https://answer.freetools.me/tags/%E8%BF%9E%E7%BB%AD%E6%89%B9%E5%A4%84%E7%90%86/ Recent content in 连续批处理 on Answer Hugo -- 0.152.2 zh-cn Thu, 12 Mar 2026 22:19:43 +0800 https://answer.freetools.me/%E5%A4%A7%E6%A8%A1%E5%9E%8B%E6%8E%A8%E7%90%86%E6%A1%86%E6%9E%B6%E7%9A%84%E6%8A%80%E6%9C%AF%E5%8D%9A%E5%BC%88%E4%BB%8Evllm%E5%88%B0tensorrt-llm%E8%A7%A3%E6%9E%90%E4%B8%89%E5%A4%A7%E6%A1%86%E6%9E%B6%E7%9A%84%E8%AE%BE%E8%AE%A1%E5%93%B2%E5%AD%A6%E4%B8%8E%E6%80%A7%E8%83%BD%E7%AA%81%E5%9B%B4/ Thu, 12 Mar 2026 22:19:43 +0800 https://answer.freetools.me/%E5%A4%A7%E6%A8%A1%E5%9E%8B%E6%8E%A8%E7%90%86%E6%A1%86%E6%9E%B6%E7%9A%84%E6%8A%80%E6%9C%AF%E5%8D%9A%E5%BC%88%E4%BB%8Evllm%E5%88%B0tensorrt-llm%E8%A7%A3%E6%9E%90%E4%B8%89%E5%A4%A7%E6%A1%86%E6%9E%B6%E7%9A%84%E8%AE%BE%E8%AE%A1%E5%93%B2%E5%AD%A6%E4%B8%8E%E6%80%A7%E8%83%BD%E7%AA%81%E5%9B%B4/ 深度解析大模型推理框架的技术演进，从vLLM的PagedAttention到TensorRT-LLM的硬件极致优化，再到llama.cpp的跨平台哲学，揭示三大框架如何在内存管理、批处理策略、算子融合等维度进行技术博弈，以及如何根据业务场景做出正确选择。 https://answer.freetools.me/%E4%B8%80%E4%B8%AA%E8%AF%B7%E6%B1%82%E5%85%88%E7%BB%93%E6%9D%9F%E4%B8%BA%E4%BD%95%E6%95%B4%E6%89%B9%E9%83%BD%E8%A6%81%E7%AD%89%E4%BB%8E%E9%9D%99%E6%80%81%E6%89%B9%E5%A4%84%E7%90%86%E5%88%B0%E8%BF%9E%E7%BB%AD%E6%89%B9%E5%A4%84%E7%90%86%E7%9A%84llm%E6%8E%A8%E7%90%86%E9%9D%A9%E5%91%BD/ Mon, 09 Mar 2026 04:21:14 +0800 https://answer.freetools.me/%E4%B8%80%E4%B8%AA%E8%AF%B7%E6%B1%82%E5%85%88%E7%BB%93%E6%9D%9F%E4%B8%BA%E4%BD%95%E6%95%B4%E6%89%B9%E9%83%BD%E8%A6%81%E7%AD%89%E4%BB%8E%E9%9D%99%E6%80%81%E6%89%B9%E5%A4%84%E7%90%86%E5%88%B0%E8%BF%9E%E7%BB%AD%E6%89%B9%E5%A4%84%E7%90%86%E7%9A%84llm%E6%8E%A8%E7%90%86%E9%9D%A9%E5%91%BD/ 深入解析连续批处理（Continuous Batching）如何解决大模型推理中的GPU利用率问题。从ORCA论文的迭代级调度到vLLM的实现，全面阐述这项让推理吞吐量提升数十倍的核心技术。涵盖prefill与decode阶段、KV缓存、选择性批处理、chunked prefill等关键技术。