serialized and deserialized across different machine learning libraries.
The system covers high-level capabilities for tensor metadata management, using a header-based approach to parse shapes and auxiliary information without loading the full data array. It also supports efficient tensor slicing by calculating byte offsets to load only specific portions of a tensor into memory.
Features
- Safe Tensor Formats - Provides a secure serialization format for neural network weights that prevents arbitrary code execution during loading.
- Universal Model Weight Distribution - Stores tensor data in a universal format that ensures compatibility across different machine learning libraries.
- Machine Learning Model Formats - Implements a portable file format for model weights that eliminates security risks associated with code execution.
- Secure Weight Formats - Saves model weights in a secure format that prevents malicious code execution during the loading process.
- Zero-Copy Weight Loaders - Employs memory mapping and zero-copy mechanisms to import model weights with maximum efficiency.
- Tensor Libraries - Provides a library for serializing and deserializing tensor data across disparate machine learning environments.
- Zero-Copy Memory Mappings - Maps file contents directly into memory addresses to eliminate redundant data copying between disk and application.
- Zero-Copy Deserialization - Reads serialized tensors using zero-copy deserialization to minimize memory usage and increase processing speed.
- Binary Execution Restrictions - Ensures tensors are stored as raw byte arrays to prevent the execution of arbitrary code during deserialization.
- Framework-Agnostic Formats - Provides a standardized binary layout enabling tensor data to be read across different machine learning libraries.
- Tensor Data Mapping - Enables loading of tensor data structures across different machine learning frameworks for environment compatibility.
- Tensor-Metadata Bundles - Extracts descriptive tensor information and shapes without loading the entire data array into system memory.
- Tensor - Implements a JSON header at the start of files to allow reading tensor shapes without loading all data.
- Byte Sequence Storage - Organizes tensor data in contiguous byte sequences to ensure compatibility with low-level memory access.
- Tensor Slicing - Calculates byte offsets to extract specific portions of a tensor directly from a file.
- Data Storage Systems - Stores and distributes tensors safely and efficiently.
