Container workloads

Submit ROCm images

Declare AMD and ROCm compatibility when a container workload needs the ROCm runtime instead of CUDA.

ROCm submission contract

✓Use an image that contains the ROCm libraries required by your framework and application.
✓Set gpu_vendor to amd and runtime to rocm so routing considers only compatible capacity.
✓Use framework hints such as pytorch-rocm, vllm-rocm, llama.cpp-rocm, or onnx-rocm when applicable.
✓CUDA images and NVIDIA-only commands are rejected when they conflict with explicit ROCm intent.

Submit a ROCm workload

ROCm uses the normal container job API. The only additional contract is explicit accelerator compatibility: use an appropriate image and declare AMD plus ROCm intent.

•gpu_vendor accepts amd for this route.
•runtime is the public request field for the ROCm accelerator runtime.
•framework is optional but helps compatibility screening when the image name is ambiguous.
•All normal job status, logs, callbacks, cancellation, and artifact endpoints remain unchanged.

Job request

POST /v1/jobs

{
  "name": "rocm-inference",
  "image": "ghcr.io/acme/pytorch-rocm:latest",
  "workload_type": "inference",
  "command": ["python", "infer.py"],
  "gpu_vendor": "amd",
  "runtime": "rocm",
  "framework": "pytorch-rocm",
  "expected_artifacts": [
    "/workspace/artifacts/result.json"
  ]
}

ROCm-aware routing and screening

Explicit AMD or ROCm requests consider only compatible capacity. CUDA images, NVIDIA labels, nvidia-smi commands, and known CUDA-only workload hints are blocked before execution.

•Use io.junglegrid.runtime=rocm or an AMD label when a build pipeline can supply OCI metadata.
•Recognized framework hints are pytorch-rocm, vllm-rocm, llama.cpp-rocm, and onnx-rocm.
•Unknown CUDA-heavy containers are never silently moved to AMD; declare ROCm compatibility or select an NVIDIA route.

Prepare the image

The image must include the ROCm user-space libraries and framework build needed by the command. Do not start from a CUDA image and expect the runtime to translate it.

•Build and test the image against the ROCm version your application supports.
•Write output files under /workspace/artifacts so Jungle Grid can collect them.
•Use /workspace/inputs and /workspace/scripts for uploaded inputs and scripts.
•Avoid hard-coded NVIDIA device paths, CUDA environment variables, and nvidia-smi probes.

Troubleshooting

Compatibility errors are returned before execution when declared intent conflicts with the image, command, labels, or framework. Capacity can still be temporarily unavailable after a request passes screening.

•If screening reports CUDA incompatibility, use a ROCm-native image and remove NVIDIA-only commands or labels.
•If rocm-smi or framework imports fail at runtime, verify that the image contains the matching ROCm user-space stack.
•If no compatible capacity is currently available, keep the same workload request and retry later.
•Inspect job status, lifecycle events, and logs before changing image or routing fields.