Compatibility
Compatibility is a tuple, not a single project version:
model revision × build configuration × bundle contract × runtime implementation
× GPU/SM × driver/CUDA/TensorRT × host architecture
| Boundary | Rule |
|---|---|
| Checkpoint | Use the exact hf_id and immutable revision retained by the evidence. Same-family fine-tunes are best-effort until tested. |
| Bundle | Inspect compatibility metadata and do not edit sections by hand. |
| Native runtime | Use the matching model/backend DSOs and a compatible TensorRT/CUDA cohort. |
| Platform specialization | Match the exact provider/profile target and options; do not generalize it to native platform support. |
| CUDA code | Compile native extensions and DSOs for the intended SM architecture. |
| Multi-device | Match the topology and world size compiled into the bundle. |
The Supported Models page lists declared checkpoint/configuration contracts. A current target-hardware E2E result is the evidence required for a verified target claim.