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

fix(strixhalo): switch base to ROCm 7.0.2 to fix libhsa segfault
Some checks failed
Build and Push Images / build (Dockerfile.ray-worker-intel, intel) (push) Has been cancelled
Details
Build and Push Images / build (Dockerfile.ray-worker-nvidia, nvidia) (push) Has been cancelled
Details
Build and Push Images / build (Dockerfile.ray-worker-rdna2, rdna2) (push) Has been cancelled
Details
Build and Push Images / build (Dockerfile.ray-worker-strixhalo, strixhalo) (push) Has been cancelled
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 / determine-version (push) Has been cancelled
Details

Browse Source 
ROCm 7.1 system libraries (libhsa-runtime64.so.1.18.70100) are ABI-
incompatible with the torch/vLLM ROCm 7.0 wheels from wheels.vllm.ai.
This caused SIGSEGV at 0x34 in libhsa-runtime64 on every GPU operation.

Switch to rocm/pytorch:rocm7.0.2_ubuntu24.04_py3.12_pytorch_release_2.9.1
which provides matching ROCm 7.0.2 system libraries while keeping
Ubuntu 24.04 (glibc 2.38) and Python 3.12.
This commit is contained in:
Billy Davies
2026-02-09 14:37:05 -05:00
parent 3a33ed387f
commit ab2a7f486e
1 changed files with 9 additions and 9 deletions
Download Patch File Download Diff File Expand all files Collapse all files

18
dockerfiles/Dockerfile.ray-worker-strixhalo
View File

@@ -6,12 +6,13 @@
# docker build -t git.daviestechlabs.io/daviestechlabs/ray-worker-strixhalo:latest \
# -f dockerfiles/Dockerfile.ray-worker-strixhalo .
#
# Uses ROCm vendor image as base (Ubuntu 24.04 / glibc 2.38) so that all
# ROCm 7.1 shared libraries (libMIOpen, libhipblas, etc.) find a compatible
# glibc. Ray 2.53.0 is installed into the vendor venv via pip.
# Uses ROCm 7.0.2 vendor image (Ubuntu 24.04 / glibc 2.38) so system ROCm
# libraries (libhsa-runtime64, libhipblas, libMIOpen, etc.) are ABI-compatible
# with torch + vLLM wheels from wheels.vllm.ai/rocm/ (compiled for ROCm 7.0).
# ROCm 7.1 base causes segfault in libhsa-runtime64.so due to ABI mismatch.
# Note: Python 3.12 required — vLLM ROCm wheel (wheels.vllm.ai/rocm) is cp312 only
FROM docker.io/rocm/pytorch:rocm7.1_ubuntu24.04_py3.12_pytorch_release_2.9.1
FROM docker.io/rocm/pytorch:rocm7.0.2_ubuntu24.04_py3.12_pytorch_release_2.9.1
# OCI Image Spec labels
LABEL org.opencontainers.image.title="Ray Worker - AMD Strix Halo"
@@ -19,7 +20,7 @@ LABEL org.opencontainers.image.description="Ray Serve worker for AMD Strix Halo
LABEL org.opencontainers.image.vendor="DaviesTechLabs"
LABEL org.opencontainers.image.source="https://git.daviestechlabs.io/daviestechlabs/kuberay-images"
LABEL org.opencontainers.image.licenses="MIT"
LABEL gpu.target="amd-rocm-7.1-gfx1151"
LABEL gpu.target="amd-rocm-7.0-gfx1151"
LABEL ray.version="2.53.0"
WORKDIR /app
@@ -63,10 +64,9 @@ RUN --mount=type=cache,target=/var/cache/apt,sharing=locked \
# Install uv for fast Python package management (ADR-0014)
COPY --from=ghcr.io/astral-sh/uv:latest /uv /usr/local/bin/uv
# Remove vendor torch — the vendor ships torch 2.9.1+rocm7.1 but vLLM
# ROCm wheel (0.15.1+rocm700) was compiled against the PyTorch ROCm 7.0
# ABI. Installing from the ROCm 7.0 index avoids undefined-symbol errors
# in vllm._C / vllm._rocm_C (e.g. silu_and_mul).
# Remove vendor torch — replace with the exact torch from wheels.vllm.ai/rocm/
# (2.9.1+git8907517) that vLLM 0.15.1+rocm700 was compiled against.
# The vendor torch is close but may differ in C++ ABI details.
RUN uv pip uninstall --python /opt/venv/bin/python3 \
torch torchaudio torchvision 2>/dev/null || true