27 lines
1.1 KiB
Docker
27 lines
1.1 KiB
Docker
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
|
|
# 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 |