
lmcache, and the HTTP services check an API key once you configure one. The ZMQ and gRPC transports still have no authentication, so keep them on a private network or the local host.The LMCache contributor team and community are pleased and relieved to announce LMCache v0.5.7.post1, released on November 12, 2026. This release has the same features as v0.5.7, with packages built against PyTorch 2.14.0. It includes safer defaults for deployments, optional API key authentication, support for more accelerators, and new multiprocess (MP) transport modes.
If you run the official Docker images, read the upgrade notes before you upgrade. The images no longer run as root.
Safer defaults for deployments
The official Docker images now run as a non-root user by default (#5685). The user is lmcache, with user ID (UID) 1000.
Important: the official images now run as a non-root user. Check writable paths and mounted-volume permissions before you upgrade.
With this change, LMCache in a container has only the access that UID 1000 has. Build changes in #6056 and #6132 make the image build work for a non-root user.
The Hugging Face cache inside the container moves to /home/lmcache/.cache/huggingface. If you mount a host directory there, UID 1000 must be able to read and write it.
Optional API key authentication
This release adds API key authentication to the MP HTTP server (#5756) and to the frontend HTTP API (#5767). Both use one shared verifier and one set of configuration options (#5779).
Authentication is off until you configure a key. To turn it on, give the server a key file:
lmcache server --api-key-file /run/secrets/lmcache-api-key
You can also set LMCACHE_API_KEY_FILE or LMCACHE_API_KEY. All key sources add to one set of accepted keys. This lets you rotate keys: add the new key, move clients to it, and then remove the old key.
Clients send the key in the Authorization: Bearer <key> header, or in the X-LMCache-API-Key header when a proxy uses Authorization for itself. The / and /healthcheck routes stay public, so liveness and readiness probes still work.
The coordinator uses the same --api-key-file option, and it sends the configured key on its own outbound requests (#6174). In the words of the change: “An unset flag leaves the coordinator open; supplying a key closes every route except /healthz.”
Coverage and non-coverage
API keys check who is calling. They do not encrypt traffic. If requests cross a network you do not trust, put Transport Layer Security (TLS) in front of LMCache at a trusted reverse proxy.
This release adds authentication to HTTP services. It does not add authentication to the MP request transports, ZeroMQ (ZMQ) and gRPC. Keep the transport port on a private network or on the local host. API key authentication for the ZMQ and gRPC transports is in v0.5.8rc1 (see “Coming in v0.5.8” below).
Upgrade notes for operators
Check these before you upgrade:
- Your container can write to every path LMCache uses as UID 1000.
- Host directories that you mount, such as a shared Hugging Face cache, are readable and writable by UID 1000.
- If you depend on a cache that was filled under
/root, either move it or run the container as root on purpose (docker run --user 0, orsecurityContext.runAsUser: 0on Kubernetes). - If you expose the MP HTTP server, the frontend, or the coordinator beyond the local host, decide whether to turn on API key authentication.
- If you used
--http-api-key-fileorLMCACHE_HTTP_API_KEY, these still work. The new--api-key-fileandLMCACHE_API_KEYnames apply to all HTTP services.
Pickle removed from the tensor codec
Engine-driven transfers in MP mode no longer use Python pickle for tensor data (#5699). A typed codec takes its place, and it rejects unknown fields. New pickle imports are now blocked outside legacy code paths (#5903).
Bounded authentication replies
LMCache now limits the length of an authentication reply in the Redis serialization protocol (RESP) (#6108).
Kubernetes example bound to localhost
The MP example Kubernetes deployment now binds the LMCache transport port to 127.0.0.1 (#5684). vLLM pods connect over the node’s loopback interface, so the transport port is not exposed beyond the node.
