Skip to main content

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
BoundaryRule
CheckpointUse the exact hf_id and immutable revision retained by the evidence. Same-family fine-tunes are best-effort until tested.
BundleInspect compatibility metadata and do not edit sections by hand.
Native runtimeUse the matching model/backend DSOs and a compatible TensorRT/CUDA cohort.
Platform specializationMatch the exact provider/profile target and options; do not generalize it to native platform support.
CUDA codeCompile native extensions and DSOs for the intended SM architecture.
Multi-deviceMatch 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.