Build Triton

Instructions to build a minimal Triton container for CMS.

Server build

1. Checkout:

   git clone git@github.com:fastmachinelearning/server -b buildpy_revamp_main

2. Build:

   ./build.py \
   --target-platform linux -j 24 --no-container-interactive \
   --version 2.68.0 --container-version r26.04 --use-buildbase \
   --enable-backend ensemble python pytorch onnxruntime tensorflow \
   --image pytorch nvcr.io/nvidia/pytorch:26.04-py3 \
   --backend-tag tensorflow r25.06 \
   --extra-backend-cmake-arg tensorflow TRITON_TENSORFLOW_DOCKER_IMAGE "nvcr.io/nvidia/tensorflow:25.02-tf2-py3" \
   --override-backend-cmake-arg onnxruntime TRITON_ENABLE_ONNXRUNTIME_OPENVINO OFF \
   --enable-endpoint grpc http \
   --enable-repoagent checksum \
   --enable-feature logging stats metrics gpu_metrics cpu_metrics tracing nvtx gpu \
   -v &> log_build.log &

3. Tag for later use:

   docker tag tritonserver:latest fastml/triton-cms:26.04-py3

PyTorch Geometric libraries

1. Add PyTorch Geometric libraries (based on triton-torchgeo-gat-example):

   docker build -t fastml/triton-torchgeo:26.04-py3-geometric -f Dockerfile.torchgeo -m 16g . &> log_build_geo.log &

2. Push to DockerHub:

   docker push fastml/triton-torchgeo:26.04-py3-geometric

This automatically triggers the Apptainer conversion and cvmfs synchronization via unpacked.