Skip to content

v0.9.1

Latest

Choose a tag to compare

@github-actions github-actions released this 12 Nov 08:28
· 27 commits to develop since this release
c871308

UCM v0.9.1 Release Notes

UCM v0.9.1 focuses on model compatibility validation and inference stability improvements for the v0.30.0 engine release. This update adds validation for MiniMax-M3.1-Flash on NVIDIA H100 and addresses two issues reported with UCM v0.9.0.

1. Newly Validated Model

  • MiniMax-M3.1-Flash (NVIDIA H100): Added compatibility validation for MiniMax-M3.1-Flash on NVIDIA H100 GPUs.

2. Bug Fixes

Compared with UCM v0.9.0, this release includes the following fixes:

  1. GLM5.3-W8A8C8: Inference timeouts on selected datasets

    Addressed inference timeouts observed with GLM5.3-W8A8C8 on certain datasets. The associated fix restores the synchronous KV-cache load flag in the vLLM Ascend Balance Scheduler so that required external cache loads are initiated before model forward execution.

    Related PR: #1606.

  2. GLM5.3-Flash-W8A8: Inference service hang on Posix cache hits

    Fixed an issue on Atlas A3 servers where the GLM5.3-Flash-W8A8 inference service could hang when accessing cached data through Posix storage. The fix separates load tasks for shared MLA KV-cache entries and rank-private Mamba/KDA state entries, applying the appropriate backend-loading distribution policy to each task.

    Related PR: #1623.

3. Known Issues and Version Constraints

Silent output corruption with hybrid GDN models, MTP, and asynchronous scheduling

In the vLLM Ascend v0.30.0 image, hybrid GDN models using mamba_cache_mode="align" may produce silently corrupted outputs when MTP speculative decoding and asynchronous scheduling are enabled together. Under KV-cache pressure, a request may be temporarily paused during scheduling, causing intermittent repeated output or garbled text without an explicit error.

Temporary workaround: Disable asynchronous scheduling when starting the vLLM service:

--no-async-scheduling

Disabling asynchronous scheduling may reduce throughput or increase latency for some workloads.

Related issue: vllm-ascend #20677.

Status: complete

Backend Wheels: 20
Chart: unified-cache-chart-0.9.1.tgz

vLLM OpenAI

Runtime: docker.io/vllm/vllm-openai

Images: GHCR ghcr.io/modelengine-group/vllm-openai · Docker Hub docker.io/unifiedcachemanager/vllm-openai — 8 image families / 16 architecture members

Runtime capability Upstream Runtime tags
docker.io/vllm/vllm-openai:
Runtime tags
GHCR: ghcr.io/modelengine-group/vllm-openai
Docker Hub: docker.io/unifiedcachemanager/vllm-openai
Python ABI Wheel
CUDA 12.9 v0.27.1-cu129
v0.27.1-cu129-ubuntu2404
v0.30.0-cu129
v0.30.0-cu129-ubuntu2404
v0.27.1-cu129-ucm-0.9.1
v0.27.1-cu129-ubuntu2404-ucm-0.9.1
v0.30.0-cu129-ucm-0.9.1
v0.30.0-cu129-ubuntu2404-ucm-0.9.1
cp312 PyPI
pip install "uc-manager[cu129]==0.9.1"
CUDA 13.0 v0.27.1
v0.27.1-ubuntu2404
v0.30.0
v0.30.0-ubuntu2404
v0.27.1-ucm-0.9.1
v0.27.1-ubuntu2404-ucm-0.9.1
v0.30.0-ucm-0.9.1
v0.30.0-ubuntu2404-ucm-0.9.1
cp312 PyPI
pip install "uc-manager[cu130]==0.9.1"

vLLM-Ascend

Runtime: quay.io/ascend/vllm-ascend

Images: GHCR ghcr.io/modelengine-group/vllm-ascend · Docker Hub docker.io/unifiedcachemanager/vllm-ascend — 12 image families / 24 architecture members

Runtime capability Upstream Runtime tags
quay.io/ascend/vllm-ascend:
Runtime tags
GHCR: ghcr.io/modelengine-group/vllm-ascend
Docker Hub: docker.io/unifiedcachemanager/vllm-ascend
Python ABI Wheel
CANN 9.1.0 / A2 v0.27.1rc1
v0.27.1rc1-openeuler
v0.27.1rc1-ucm-0.9.1
v0.27.1rc1-openeuler-ucm-0.9.1
cp312 PyPI
pip install "uc-manager[cann910-a2]==0.9.1"
CANN 9.1.0 / A3 v0.27.1rc1-a3
v0.27.1rc1-a3-openeuler
v0.27.1rc1-a3-ucm-0.9.1
v0.27.1rc1-a3-openeuler-ucm-0.9.1
cp312 PyPI
pip install "uc-manager[cann910-a3]==0.9.1"
CANN 9.1.0 / A5 v0.27.1rc1-a5
v0.27.1rc1-a5-openeuler
v0.27.1rc1-a5-ucm-0.9.1
v0.27.1rc1-a5-openeuler-ucm-0.9.1
cp312 PyPI
pip install "uc-manager[cann910-a5]==0.9.1"
CANN 9.2.1 / A2 v0.30.0
v0.30.0-openeuler
v0.30.0-ucm-0.9.1
v0.30.0-openeuler-ucm-0.9.1
cp312 PyPI
pip install "uc-manager[cann921-a2]==0.9.1"
CANN 9.2.1 / A3 v0.30.0-a3
v0.30.0-a3-openeuler
v0.30.0-a3-ucm-0.9.1
v0.30.0-a3-openeuler-ucm-0.9.1
cp312 PyPI
pip install "uc-manager[cann921-a3]==0.9.1"
CANN 9.2.1 / A5 v0.30.0-a5
v0.30.0-a5-openeuler
v0.30.0-a5-ucm-0.9.1
v0.30.0-a5-openeuler-ucm-0.9.1
cp312 PyPI
pip install "uc-manager[cann921-a5]==0.9.1"

SGLang

Runtime: docker.io/lmsysorg/sglang

Images: GHCR ghcr.io/modelengine-group/sglang · Docker Hub docker.io/unifiedcachemanager/sglang — 6 image families / 12 architecture members

Runtime capability Upstream Runtime tags
docker.io/lmsysorg/sglang:
Runtime tags
GHCR: ghcr.io/modelengine-group/sglang
Docker Hub: docker.io/unifiedcachemanager/sglang
Python ABI Wheel
CANN 9.0.0 / A2 v0.5.20-cann9.0.0-910b v0.5.20-cann9.0.0-910b-ucm-0.9.1 cp311 PyPI
pip install "uc-manager[cann900-a2]==0.9.1"
CANN 9.0.0 / A3 v0.5.20-cann9.0.0-a3 v0.5.20-cann9.0.0-a3-ucm-0.9.1 cp311 PyPI
pip install "uc-manager[cann900-a3]==0.9.1"
CANN 9.1.0 / A3 v0.5.22-cann9.1.0-a3 v0.5.22-cann9.1.0-a3-ucm-0.9.1 cp312 PyPI
pip install "uc-manager[cann910-a3]==0.9.1"
CUDA 13.0 v0.5.20-cu130
v0.5.21-cu130
v0.5.22-cu130
v0.5.20-cu130-ucm-0.9.1
v0.5.21-cu130-ucm-0.9.1
v0.5.22-cu130-ucm-0.9.1
cp312 PyPI
pip install "uc-manager[cu130]==0.9.1"

Toolkit

ucm_toolkit-0.9.1-py3-none-any.whl
Install the optional toolkit extra with the same-version meta package, or install this Wheel independently.