Handbook

Granite NN worker images (LCDL)

Forge LCDL NN tasks (nn_fleet_infer, nn_train, nn_eval, nn_generate) submit Fleet docker_argv workspace jobs with NVIDIA GPU flags.

Updated 2026-10-22

Requirements

Granite host with NVIDIA Container Toolkit
Fleet bearer token and POST /v1/jobs access
Worker image with Python 3.11+, TensorFlow GPU for MNIST training

Images

Image	Role
`forge-lcdl-nn-infer-worker:local`	Default infer stub / light worker
`forge-lcdl-nn-gpu-worker:local`	CUDA MNIST train/eval/infer (`deploy/nn-gpu-worker/Dockerfile`)

Build GPU worker from forge-lcdl root:

docker build -f deploy/nn-gpu-worker/Dockerfile -t forge-lcdl-nn-gpu-worker:local .

Entry: python -m forge_lcdl.nn.fleet_worker

Override with FORGE_LCDL_NN_FLEET_INFER_IMAGE.

Fleet template meta

LCDL sets job meta:

{
  "container_class": "docker_argv_workspace",
  "workspace_upload_required": true,
  "nn_worker_kind": "train",
  "use_fleet_template_image": true,
  "requirements": ["lcdl_nn_gpu_worker"]
}

Set FORGE_LCDL_NN_FLEET_TEMPLATE_REQUIREMENTS=lcdl_nn_gpu_worker for Granite template resolution.

docker argv shape

docker run --rm --gpus all <image> python -m forge_lcdl.nn.fleet_worker

Workspace tarball contains request.json with op: infer | train | eval | generate.

Worker POSTs /v1/jobs/{id}/workspace-worker-complete with metrics, artifact_path, device_used, architecture.

MNIST Granite learning

Prepare categorized digits on operator host: python scripts/prepare-handwritten-digits.py --full
export FORGE_LCDL_GRANITE_GPU=1
./scripts/run-mnist-granite-gpu.sh cnn_lenet or --full for all Tier A

Manual Granite smoke

Confirm /v1/health reports host.gpu.nvidia.available: true
export FORGE_LCDL_TASK_PACKS=nn
Run nn_fleet_infer with keras model and require_gpu: true
Poll job until completed; verify worker_result.device_used is GPU

See LCDL NN backends in forge-lcdl.

Forge Fleet