功能可用性与存储保证取决于所选引擎、模型、平台和后端,需先对照支持矩阵;安装命令、镜像标签和引擎版本不在 README 中,需查阅官方文档;许可证为 MIT 附加额外条件,使用前需阅读 LICENSE
unified-cache-management
ModelEngine-GroupMIT活跃维护把 LLM 推理中的 KV Cache 持久化到 DRAM、SSD 或 NFS 并跨请求复用,减少重复计算,加快推理速度,兼容昇腾 NPU 和 CUDA GPU。
Persist and reuse KV Cache to speedup your LLM.
README 未给出安装命令,安装产物与命令维护在文档 Quickstart 中,需先查阅支持矩阵选择兼容配置README 未说明具体显存、CPU 或 GPU 型号要求,仅提到通过外部存储缓解加速器内存压力,实际可用配置需查阅支持矩阵。
在推理引擎与存储之间统一管理 KV Cache,让兼容请求和实例复用已有计算
把 LLM 推理中的 KV Cache 持久化到 DRAM、SSD 或 NFS 并跨请求复用,减少重复计算,加快推理速度,兼容昇腾 NPU 和 CUDA GPU。
核心能力
- 在推理引擎与存储之间提供统一 KV Cache 层
- 跨请求与跨推理实例复用已有 KV Cache
- 通过外部存储扩展缓存容量,缓解显存压力
- 引擎适配器、缓存机制与存储后端可插拔独立演进
- 支持 PD 传输等缓存移动形式与稀疏注意力加载策略
容易踩的坑
- nightly/v0.9.0-20261109-1pre-release2026-11-08 · 2026-11-08 22:18 UTC
查看更新说明
Status:
completeBackend Wheels: 20
vLLM OpenAI
Runtime:
docker.io/vllm/vllm-openaiRuntime capability Upstream Runtime tagsdocker.io/vllm/vllm-openai: Runtime tags Python ABI Wheel CUDA 12.9 v0.27.1-cu129v0.27.1-cu129-ubuntu2404v0.30.0-cu129v0.30.0-cu129-ubuntu2404— cp312aarch64x86_64 CUDA 13.0 v0.27.1v0.27.1-ubuntu2404v0.30.0v0.30.0-ubuntu2404— cp312aarch64x86_64 vLLM-Ascend
Runtime:
quay.io/ascend/vllm-ascendRuntime capability Upstream Runtime tagsquay.io/ascend/vllm-ascend: Runtime tags Python ABI Wheel CANN 9.1.0 / A2 v0.27.1rc1v0.27.1rc1-openeuler— cp312aarch64x86_64 CANN 9.1.0 / A3 v0.27.1rc1-a3v0.27.1rc1-a3-openeuler— cp312aarch64x86_64 CANN 9.1.0 / A5 v0.27.1rc1-a5v0.27.1rc1-a5-openeuler— cp312aarch64x86_64 CANN 9.2.1 / A2 v0.30.0v0.30.0-openeuler— cp312aarch64x86_64 CANN 9.2.1 / A3 v0.30.0-a3v0.30.0-a3-openeuler— cp312aarch64x86_64 CANN 9.2.1 / A5 v0.30.0-a5v0.30.0-a5-openeuler— cp312[aarch64](https://github.com/ModelEngine- - nightly/v0.9.0-20261108-1pre-release2026-11-07 · 2026-11-07 21:59 UTC
查看更新说明
Status:
completeBackend Wheels: 20
vLLM OpenAI
Runtime:
docker.io/vllm/vllm-openaiRuntime capability Upstream Runtime tagsdocker.io/vllm/vllm-openai: Runtime tags Python ABI Wheel CUDA 12.9 v0.27.1-cu129v0.27.1-cu129-ubuntu2404v0.30.0-cu129v0.30.0-cu129-ubuntu2404— cp312aarch64x86_64 CUDA 13.0 v0.27.1v0.27.1-ubuntu2404v0.30.0v0.30.0-ubuntu2404— cp312aarch64x86_64 vLLM-Ascend
Runtime:
quay.io/ascend/vllm-ascendRuntime capability Upstream Runtime tagsquay.io/ascend/vllm-ascend: Runtime tags Python ABI Wheel CANN 9.1.0 / A2 v0.27.1rc1v0.27.1rc1-openeuler— cp312aarch64x86_64 CANN 9.1.0 / A3 v0.27.1rc1-a3v0.27.1rc1-a3-openeuler— cp312aarch64x86_64 CANN 9.1.0 / A5 v0.27.1rc1-a5v0.27.1rc1-a5-openeuler— cp312aarch64x86_64 CANN 9.2.1 / A2 v0.30.0v0.30.0-openeuler— cp312aarch64x86_64 CANN 9.2.1 / A3 v0.30.0-a3v0.30.0-a3-openeuler— cp312aarch64x86_64 CANN 9.2.1 / A5 v0.30.0-a5v0.30.0-a5-openeuler— cp312[aarch64](https://github.com/ModelEngine- - nightly/v0.9.0-20261107-1pre-release2026-11-06 · 2026-11-06 23:18 UTC
查看更新说明
Status:
completeBackend Wheels: 20
vLLM OpenAI
Runtime:
docker.io/vllm/vllm-openaiRuntime capability Upstream Runtime tagsdocker.io/vllm/vllm-openai: Runtime tags Python ABI Wheel CUDA 12.9 v0.27.1-cu129v0.27.1-cu129-ubuntu2404v0.30.0-cu129v0.30.0-cu129-ubuntu2404— cp312aarch64x86_64 CUDA 13.0 v0.27.1v0.27.1-ubuntu2404v0.30.0v0.30.0-ubuntu2404— cp312aarch64x86_64 vLLM-Ascend
Runtime:
