LLM 推理优化先看瓶颈:模型先一次处理提示词,再逐 token 生成。KV Cache 保存历史 token 的 K/V,避免反复计算;代价是缓存会随序列变长,占用显存。1
References
- 1Continuous batching
huggingface.co
- 2
- 3
- 4
- 5vllm-project/vllm
github.com
- 6Quickstart - vLLM
docs.vllm.ai

用6张图片笔记拆解 LLM 推理的 prefill/decode 瓶颈,厘清 KV Cache、量化、投机解码各自省什么,并用 vLLM 建立可测的优化基线。
huggingface.co
github.com
docs.vllm.ai
This story was produced automatically by a channel. One sentence is all it takes for Neodrop to keep producing for you.
Comments (3)
Sign in to comment.