github-featured / lmcache--lmcache

发布

0 发布 86 Git 标签

RSS 订阅源
  • v0.5.3.rc3 a7afadebb9
    比较

    [CI] fix abi mismatch and smoke checks (#4423)

    frostbyte_neo 发布于 2026-08-04 22:32:03 +00:00 | -57 次提交 在此版本后已推送到 automation/sync-torch-version

    • [Build] pin torch to vLLM stable, not vLLM main

    The weekly sync_torch_version bot reads the torch pin from
    vllm-project/vllm@main. The published lmcache wheel is installed next to
    stable vLLM (docker/Dockerfile image-release, and plain
    pip install vllm lmcache), so whenever main runs ahead of the latest
    release the wheel is built against a torch that is not the one installed
    at runtime.

    That is what shipped in v0.5.3rc2: the bot bumped the pin to 2.13.0
    (#4255) while stable vLLM 0.26.0 still pins torch 2.11.0, so
    lmcache.c_ops in lmcache/vllm-openai:v0.5.3rc2 fails to load with

    undefined symbol: _ZN3c1019NotImplementedErrorC1ENS_14SourceLocation...
(c10::NotImplementedError::NotImplementedError)

    CudaDeviceOps.ensure_native() swallows the ImportError, so the image
    serves normally with every CUDA kernel silently disabled.

    Point the sync at the latest stable vLLM release (/releases/latest
    excludes drafts and pre-releases) and revert the pin to 2.11.0 to match
    vLLM 0.26.0. Nightly and dev images are unaffected either way: they
    build from source against whatever torch is already installed.

    Signed-off-by: deng451e 838677410@qq.com

    • [CI] make the c_ops smoke check able to fail

    python -c 'import lmcache.c_ops' was the smoke check in four workflows,
    but it cannot detect a broken extension. lmcache/init registers a
    backward-compat shim under that name, and CudaDeviceOps.ensure_native()
    swallows the ImportError and falls back to the torch baseline. Either one
    alone is enough to make the import succeed. Verified against
    lmcache/vllm-openai:v0.5.3rc2, whose c_ops does not load:

    $ python -c 'import lmcache.c_ops'; echo $?
0

    Replace it with .github/scripts/assert_native_ops.py, which loads the
    extension file directly so dlopen and symbol resolution actually run. On
    the same image it reports the real failure and exits 1:

    lmcache.c_ops is installed but failed to load against torch
2.11.0+cu130 ... undefined symbol: _ZN3c1019NotImplementedError...

    The script imports no lmcache internals, so it also works as a one-shot
    check against any released wheel or container image.

    Signed-off-by: deng451e 838677410@qq.com

    下载附件