memcache-hybrid

What it is and what it does

memcache-hybrid is a Python wrapper around a distributed key-value cache engine built for LLM and graph reasoning inference acceleration. It exposes object-oriented APIs for batch and non-batch put/get/exist/remove operations on KV blocks, backed by a multi-tier memory pool spanning HBM, DDR, and SSD with automatic data migration between tiers. The package depends on memfabric_hybrid for low-latency, high-bandwidth cross-device and cross-machine data access using hardware-specific paths (RDMA, SDMA, shared memory) on Ascend and Kunpeng processors.

The system is architected around two core components: MetaService, which manages cluster-wide memory allocation and node membership, and LocalService, which provides memory regions and acts as the client-side library. MetaService supports both single-node and Kubernetes-based HA deployments. The package is extremely new—first released in February 2026—and targets inference workloads where KV cache pooling and efficient multi-node memory access are critical bottlenecks.

Use it for:

• Accelerate LLM inference by pooling KV cache across multiple nodes with direct cross-device memory access.
• Implement vLLM-Ascend backend KV pool storage for large-language-model serving.
• Enable prefix-cache acceleration for inference workloads on Ascend hardware.
• Build distributed graph reasoning inference systems requiring high-bandwidth, low-latency KV storage.
• Manage multi-tier memory (HBM/DDR/SSD) for inference workloads with automatic hot/cold data migration.

Worth the install?

AI-flagged interpretation of the facts on this page — verify before relying

Provides Python bindings for a high-performance distributed key-value cache engine designed for LLM and graph reasoning inference workloads, with multi-tier memory pooling and cross-device data access.

Yes, with conditions. Install if you are deploying LLM or graph reasoning inference on Ascend or Kunpeng hardware and need distributed KV cache pooling with multi-tier memory support. Do not install if: (1) your organization cannot adopt Mulan PSL v2 licensing, (2) you require production-grade stability (the package is 6 months old with no long-term track record), (3) you lack the infrastructure to deploy and manage MetaService and LocalService components, or (4) you are not targeting Ascend/Kunpeng hardware where the performance benefits are realized.

Install

memcache-hybrid on PyPI

pip install memcache-hybrid

uv add memcache-hybrid

poetry add memcache-hybrid

Installing memcache-hybrid

Before you install

Medium install friction due to compiled wheels for specific Python versions and Linux architectures. Package is very new (first release February 2026, latest August 2026) with active maintenance status, but the nascent release history and reliance on memfabric_hybrid as a runtime dependency warrant caution in production environments.

License in practice

Licensed under Mulan PSL v2, which is a permissive open-source license. License treatment is marked unclear in the fact sheet, so review the actual LICENSE file and your organization's policy before adopting in proprietary or commercial contexts.

Quickstart

pip install memcache-hybrid
import memcache_hybrid
# Requires MetaService and LocalService components running; see docs for deployment

Requires MemFabric (memfabric_hybrid) as a runtime dependency; MetaService and LocalService processes must be deployed and running before client code can connect; Linux x86_64 or aarch64 with manylinux_2_26+ or manylinux_2_27+ required.

Verify before relying

• Whether Mulan PSL v2 is compatible with your organization's licensing requirements (marked unclear in metadata).
• Performance characteristics and scalability limits for your specific workload size and cluster topology.
• Stability and breaking-change policy given the package's very recent release (first release Feb 2026).
• Whether memfabric_hybrid is available and compatible in your target deployment environment.

Package facts

Evidence: memcache_hybrid-1.2.0-cp310-cp310-manylinux_2_26_aarch64.manylinux_2_28_aarch64.whl; memcache_hybrid-1.2.0-cp310-cp310-manylinux_2_27_x86_64.manylinux_2_28_x86_64.whl; memcache_hybrid-1.2.0-cp311-cp311-manylinux_2_26_aarch64.manylinux_2_28_aarch64.whl; memcache_hybrid-1.2.0-cp311-cp311-manylinux_2_27_x86_64.manylinux_2_28_x86_64.whl; memcache_hybrid-1.2.0-cp312-cp312-manylinux_2_26_aarch64.manylinux_2_28_aarch64.whl; memcache_hybrid-1.2.0-cp312-cp312-manylinux_2_27_x86_64.manylinux_2_28_x86_64.whl; memcache_hybrid-1.2.0-cp313-cp313-manylinux_2_26_aarch64.manylinux_2_28_aarch64.whl; memcache_hybrid-1.2.0-cp313-cp313-manylinux_2_27_x86_64.manylinux_2_28_x86_64.whl

Tags

More Distributed Computing packages

Further reading