#FROM vllm/vllm-openai:v0.19.0-cu130
#vllm says version 0.20.2rc1.dev9+g01d4d1ad3
FROM  vllm/vllm-openai:glm52
#FROM vllm/vllm-openai:nightly
#FROM vllm/vllm-openai:glm51-cu130

# Install LMCache for KV cache offloading / sharing across nodes
# Build with system CUDA 13.0 for Blackwell (B200)
RUN apt-get update && apt-get install -y git \
    libcusolver-dev-13-0 \
    libcusparse-dev-13-0 \
    libcublas-dev-13-0 \
    libcurand-dev-13-0 \
    libcufft-dev-13-0 \
    libnvjitlink-dev-13-0 && \
    git clone https://github.com/biondizzle/LMCache.git /tmp/lmcache && \
    cd /tmp/lmcache && \
    git checkout dream-build && \
    CUDA_HOME=/usr/local/cuda \
    TORCH_CUDA_ARCH_LIST="10.0" \
    pip install --no-cache-dir --no-build-isolation . && \
    rm -rf /tmp/lmcache && export CACHE_BUSTER=4 && \
    pip uninstall -y cupy-cuda12x || true


# Make sure we have patch to make MTP work on GLM (I did this based on vllm-project/vllm#40989)
#COPY indexer.py /usr/local/lib/python3.12/dist-packages/vllm/v1/attention/backends/mla/indexer.py

# These were from https://github.com/vllm-project/vllm/pull/41357/changes#diff-75b8ca6d854db6a47e75db6507afd20c15624f35229e2fd0d71642bffd70b11c
#COPY shm_broadcast.py /usr/local/lib/python3.12/dist-packages/vllm/distributed/device_communicators/shm_broadcast.py
#COPY multiproc_executor.py /usr/local/lib/python3.12/dist-packages/vllm/v1/executor/multiproc_executor.py

# Make sure we have the latest up to date chat template
#COPY glm_5.1_chat_template.jinja /opt/chat_template.jinja

# GLM 5.1 LMCache config
COPY lmcache-config-glm-51.yaml /opt/lmcache-config-glm-51.yaml

# DEEPSEEK v4 LMCache config
#COPY lmcache-config-dsv4.yaml /opt/lmcache-config-dsv4.yaml