Revert to original implementation for testing

Author: XkunW

Dockerfile

@@ -55,11 +55,15 @@ COPY . /vec-inf
 RUN PIP_INDEX_URL="https://download.pytorch.org/whl/cu128" uv pip install --system -e .[dev]
 
 # Install a single, system NCCL (from NVIDIA CUDA repo in base image)
-RUN apt-get update && apt-get install -y --allow-change-held-packages\
-    libnccl2 libnccl-dev \
-    && rm -rf /var/lib/apt/lists/*
+# RUN apt-get update && apt-get install -y --allow-change-held-packages\
+#     libnccl2 libnccl-dev \
+#     && rm -rf /var/lib/apt/lists/*
 
 # Final configuration
+RUN mkdir -p /vec-inf/nccl && \
+     mv /root/.config/vllm/nccl/cu12/libnccl.so.2.18.1 /vec-inf/nccl/libnccl.so.2.18.1
+
+ENV VLLM_NCCL_SO_PATH=/vec-inf/nccl/libnccl.so.2.18.1
 ENV NCCL_DEBUG=INFO
 
 # Set the default command to start an interactive shell

pyproject.toml

@@ -42,6 +42,7 @@ dev = [
     "xgrammar>=0.1.11",
     "torch>=2.7.0",
     "vllm>=0.10.0",
+    "vllm-nccl-cu12>=2.18,<2.19",
     "ray>=2.40.0",
     "cupy-cuda12x==12.1.0"
 ]