Chitu

Tracked

A production-focused inference framework for large language models, offering high performance, multi-hardware support, and scalable deployment.

Author thu-pacman Open Sourced 2025-02-20 Last Commit Unknown

Chitu is a high-performance inference framework for large language models developed by Tsinghua University, focusing on efficiency, flexibility, and availability. It delivers production-grade, low-latency LLM inference across a wide range of deployment scenarios from CPU-only and single-GPU setups to large-scale distributed clusters, with multi-vendor hardware compatibility for enterprise adoption.

Multi-Hardware Support

  • Optimized implementations for NVIDIA GPUs and various domestic AI accelerators
  • Mixed-hardware optimization for regulated or cost-sensitive environments
  • Quantization and mixed-precision support including FP4, FP8, and BF16 formats
  • Extensible plugin and adapter architecture ensuring compatibility with diverse backend hardware

Scalable Deployment

  • Deployment spanning single-node heterogeneous CPU/GPU configurations to full distributed cluster environments
  • Streaming and batch optimizations for maximizing throughput in production serving scenarios
  • Production stability engineering for long-term concurrent operation
  • Official container images and comprehensive deployment guides for enterprise adoption

Performance Engineering

  • High-performance operator implementations tuned for mainstream LLMs
  • Batched serving optimizations for high-throughput inference endpoints
  • Developer guides and performance benchmarks for hardware selection and capacity planning
  • Compatible with enterprise Q&A systems, real-time online inference services, and batched model serving