quay.io/ascend/vllm-ascendRuntime capability Upstream Runtime tagsquay.io/ascend/vllm-ascend: Runtime tags Python ABI Wheel CANN 9.1.0 / A2 v0.27.1rc1v0.27.1rc1-openeuler— cp312aarch64x86_64 CANN 9.1.0 / A3 v0.27.1rc1-a3v0.27.1rc1-a3-openeuler— cp312aarch64x86_64 CANN 9.1.0 / A5 v0.27.1rc1-a5v0.27.1rc1-a5-openeuler— cp312aarch64x86_64 CANN 9.2.1 / A2 v0.30.0v0.30.0-openeuler— cp312aarch64x86_64 CANN 9.2.1 / A3 v0.30.0-a3v0.30.0-a3-openeuler— cp312aarch64x86_64 CANN 9.2.1 / A5 v0.30.0-a5v0.30.0-a5-openeuler— cp312[aarch64](https://github.com/ModelEngine- - nightly/v0.9.0-20261104-1pre-release2026-11-03 · 2026-11-03 23:33 UTC
查看更新说明
Status:
completeBackend Wheels: 14
vLLM OpenAI
Runtime:
docker.io/vllm/vllm-openaiRuntime capability Upstream Runtime tagsdocker.io/vllm/vllm-openai: Runtime tags Python ABI Wheel CUDA 12.9 v0.30.0-cu129v0.30.0-cu129-ubuntu2404v0.31.0-cu129v0.31.0-cu129-ubuntu2404— cp312aarch64x86_64 CUDA 13.0 v0.30.0v0.30.0-ubuntu2404v0.31.0v0.31.0-ubuntu2404— cp312aarch64x86_64 vLLM-Ascend
Runtime:
quay.io/ascend/vllm-ascendRuntime capability Upstream Runtime tagsquay.io/ascend/vllm-ascend: Runtime tags Python ABI Wheel CANN 9.2.1 / A2 v0.30.0v0.30.0-openeuler— cp312aarch64x86_64 CANN 9.2.1 / A3 v0.30.0-a3v0.30.0-a3-openeuler— cp312aarch64x86_64 CANN 9.2.1 / A5 v0.30.0-a5v0.30.0-a5-openeuler— cp312aarch64x86_64 SGLang
Runtime:
docker.io/lmsysorg/sglangRuntime capability Upstream Runtime tagsdocker.io/lmsysorg/sglang: Runtime tags Python ABI Wheel CANN 9.0.0 / A2 v0.5.18-cann9.0.0-910bv0.5.19-cann9.0.0-910bv0.5.20-cann9.0.0-910b— cp311aarch64x86_64 CANN 9.0.0 / A3 v0.5.18-cann9.0.0-a3v0.5.19-cann9.0.0-a3v0.5.20-cann9.0.0-a3— cp311aarch64[x86_64](https://github.com/ModelEngine-Group/unified- - v0.9.0最新正式版2026-11-03 · 2026-11-03 01:03 UTC
Status:
completeBackend Wheels: 20 Chart: unified-cache-chart-0.9.0.tgz
vLLM OpenAI
Runtime:
docker.io/vllm/vllm-openaiImages: GHCR
ghcr.io/modelengine-group/vllm-openai· Docker Hubdocker.io/unifiedcachemanager/vllm-openai— 8 image families / 16 architecture membersRuntime capability Upstream Runtime tagsdocker.io/vllm/vllm-openai: Runtime tagsGHCR: ghcr.io/modelengine-group/vllm-openaiDocker Hub: docker.io/unifiedcachemanager/vllm-openai Python ABI Wheel CUDA 12.9 v0.27.1-cu129v0.27.1-cu129-ubuntu2404v0.30.0-cu129v0.30.0-cu129-ubuntu2404v0.27.1-cu129-ucm-0.9.0v0.27.1-cu129-ubuntu2404-ucm-0.9.0v0.30.0-cu129-ucm-0.9.0v0.30.0-cu129-ubuntu2404-ucm-0.9.0cp312PyPI pip install "uc-manager[cu129]==0.9.0"CUDA 13.0 v0.27.1v0.27.1-ubuntu2404v0.30.0v0.30.0-ubuntu2404v0.27.1-ucm-0.9.0v0.27.1-ubuntu2404-ucm-0.9.0v0.30.0-ucm-0.9.0v0.30.0-ubuntu2404-ucm-0.9.0cp312PyPI pip install "uc-manager[cu130]==0.9.0"vLLM-Ascend
Runtime:
quay.io/ascend/vllm-ascendImages: GHCR
ghcr.io/modelengine-group/vllm-ascend· Docker Hubdocker.io/unifiedcachemanager/vllm-ascend— 12 image families / 24 architecture membersRuntime capability Upstream Runtime tagsquay.io/ascend/vllm-ascend: Runtime tagsGHCR: ghcr.io/modelengine-group/vllm-ascendDocker Hub: docker.io/unifiedcachemanager/vllm-ascend Python ABI Wheel CANN 9.1.0 / A2 v0.27.1rc1v0.27.1rc1-openeulerv0.27.1rc1-ucm-0.9.0v0.27.1rc1-openeuler-ucm-0.9.0cp312PyPI pip install "uc-manager[cann910-a2]==0.9.0"CANN 9.1.0 / A3 v0.27.1rc1-a3v0.27.1rc1-a3-openeulerv0.27.1rc1-a3-ucm-0.9.0v0.27.1rc1-a3-openeuler-ucm-0.9.0cp312PyPI