vllm-glm/Dockerfile

ARG BASE_IMAGE=vllm/vllm-openai:glm51-cu130
FROM ${BASE_IMAGE}

# Patch tool parser for GLM regex fix
COPY glm4_moe_tool_parser.py /usr/local/lib/python3.12/dist-packages/vllm/tool_parsers/glm4_moe_tool_parser.py
COPY utils.py /usr/local/lib/python3.12/dist-packages/vllm/tool_parsers/utils.py

# Patch hf renderer to force string content format for GLM models
# This fixes the issue where tool response content is dropped
COPY vllm_patches/hf.py /usr/local/lib/python3.12/dist-packages/vllm/renderers/hf.py

# Install LMCache for KV cache offloading / sharing across nodes
# Reinstall torch with CUDA 13 support, then build LMCache with CUDA extensions
RUN apt-get update && apt-get install -y git && \
    # Reinstall torch with CUDA 13.x support to match base image CUDA version
    pip install --no-cache-dir --upgrade torch --index-url https://download.pytorch.org/whl/cu131 && \
    # Clone and build LMCache from source with CUDA extensions
    git clone --depth 1 https://github.com/LMCache/LMCache.git /tmp/lmcache && \
    cd /tmp/lmcache && \
    pip install --no-cache-dir . && \
    rm -rf /tmp/lmcache