Non-blocking follow-up to completed epic #2951 — the extensibility hook for NVIDIA/CUDA hosts (federated Linux+CUDA peers or future hardware). This is outside #2951's shipped Apple-Silicon scope and does not block that epic. The EXECUTION_LANE.LOCAL_CUDA lane already exists in the registry (server/services/imageTo3d/targets.js) with a requires-cuda gate; this wires an actual CUDA installer/runner behind it.
Scope
- A CUDA install path (conda/PyTorch 12.4, upstream
microsoft/TRELLIS.2 setup.sh) + runner, surfaced ONLY when detectHostCapabilities().cuda is true.
- Real CUDA GPU capability detection to populate
caps.cuda (currently hardcoded false in detectHostCapabilities).
Why blocked
Cannot be built/validated on this install's Apple Silicon hardware — needs a Linux + NVIDIA (24 GB+ VRAM, CUDA 12.4) box to test end-to-end. Unblocks when such hardware / a CUDA sync peer is available. See #2951.
Non-blocking follow-up to completed epic #2951 — the extensibility hook for NVIDIA/CUDA hosts (federated Linux+CUDA peers or future hardware). This is outside #2951's shipped Apple-Silicon scope and does not block that epic. The
EXECUTION_LANE.LOCAL_CUDAlane already exists in the registry (server/services/imageTo3d/targets.js) with arequires-cudagate; this wires an actual CUDA installer/runner behind it.Scope
microsoft/TRELLIS.2setup.sh) + runner, surfaced ONLY whendetectHostCapabilities().cudais true.caps.cuda(currently hardcoded false indetectHostCapabilities).Why blocked
Cannot be built/validated on this install's Apple Silicon hardware — needs a Linux + NVIDIA (24 GB+ VRAM, CUDA 12.4) box to test end-to-end. Unblocks when such hardware / a CUDA sync peer is available. See #2951.