vllm-with-lmcache/Dockerfile

FROM vllm/vllm-openai:v0.19.0-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/neuralwatt/LMCache.git /tmp/lmcache && \
    cd /tmp/lmcache && \
    git checkout fix/mla-multi-group-kv-cache && \
    CUDA_HOME=/usr/local/cuda \
    TORCH_CUDA_ARCH_LIST="10.0" \
    pip install --no-cache-dir --no-build-isolation . && \
    rm -rf /tmp/lmcache

COPY ./super_v3_reasoning_parser.py /opt/super_v3_reasoning_parser.py

# Monkey patch more vllm stuff - https://github.com/vllm-project/vllm/pull/38237/changes#diff-bee6813076031d3ca1edc903c1b02b81e4676519afc562ce3fefe37f20c7b650
RUN sed -i "s/if self\.kv_events_config is not None:/if self.kv_events_config is not None and self.kv_events_config.enable_kv_cache_events:/" /usr/local/lib/python3.12/dist-packages/vllm/config/vllm.py

# Patch LMCacheConnectorV1 to support HMA (Hybrid Mamba/Attention KV cache manager)
# This is required for hybrid models like Nemotron that use both Mamba and Attention layers.
# Without this patch, LMCacheConnectorV1 fails with:
#   "Connector LMCacheConnectorV1 does not support HMA but HMA is enabled"
COPY ./lmcache_connector.py /usr/local/lib/python3.12/dist-packages/vllm/distributed/kv_transfer/kv_connector/v1/lmcache_connector.py