TomoCloud Pipeline Container

GPU-accelerated tomographic reconstruction and AI segmentation container for DOE facility pipelines.

Stack

• tomocupy -- GPU reconstruction (built from source, CUDA 12.2)
• SAM 2.1 + SAM3 -- AI segmentation (Facebook Research)
• PyTorch -- cu121 (CUDA 12.1 runtime)
• CuPy, Zarr, OME-Zarr -- GPU array processing + cloud-native storage
• Globus SDK -- data transfer integration
• Python 3.11, numpy < 2.0

Repository Structure

tomocloud-container/
  apptainer/              # Apptainer/Singularity container
    txm-pipeline.def      # Definition file
    build.sh              # Build script (fakeroot, no root needed)

  docker/                 # OCI/Docker container
    Dockerfile            # Multi-stage build (builder + runtime)
    entrypoint.sh         # Container entrypoint
    .dockerignore

  deploy/                 # Deployment and registry scripts
    push-oci.sh           # Push SIF to OCI registry (ORAS)
    build-and-push-docker.sh  # Build Docker image + push to registry
    run-with-globus.sh    # Run with Globus/HuggingFace bind mounts

Quick Start

Option 1: Apptainer (HPC / Slurm)

cd apptainer/
bash build.sh
# Output: txm-pipeline.sif (~6-7 GB)

Option 2: Docker / OCI (AWS, Kubernetes, Cloud)

cd docker/
docker build -t tomocloud-pipeline:latest .

Push to Registry

From Apptainer SIF:

bash deploy/push-oci.sh txm-pipeline.sif ghcr.io/<org>/tomocloud-pipeline:latest

From Docker build:

bash deploy/build-and-push-docker.sh ghcr.io/<org>/tomocloud-pipeline:latest

Running

Apptainer + Slurm

srun --gpus=4 --time=03:00:00 \
  apptainer exec --nv \
    --bind /data:/data \
    --bind /models:/models \
    --bind ~/.cache/huggingface:/hf_cache \
    txm-pipeline.sif python /app/reconstruct_gpu.py ...

Docker / AWS

docker run --gpus all \
  -v /data:/data \
  -v /models:/models \
  -v ~/.cache/huggingface:/hf_cache \
  tomocloud-pipeline:latest python /app/reconstruct_gpu.py ...

With Globus Integration

bash deploy/run-with-globus.sh apptainer \
  --sif txm-pipeline.sif \
  --data-dir /path/to/data \
  --output-dir /path/to/output \
  --model-dir /path/to/models \
  -- python /app/reconstruct_gpu.py ...

CUDA Compatibility

• Base image: nvidia/cuda:12.2.2 (devel for build, runtime for final)
• Tested drivers: 535.x+ (S3DF L40S), 550.x+ (AWS)
• PyTorch cu121 wheels are forward-compatible with CUDA 12.2+ drivers

Build Notes

• tomocupy MUST be built with --no-build-isolation -- pip's isolated build env produces .so files with no embedded CUDA device code
• numpy must stay < 2.0 -- tomocupy and scipy 1.13 are compiled against numpy 1.x ABI
• SAM3 requires HuggingFace token for gated model access -- bind mount ~/.cache/huggingface into the container

Tested Platforms

Platform	GPUs	Status
SLAC S3DF (ada)	L40S 46GB	Validated
ALCF Polaris	A100 40GB	Validated
NERSC Perlmutter	A100 40GB	Validated
AWS (via IRI API)	TBD	In progress

Facilities

This container is designed for use across D4 tomography pipelines:

• SSRL (SLAC) -- beamline 6-2 TXM
• APS (ANL) -- tomography beamlines
• ALS (LBNL) -- tomography beamlines

Data transfer via Globus enables seamless integration with any facility endpoint.