Hardware and software in this release
Hardware
| Platform | What is new or improved | PRs |
|---|---|---|
| Biren SUPA | New platform support | #5420 |
| AWS Trainium (Neuron) | KV caching in MP mode | #5359 |
| Cambricon MLU | New platform support | #6023 |
| Ascend | Mamba model support; KV groups split by physical block stride | #5702, #6127 |
| Rebellions neural processing unit (NPU, RBLN) | Blocks-first KV layout in vLLM-RBLN; wheels built and published on GitHub Releases | #5748, #6140 |
| MUSA | Unified KV format | #5573 |
| Intel XPU | Release wheels aligned with vLLM v0.32.0 | #6151 |
| AMD ROCm | Wheels built against PyTorch 2.13; fixes for registered-L1 compaction and hipHostRegister faults | #6080, #5710, #5534 |
| Device-DAX and CXL memory | L1 offsets on Device-DAX and Compute Express Link (CXL) memory; range-based DAX L2 ownership | #5874, #5878 |
Engines, storage, and platforms
| Platform | What is new or improved | PRs |
|---|---|---|
| SGLang | Kimi Linear support; direct-copy memory kernels | #5892, #6062 |
| DeepSeek V4 | Compressed KV groups in engine-driven transfers | #6124 |
| TensorRT-LLM | Respects the kernel-object batch limit | #5762 |
| NIXL | Pool capacity reported in allocation-slot units | #5800 |
| HF3FS | Atomic object writes through temporary files | #5808 |
| Amazon S3 | Wrongly sized L2 GET results treated as misses | #6015 |
| Bigtable | Rejected L2 stores reported as failures | #5471 |
| Mooncake | SDK loaded only when its backend is selected; request tags in lookup keys | #6134, #5449 |
| gRPC (MP transport) | Native and grpc.aio server modes, automatic server start for grpc:// endpoints, native-mode metrics, and grpc+unix | #6031, #5799, #6001, #6061, #6154 |
| Python | Python 3.14 support | #5289 |
Builds and images
Install or upgrade from the Python Package Index (PyPI):
pip install --upgrade lmcache==0.5.7.post1
The PyPI packages are built against PyTorch 2.14.0, to match the vLLM runtime (#6465). If you installed v0.5.7 from an earlier build, upgrade to v0.5.7.post1.
These images are on Docker Hub, for both amd64 and arm64:
docker pull lmcache/vllm-openai:v0.5.7.post1docker pull lmcache/vllm-openai:v0.5.7.post1-lightweightdocker pull lmcache/standalone:v0.5.7.post1docker pull lmcache/lmcache-payload:v0.5.7.post1
This release no longer provides CUDA 12.9 (cu129) wheels or images (#6533). It also does not include an SGLang image, because SGLang compatibility with PyTorch 2.14 is not yet validated.
RBLN wheels are published on their own GitHub release page (#6140).
Other changes
- A shared-L1 memory orchestrator (#5559) and a per-request policy to load KV cache or recompute it (#5536).
- Coordinator visibility: per-server cache-event loss, view-freshness tracking, persistence and prefetch metrics, a
/statussnapshot, and a Grafana dashboard (#5560, #5681, #5881, #5879, #6179). - Groundwork for token dropping: request-local state, request-scoped algorithm plugins, and paged GPU KV compaction (#5870, #5736, #5737).
- Dynamic prefetch admission (#5889), and no L2 lookup when L1 already holds the full prefetch request (#5652).
- Retryable peer-to-peer (P2P) adapter teardown (#6208).
- In-process components moved under legacy namespaces, and in-process guides moved to the legacy docs (#5923).
The v0.5.7 release notes list every change, and the v0.5.7.post1 release notes describe the rebuild.
Coming in v0.5.8
These changes are in v0.5.8rc1, the first release candidate for v0.5.8, and they are planned for v0.5.8. They are not in v0.5.7.post1. You can test v0.5.8rc1 from PyPI with pip install --pre lmcache.
Security:
- API key authentication for the ZMQ request transport (#5769), and for the sync, aio, and native gRPC servers (#5770, #6439)
Features and integrations:
- vLLM sleep mode with VMM-shared KV caches (#6358)
- The
valkey_rdmaL2 adapter (#5556) - Request matching tags through object identity (#6199, #5843)
- File descriptor (fd) based sharing of VMM-mapped memory regions (#6266)
- TP-replicated KV groups stored once per tensor parallel (TP) group (#6125)
- HF3FS
shard_dirsto spread stores across directories (#6261) --isolated-ipcforlmcache bench server(#5931)- Token dropping that allocates from the resident KV length (#5871)
- A GLM-5.3-Flash recipe (#6372), and Kimi K3.1 KDA support with correctness checks (#6318)
Refactoring:
- The in-tree in-process implementation is removed (#6254)
Bug fixes:
- CLI profiling tests isolated from host dependencies (#6350)
- Padded per-layer (K, V) KV tuples on Ascend (#6241)
- The Intel compiler runtime is no longer bundled into the XPU wheel (#6281)
- The engine stays alive when gRPC or ZMQ requests fail (#6275, #6331), and RESP settings fall back to environment variables (#6294)
Thank you, new contributors
This release includes work from 34 new contributors. Thank you, and welcome to the LMCache community.