Software-Delivered AI Inference

Neural Magic helps developers in accelerating deep learning performance using automated model compression technologies and inference engines. Download our compression-aware inference engines and open source tools for fast model inference.

• nm-vllm: A high-throughput and memory-efficient inference engine for LLMs, our supported enterprise distribution of vLLM.
• DeepSparse: Inference runtime offering accelerated performance on CPUs and APIs to integrate ML into your application
• llm-compressor: HF-compatible library for applying various quantization and sparsity algorithms to llms for optimized deployment with vLLM

In this profile we provide accurate model checkpoints compressed with SOTA methods ready to run in vLLM such as W4A16, W8A16, W8A8 (int8 and fp8), and many more! If you would like help quantizing a model or have a request for us to add a checkpoint, please open an issue in https://github.com/vllm-project/llm-compressor.