ZSE (Zyora Server Engine) is an ultra memory-efficient inference engine tailored for large language models (LLMs). It enables running advanced models such as Qwen, Llama, and Mistral while significantly reducing memory usage without sacrificing performance. At the core of ZSE lies the Intelligence Orchestrator, which offers smart recommendations based on available memory, ensuring optimal utilization.

Key Features

• zAttention: Implements custom CUDA kernels for paged, flash, and sparse attention mechanisms.
• zQuantize: Adopts per-tensor INT2-8 mixed precision quantization for efficient processing.
• zKV: Provides a quantized key-value cache that achieves a fourfold memory reduction.
• zStream: Allows for layer streaming with asynchronous prefetching, enabling substantial performance on limited hardware, for example, running 70 billion parameters on a 24GB GPU.
• zOrchestrator: Delivers intelligent recommendations based on available free memory.
• Efficiency Modes: Offers various operational modes including speed, balanced, memory, and ultra to cater to differing requirements.

Cold Start Benchmark

ZSE boasts rapid cold start times, achieving 3.9 seconds for 7B and 21.4 seconds for 32B models utilizing the .zse format on A100-80GB hardware.

Memory Benchmarks (Verified on A100-80GB)

Quick Start Example

To initiate a server for any HuggingFace model:

zse serve Qwen/Qwen2.5-7B-Instruct

Leverage memory optimization to run larger models by specifying max memory:

zse serve Qwen/Qwen2.5-32B-Instruct --max-memory 24GB

Interactive Chat

Engage with models in a chat format:

zse chat Qwen/Qwen2.5-7B-Instruct

API Server Compatibility

ZSE provides compatibility with OpenAI's API, allowing easy integration in applications:

zse serve Qwen/Qwen2.5-7B-Instruct --port 8000

Utilize the API to generate completions seamlessly.

Deployment Options

ZSE supports various deployment options, including Docker and cloud platforms. Commands are provided for both CPU and GPU setups, including support for model pre-loading. Comprehensive deployment documentation is available for environments such as Railway, Render, and Kubernetes.

In summary, ZSE is engineered for advanced LLM applications, dramatically improving inference efficiency and operational speed, making it suitable for developers and enterprises looking to harness the capabilities of large language models.