Billy D.
3c788fe2b6
Some checks failed
Build and Push Images / build-strixhalo (push) Has been cancelled
Build and Push Images / build-nvidia (push) Has been cancelled
Build and Push Images / build-intel (push) Has been cancelled
Build and Push Images / Release (push) Has been cancelled
Build and Push Images / Notify (push) Has been cancelled
Build and Push Images / build-rdna2 (push) Has been cancelled
Ray base image has pandas 1.5.3 compiled against numpy 1.x, but TheRock PyTorch ROCm wheels require numpy 2.x. This causes: ValueError: numpy.dtype size changed, may indicate binary incompatibility Fix by installing pandas 2.x which is compatible with numpy 2.x.
105 lines
4.0 KiB
Docker
105 lines
4.0 KiB
Docker
# syntax=docker/dockerfile:1.7
|
|
# AMD Strix Halo Ray Worker for khelben (gfx1151 / RDNA 3.5)
|
|
# Used for: vLLM (Llama 3.1 70B)
|
|
#
|
|
# Build:
|
|
# docker build -t git.daviestechlabs.io/daviestechlabs/ray-worker-strixhalo:latest \
|
|
# -f dockerfiles/Dockerfile.ray-worker-strixhalo .
|
|
#
|
|
# Multi-stage build: Extract ROCm 7.1 from vendor image, use Ray base for Python 3.11
|
|
# Note: Uses TheRock gfx110X wheels due to ROCm/ROCm#5853 segfault issue
|
|
|
|
# Stage 1: ROCm 7.1 libraries from AMD vendor image
|
|
FROM docker.io/rocm/pytorch:rocm7.1_ubuntu24.04_py3.12_pytorch_release_2.9.1 AS rocm-source
|
|
|
|
# Stage 2: Production image
|
|
FROM rayproject/ray:2.53.0-py311 AS production
|
|
|
|
# OCI Image Spec labels
|
|
LABEL org.opencontainers.image.title="Ray Worker - AMD Strix Halo"
|
|
LABEL org.opencontainers.image.description="Ray Serve worker for AMD Strix Halo (vLLM LLM inference)"
|
|
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 ray.version="2.53.0"
|
|
|
|
WORKDIR /app
|
|
|
|
# Copy ROCm stack from vendor image
|
|
COPY --from=rocm-source /opt/rocm /opt/rocm
|
|
|
|
# ROCm environment variables
|
|
ENV ROCM_HOME=/opt/rocm \
|
|
PATH="${ROCM_HOME}/bin:${ROCM_HOME}/llvm/bin:${PATH}" \
|
|
LD_LIBRARY_PATH="${ROCM_HOME}/lib:${ROCM_HOME}/lib64:${LD_LIBRARY_PATH}" \
|
|
HSA_PATH="${ROCM_HOME}/hsa" \
|
|
HIP_PATH="${ROCM_HOME}/hip" \
|
|
# Strix Halo (gfx1151) specific settings
|
|
HIP_VISIBLE_DEVICES=0 \
|
|
HSA_ENABLE_SDMA=0 \
|
|
PYTORCH_HIP_ALLOC_CONF="expandable_segments:True,max_split_size_mb:512" \
|
|
HSA_OVERRIDE_GFX_VERSION="11.0.0" \
|
|
ROCM_TARGET_LST="gfx1151,gfx1100"
|
|
|
|
# Install system dependencies
|
|
USER root
|
|
RUN --mount=type=cache,target=/var/cache/apt,sharing=locked \
|
|
--mount=type=cache,target=/var/lib/apt,sharing=locked \
|
|
apt-get update && apt-get install -y --no-install-recommends \
|
|
libelf1 \
|
|
libnuma1 \
|
|
libdrm2 \
|
|
libdrm-amdgpu1 \
|
|
kmod \
|
|
&& rm -rf /var/lib/apt/lists/*
|
|
|
|
# Install uv for fast Python package management (ADR-0014)
|
|
COPY --from=ghcr.io/astral-sh/uv:latest /uv /usr/local/bin/uv
|
|
|
|
USER ray
|
|
|
|
# WORKAROUND: ROCm/ROCm#5853 - Standard PyTorch ROCm wheels cause segfault
|
|
# in libhsa-runtime64.so during VRAM allocation on gfx1151 (Strix Halo).
|
|
# TheRock gfx110X-all packages provide compatible Python 3.11 wheels.
|
|
RUN --mount=type=cache,target=/home/ray/.cache/uv,uid=1000,gid=1000 \
|
|
uv pip install --system \
|
|
--index-url https://rocm.nightlies.amd.com/v2/gfx110X-all/ \
|
|
torch torchaudio torchvision
|
|
|
|
# Install vLLM and inference dependencies (uv is 10-100x faster than pip)
|
|
RUN --mount=type=cache,target=/home/ray/.cache/uv,uid=1000,gid=1000 \
|
|
uv pip install --system \
|
|
'vllm>=0.5.0' \
|
|
'transformers>=4.35.0,<5.0' \
|
|
'accelerate>=0.25.0,<1.0' \
|
|
'sentence-transformers>=2.3.0,<3.0' \
|
|
'httpx>=0.27.0,<1.0' \
|
|
'scipy>=1.11.0,<2.0'
|
|
|
|
# FIX: Ray base image has pandas 1.5.3 which is incompatible with numpy 2.x
|
|
# The TheRock PyTorch wheels require numpy 2.x, so upgrade pandas to match
|
|
RUN --mount=type=cache,target=/home/ray/.cache/uv,uid=1000,gid=1000 \
|
|
uv pip install --system 'pandas>=2.0.0,<3.0'
|
|
|
|
# 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')"
|
|
|
|
# Copy application code
|
|
COPY --chown=ray:ray ray-serve/ /app/ray_serve/
|
|
COPY --chown=ray:ray --chmod=755 dockerfiles/ray-entrypoint.sh /app/ray-entrypoint.sh
|
|
|
|
# Environment configuration
|
|
ENV PYTHONPATH=/app \
|
|
PYTHONUNBUFFERED=1 \
|
|
PYTHONDONTWRITEBYTECODE=1 \
|
|
RAY_HEAD_SVC="ai-inference-raycluster-head-svc" \
|
|
GPU_RESOURCE="gpu_amd_strixhalo" \
|
|
NUM_GPUS="1"
|
|
|
|
# Health check
|
|
HEALTHCHECK --interval=30s --timeout=10s --start-period=120s --retries=3 \
|
|
CMD ray status --address=localhost:6379 || exit 1
|
|
|
|
ENTRYPOINT ["/app/ray-entrypoint.sh"]
|