IPC coupler with Genesis crashes on RTX 5070 (sm_120)

[55155]

I'm trying out the IPC coupler. When I ran the example provided by Genesis, the runtime kept dying with an error. The diagnosis from Claude Code is as follows.

---

cudaErrorInvalidDevice in cub::DeviceScan on RTX 5070 (sm_120) — occurs only via the Genesis IPC coupler

Environment: RTX 5070 (sm_120, Blackwell), driver 581.80, Win11, Python 3.11.9, pyuipc 0.0.25, genesis-world 1.1.0, torch 2.11.0+cu128, warp-lang 1.13.0 (no system CUDA Toolkit; the CUDA runtime DLLs are supplied into the process by torch/lib).

What works: Every official uipc.assets example (cube_ground, fem_link_drop, fem_bouncing_cubes, ipc_2cubes_fall, abd_fem_tower) runs for dozens of frames without issue on sm_120 — so the prebuilt CUDA kernels do in fact execute on Blackwell.

What fails: It only fails when libuipc is driven through Genesis's IPCCoupler. Even for a tiny scene (Steiner points: 52), the very first World::advance() dies with cub::DeviceScan::ExclusiveSum(...) code=101 cudaErrorInvalidDevice (muda/cub/device/device_scan.h:36), followed by a C++ terminate. This is not a scan-size issue.

Questions: Is this a known Blackwell symptom caused by _stream being created on a different CUDA context than libuipc's engine (i.e. Warp/Torch initializing the device context first)? Does advance() assume it owns the current device/context? Is there a recommended way to bind libuipc to a host-created context/stream? Can you reproduce this on a Blackwell card? I can provide a self-contained Genesis repro and full logs.

---

The key point is that when uipc is run standalone, it is fully compatible with my environment. I'm wondering whether this error is something only I am hitting, or whether it's reproducible on your side. Thank you for providing such an excellent simulator as open source.