Инфраструктура
Serving, scaling and caching the parts around a model: inference servers, containers, orchestration, and the cost structures each one implies at load.
15 technologies
- vLLM tool High-throughput inference server built around paged attention and continuous batching.
- Ollama tool Runs open-weight models locally behind a simple command and HTTP API.
- Inference concept Running a trained model to produce output, as opposed to training it.
- Tokenizer tool Component that converts text to tokens and back.
- Prompt Caching technique Reusing the computed state of an unchanged prompt prefix.
- Safetensors infrastructure File format for model weights designed to avoid arbitrary code execution.
- Ray framework Distributed execution framework for scaling Python workloads.
- llama.cpp tool CPU-first inference for quantised models, in portable C++.
- ONNX infrastructure Interchange format for models, so a trained model runs on another runtime.
- Triton Inference Server tool NVIDIA's server for deploying models from several frameworks behind one API.
- SGLang tool Serving engine built around reusing shared prompt prefixes.
- GGUF infrastructure Quantised model format designed for fast loading and CPU inference.
- TensorRT-LLM tool NVIDIA's compiler and runtime for optimised LLM inference on its GPUs.
- Text Generation Inference tool Hugging Face's serving stack for transformer models.
- KV Cache technique Cached attention keys and values that make generating one token at a time affordable.