daviestechlabs/kuberay-images
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

feat(strixhalo): add amdsmi sysfs shim to bypass glibc 2.38 requirement
Some checks failed
Build and Push Images / determine-version (push) Successful in 58s
Details
Build and Push Images / Release (push) Has been cancelled
Details
Build and Push Images / Notify (push) Has been cancelled
Details
Build and Push Images / build-strixhalo (push) Has been cancelled
Details
Build and Push Images / build-intel (push) Has been cancelled
Details
Build and Push Images / build-nvidia (push) Has been cancelled
Details
Build and Push Images / build-rdna2 (push) Has been cancelled
Details

Browse Source 
The native amdsmi from ROCm 7.1 requires libamd_smi.so linked against
glibc 2.38 (Ubuntu 24.04), but the Ray base image is Ubuntu 22.04
(glibc 2.35). This caused vLLM to fail ROCm platform detection with
'No module named amdsmi' / GLIBC_2.38 not found errors.

Solution: Pure-Python amdsmi shim that reads GPU info from sysfs
(/sys/class/drm/*) instead of the native library. Provides the full
API surface used by both vLLM (platform detection, device info) and
PyTorch (device counting, memory/power/temp monitoring).

Tested in-container: vLLM detects RocmPlatform, PyTorch sees GPU
(Radeon 8060S, 128GB, HIP 7.3), DeviceConfig resolves to 'cuda'.

Changes:
- Add amdsmi-shim/ package with sysfs-backed implementation
- Update Dockerfile to install shim after vLLM/torch
- Add amdsmi-shim/ to .dockerignore explicit includes
This commit is contained in:
Billy Davies
2026-02-06 08:28:07 -05:00
parent 5f1873908f
commit 300582a520
4 changed files with 359 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files

11
dockerfiles/Dockerfile.ray-worker-strixhalo
View File

@@ -32,7 +32,7 @@ COPY --link --from=rocm-source /opt/rocm /opt/rocm
# ROCm environment variables - split to ensure ROCM_HOME is set first
ENV ROCM_HOME=/opt/rocm
ENV PATH="/opt/rocm/bin:/opt/rocm/llvm/bin:/home/ray/anaconda3/bin:/home/ray/.local/bin:/usr/local/sbin:/usr/local/bin:/usr/sbin:/usr/bin:/sbin:/bin" \
LD_LIBRARY_PATH="/opt/rocm/lib:/opt/rocm/lib64" \
LD_LIBRARY_PATH="/opt/rocm/lib:/opt/rocm/lib64:/home/ray/anaconda3/lib" \
HSA_PATH="/opt/rocm/hsa" \
HIP_PATH="/opt/rocm/hip" \
# Strix Halo (gfx1151) specific settings
@@ -84,6 +84,15 @@ RUN --mount=type=cache,target=/home/ray/.cache/uv,uid=1000,gid=1000 \
RUN --mount=type=cache,target=/home/ray/.cache/uv,uid=1000,gid=1000 \
uv pip install --system 'pandas>=2.0.0,<3.0'
# Install amdsmi sysfs shim LAST (required for vLLM ROCm platform detection).
# The native amdsmi from ROCm 7.1 requires glibc 2.38 (Ubuntu 24.04),
# but the Ray base image is Ubuntu 22.04 (glibc 2.35). This pure-Python
# shim reads GPU info from /sys/class/drm/* instead of libamd_smi.so.
# Must be installed after vLLM/torch to prevent PyPI amdsmi from overwriting it.
COPY --chown=1000:100 amdsmi-shim /tmp/amdsmi-shim
RUN --mount=type=cache,target=/home/ray/.cache/uv,uid=1000,gid=1000 \
uv pip install --system /tmp/amdsmi-shim && rm -rf /tmp/amdsmi-shim
# Pre-download common models for faster cold starts (optional, increases image size)
# RUN python3 -c "from sentence_transformers import SentenceTransformer; SentenceTransformer('BAAI/bge-large-en-v1.5')"