# Lightweight ROCm image: latest LMCache release on top of vLLM-ROCm.
# No NIXL, no Prefill-Decode Disaggregation.
# For AMD Instinct GPUs (MI300X/MI325X, MI350X/MI355X).
#
# Note: lmcache is built from source with HIP extensions using
# --no-build-isolation to link against the ROCm torch in the base image.
FROM vllm/vllm-openai-rocm:latest

# ROCm GPU architectures — required during build (no GPU present)
ARG PYTORCH_ROCM_ARCH="gfx942,gfx950"
ENV PYTORCH_ROCM_ARCH=${PYTORCH_ROCM_ARCH}

# Build LMCache from source with HIP extensions.
# PyPI lmcache>=0.4 ships only binary wheels (no sdist), so `--no-binary`
# cannot force a source build.  Cloning the repo ensures HIP extensions
# are compiled against the ROCm torch already present in the base image.
# Build deps are installed from the cloned source to stay version-consistent.
ARG max_jobs=2
ENV MAX_JOBS=${max_jobs}

ARG LMCACHE_VERSION=dev
RUN git clone --depth 1 -b ${LMCACHE_VERSION} https://github.com/LMCache/LMCache.git /tmp/lmcache && \
    cd /tmp/lmcache && \
    uv pip install --system --no-cache -r requirements/build.txt && \
    BUILD_WITH_HIP=1 CXX=hipcc \
    uv pip install --system --no-cache --no-build-isolation . && \
    rm -rf /tmp/lmcache


# Build (no repo context needed — everything is cloned from git):
# docker build -f docker/Dockerfile.rocm-lightweight -t 'lmcache/vllm-openai-rocm:light' docker/

# Run:
# export HF_TOKEN=<your_huggingface_token>
# docker run --device /dev/kfd --device /dev/dri --group-add video \
#    -v ~/.cache/huggingface:/root/.cache/huggingface \
#    --env "HF_TOKEN=$HF_TOKEN" \
#    -p 8000:8000 \
#    --ipc=host \
#    lmcache/vllm-openai-rocm:light \
#    --model Qwen/Qwen3-0.6B \
#    --kv-transfer-config \
#    '{"kv_connector":"LMCacheConnectorV1","kv_role":"kv_both"}'