The hub dispatches each op on a single operand's device.device_type and forwards all tensors to that backend: distance.cpp keys on loc (:55,77,98,118), pushpull.cpp on out (:30-34), and identically posdef.cpp, resize.cpp, restrict.cpp, splinc.cpp, reg_field.cpp, reg_flow.cpp.
If operands disagree — e.g. out on CPU but inp/grid on CUDA (or vice-versa) — the wrong backend runs and reads a device pointer as host memory (or vice-versa) → segfault / silent garbage. Nothing upstream guards it (the Python wrappers don't either).
Fix: add a shared require_same_device(ref, others…) helper (compare device_typeanddevice_id) and call it at the top of every public entry, listing that function's tensors; throw std::invalid_argument on mismatch. Also throw a clear "not built with CUDA" error when a CUDA operand is seen but !FF_WITH_CUDA (instead of silently falling to the CPU branch or "unsupported device").
Verifiable on CPU: construct two DLTensors with different device.device_type and assert the throw (no GPU needed — it's the pre-dispatch guard).
Found in the cross-repo audit.
The hub dispatches each op on a single operand's
device.device_typeand forwards all tensors to that backend:distance.cppkeys onloc(:55,77,98,118),pushpull.cpponout(:30-34), and identicallyposdef.cpp,resize.cpp,restrict.cpp,splinc.cpp,reg_field.cpp,reg_flow.cpp.If operands disagree — e.g.
outon CPU butinp/gridon CUDA (or vice-versa) — the wrong backend runs and reads a device pointer as host memory (or vice-versa) → segfault / silent garbage. Nothing upstream guards it (the Python wrappers don't either).Fix: add a shared
require_same_device(ref, others…)helper (comparedevice_typeanddevice_id) and call it at the top of every public entry, listing that function's tensors; throwstd::invalid_argumenton mismatch. Also throw a clear "not built with CUDA" error when a CUDA operand is seen but!FF_WITH_CUDA(instead of silently falling to the CPU branch or "unsupported device").Verifiable on CPU: construct two
DLTensors with differentdevice.device_typeand assert the throw (no GPU needed — it's the pre-dispatch guard).Found in the cross-repo audit.