pip install "uc-manager[cann910-a3]==0.9.0"CANN 9.1.0 / A5 v0.27.1rc1-a5v0.27.1rc1-a5-openeulerv0.27.1rc1-a5-ucm-0.9.0v0.27.1rc1-a5-openeuler-ucm-0.9.0cp312PyPI pip install "uc-manager[cann910-a5]==0.9.0"CANN 9.2.1 / A2 v0.30.0v0.30.0-openeulerv0.30.0-ucm-0.9.0v0.30.0-openeuler-ucm-0.9.0cp312PyPI pip install "uc-manager[cann921-a2]==0.9.0"CANN 9.2.1 / A3 v0.30.0-a3v0.30.0-a3-openeulerv0.30.0-a3-ucm-0.9.0v0.30.0-a3-openeuler-ucm-0.9.0cp312PyPI pip install "uc-manager[cann921-a3]==0.9.0"CANN 9.2.1 / A5 v0.30.0-a5v0.30.0-a5-openeulerv0.30.0-a5-ucm-0.9.0v0.30.0-a5-openeuler-ucm-0.9.0cp312PyPI pip install "uc-manager[cann921-a5]==0.9.0"SGLang
Runtime:
docker.io/lmsysorg/sglangImages: GHCR
ghcr.io/modelengine-group/sglang· Docker Hubdocker.io/unifiedcachemanager/sglang— 6 image families / 12 architecture members| Runtime capability | Upstream Runtime tagsdocker.io/lmsysorg/sglang: | Runtime tagsGHCR: ghcr.io/modelengine-group/sglang<
用户点评
登录用户的真实使用记录,可按版本和硬件筛选还没有人点评。第一个说出真实体验的人最有价值。
What is UCM?
Unified Cache Manager (UCM) is a unified KV Cache layer between inference engines and storage. It manages cache reuse, placement, and movement so that compatible requests and inference instances can reuse existing computation.
Agent workflows, multi-turn conversations, and multimodal applications often revisit the same context. UCM makes reusable KV Cache available beyond an individual inference process and extends cache capacity through external storage. This helps reduce repeated computation and the pressure on accelerator memory.
Engine adapters, cache mechanisms, and storage backends can evolve independently through pluggable interfaces.
Architecture
- Applications and inference engines produce and consume KV Cache. Reuse patterns include multimodal context, shared prefixes, and cached chunks used by mechanisms such as KV Bridge.
- Unified Cache manages the cache's usage lifecycle: what to reuse, where to retain it, and when to move it. Prefill/decode (PD) transfer is one form of cache movement. Pluggable retrieval and loading strategies also support sparse attention.
- Global KV Store provides a common storage abstraction over file systems, memory pools, and other backends. Store implementations handle retention, reclamation, and reliability according to their capabilities.
Feature availability and storage guarantees depend on the selected engine, model, platform, and backend; see supported configurations.
Getting Started
- Check the support matrix for a compatible configuration.
- Follow the Quickstart to select installation artifacts, configure storage, and run an engine with UCM.
- Use the operations guide to verify cache reuse and inspect runtime behavior.
Installation commands, image tags, engine versions, and configuration examples are maintained in the documentation. For deployment problems, see troubleshooting.
Contributing & Community
Contributions to engine adapters, storage backends, cache mechanisms, tests, and documentation are welcome.
- Start with the Contributing Guide.
- Add a backend with Extending Store.
- Report bugs and propose features through GitHub Issues.
License
UCM is licensed under the MIT license with additional conditions. Read LICENSE for the full terms.