TEST DATALive backend is not connected. Values below come from contract fixtures and are not real supply, balances or evidence.
Templates
A small catalog of tested images. Every template — and every compatible custom OCI image — runs under the same signed launch profile: non-root user, dropped capabilities, read-only root filesystem, no privileged mode, no host mounts or namespaces, exact inbound ports and restricted egress.
PyTorch + CUDA
- Version
- pytorch-cuda-v1.4.0
- Image digest
registry.amplerun.com/templates/pytorch-cuda@sha256:aaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaa- Minimum GPUs
- 1
- Minimum VRAM
- 16 GiB
- Runtime
- cuda
- Ports
- 22, 8888
- Launch profile
- launch-v1.2.0
Readiness: External SSH probe plus Jupyter health check on the leased endpoint.
Tenant SSH key is injected into the declared tenant directory. Jupyter binds container-local access with a short-lived token; connect via SSH local forwarding.
Jupyter + CUDA (minimal)
- Version
- jupyter-cuda-v1.2.1
- Image digest
registry.amplerun.com/templates/jupyter-cuda@sha256:bbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbb- Minimum GPUs
- 1
- Minimum VRAM
- 8 GiB
- Runtime
- cuda
- Ports
- 22, 8888
- Launch profile
- launch-v1.2.0
Readiness: External SSH probe; notebook server answers on the leased HTTPS endpoint.
Minimal notebook image for interactive work. Data directory is the job-owned writable volume.
PyTorch + ROCm
- Version
- pytorch-rocm-v0.9.2
- Image digest
registry.amplerun.com/templates/pytorch-rocm@sha256:cccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccc- Minimum GPUs
- 1
- Minimum VRAM
- 32 GiB
- Runtime
- rocm-hip
- Ports
- 22
- Launch profile
- launch-v1.2.0
Readiness: External SSH probe bound to the leased endpoint, job and fence.
Community qualification track. OpenCL discovery alone does not prove PyTorch compatibility.
Custom OCI images
A custom image is accepted when it satisfies the same launch profile and pinned digest checks. Images that request privileged containers, arbitrary mounts or unrestricted host networking fail qualification with a concrete compatibility message. Discovery of OpenCL alone does not prove PyTorch compatibility.
