Loading…
大模型推理的模型选型、部署规划与硬件后端适配 —— 判断「这个模型能不能跑在这台设备上」「换硬件/换后端后性能差多少」「该选哪个后端」。当用户要求评估设备可行性、做模型选型、算内存预算、移植模型到新硬件（GPU/CPU/NPU/移动 SoC/Apple Silicon）、对比不同推理后端的性能、或问「这个模型跑得动吗」「显存够不够」「换台机器要改什么」时，务必使用本技能。它基于七步闭环，把可行性判断变成算术问题而不是经验判断。
llm-backend-adaptation is an agent skill entry in ChatHome Skills, sourced from GitHub Skill Search and scored 58/100. # 模型选型、部署规划与后端适配 ## 这个技能的核心判断 选型与适配失败的典型形态不是「跑不起来」，而是**跑到一半才发现不行**： - 权重装得下，但长上下文时 KV Cache 把显存挤爆 - 换了硬件后性能掉一半，但说不清是硬件差异还是配置没适配 - 把 GPU 的调优经验直接搬到 CPU 上，得出方向相反的结. As an agent skill, it packages a reusable instruction set that AI coding agents such as Codex and Claude Code can load to perform this task without hand-written glue code. ChatHome Skills discovered this entry through its daily ingest of official registries and GitHub discovery queries, and re-verifies repository facts, freshness signals, and install metadata on every run. The quality score weighs official-registry presence, a working install command, license clarity, GitHub star trajectory, and repository freshness — the full formula is on the Sources page. The install command above is ready to copy. Review the last-pushed date and repository status before adding llm-backend-adaptation to a production agent workflow; the snapshot facts below are point-in-time values from the most recent ingest.
codex skill install zhuzhichaoTM/hello_inference_engineReview the upstream source before running this command.
Scan coverage: Install metadata and the indexed document body.
Automated signals from install metadata and a bounded skill body — not a full code security audit and not a recursive source or dependency review.
# 模型选型、部署规划与后端适配 ## 这个技能的核心判断 选型与适配失败的典型形态不是「跑不起来」，而是**跑到一半才发现不行**： - 权重装得下，但长上下文时 KV Cache 把显存挤爆 - 换了硬件后性能掉一半，但说不清是硬件差异还是配置没适配 - 把 GPU 的调优经验直接搬到 CPU 上，得出方向相反的结论 共同根因是：**没有先算再测**。本