[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"$f2zhnl9w4vbrw7":3},{"code":4,"message":5,"data":6,"timestamp":72},10000,"success",{"full_name":7,"position":8,"description":9,"language":10,"stars":11,"forks":12,"stars_gained":13,"age_days":14,"hn_threads":15,"one_liner":42,"summary_md":43,"topic":44,"github_topics":47,"license":68,"homepage":69,"updated_at":70,"history":71},"vllm-project\u002Fvllm",6,"A high-throughput and memory-efficient inference and serving engine for LLMs","Python",95322,23894,1363,1353,[16,22,28,33,38],{"id":17,"title":18,"points":19,"comments":20,"posted_at":21},49492067,"vLLM v0.28.0",108,43,"2026-08-29",{"id":23,"title":24,"points":25,"comments":26,"posted_at":27},47643924,"vLLM introduces memory optimizations for long-context inference",5,0,"2026-04-04",{"id":29,"title":30,"points":31,"comments":26,"posted_at":32},47681076,"vLLM IR: A Functional Intermediate Representation for vLLM",4,"2026-04-07",{"id":34,"title":35,"points":36,"comments":26,"posted_at":37},49625693,"vLLM 0.29.0: Model Runner V2 is now the default",3,"2026-09-09",{"id":39,"title":40,"points":36,"comments":26,"posted_at":41},37453478,"Vllm: High-throughput and memory-efficient inference and serving engine for LLMs","2023-09-10","给自建 LLM 服务的团队提供高吞吐推理与在线批处理能力。","## 解决什么问题\n\nDemo 跑通之后，瓶颈往往不在模型，而在服务层：一张 GPU 能扛多少并发、长 prompt 怎么切、KV cache 涨起来时延迟怎么稳住。直接用原生 PyTorch 循环写生成，硬件大部分时间是闲着的。vLLM 就是针对这个问题做的推理与 serving 引擎，核心是 PagedAttention 管理 KV 内存，加上对进来的请求做 continuous batching。\n\n## 有什么不同\n\n和 Hugging Face `transformers` 的 generate 相比，后者简单易用但不是一个 serving 系统；vLLM 面向并发设计：分页 KV cache、chunked prefill、prefix caching、CUDA\u002FHIP graph，以及一整套量化 kernel。和闭源 serving 方案相比，它开放且与模型解耦——README 称支持 200+ Hugging Face 架构，覆盖 dense、MoE、Mamba 混合、多模态、embedding 和 reward 模型，并提供 OpenAI 兼容 API 以及 Anthropic Messages \u002F gRPC 接口。代价是硬件面铺得很广（NVIDIA、AMD、Intel GPU、CPU，还有 TPU、Gaudi、Ascend 等插件），kernel 面也很宽；这正是它的价值，但也意味着实际表现高度依赖你的 GPU 型号和模型组合。\n\n## 适合谁\n\n自己托管 LLM、在意单位成本吞吐的团队：搭推理服务的平台\u002FMLOps 工程师、做模型评测的研究者，以及跑 MoE 或多模态模型、在其他引擎上比较别扭的人。只调 API 不碰权重的人，或者想要一个有厂商兜底的成品设备而不是自己运维项目的团队，不适合。\n\n## 上手成本\n\n需要 Python 和匹配版本的 PyTorch\u002FCUDA（或 ROCm）工具链，显存要按模型加 KV cache 预留；安装命令是 `uv pip install vllm`。依赖树不小，需要和驱动版本对齐；启动速度可以靠预编译 kernel（`vllm download-kernels`）改善。Apache-2.0 许可宽松，但项目迭代很快：小版本之间 API 和参数常变，HotRepo 窗口内的讨论热度更多反映发布节奏，而不是沉寂。\n\n## 点评\n\nvLLM 不是这个赛道的新玩家，而是定义了这条赛道的那批项目之一。2023-02 创建至今积累了约 95k star、24k fork，2026-08 关于 v0.28.0 的 Hacker News 讨论说明它仍在持续发布有分量的版本。单周 +1,363 star（周榜 #15）是健康的动能，但追踪期内它没上过日榜——这里的关注是工程驱动的稳定增长，不是病毒式冲高。这是一个成熟且绑定较深的选择：认真做自建推理就值得投入，但不适合当周末玩具。",{"slug":45,"name":46},"machine-learning","机器学习",[48,49,50,51,52,53,54,55,56,57,58,59,60,61,62,63,64,65,66,67],"amd","blackwell","cuda","deepseek","deepseek-v3","gpt","gpt-oss","inference","kimi","llama","llm","llm-serving","model-serving","moe","openai","pytorch","qwen","qwen3","tpu","transformer","Apache-2.0","https:\u002F\u002Fvllm.ai","2026-10-24T18:07:06Z",[],1792879622]