nx_nv.nx

nx_nv.nx -- the native-Linux NVIDIA OPEN-KERNEL-MODULE device shim (TIER-1 sovereign executor), sibling to nx_dxg (the WSL2 GPU-PV shim). This is the path that ACTUALLY EXECUTES on real silicon -- modeled on tinygrad's ops_nv.py, which submits via /dev/nvidia* + a GPFIFO ring + a usermode doorbell, bypassing the CUDA runtime (and bypassing the proprietary dxg priv-data handshake that walls WSL2). Runs on a native-Linux / rented cloud GPU instance (brick-safe, rule #26). LOCALLY-VERIFIABLE (gated bit-exact here): the GPFIFO entry encoder + the doorbell offset. CLOUD-PENDING (structured, completed against the instance's open-gpu-kernel-modules headers on first run): the full RM-API object tree -- NV_ESC_RM_ALLOC on /dev/nvidiactl for root client -> device -> subdevice -> channel-group -> GPFIFO channel (NV_CHANNELGPFIFO_ALLOCATION_PARAMETERS) -> compute object; NV_ESC_RM_CONTROL for GPFIFO_SCHEDULE + GET_WORK_SUBMIT_TOKEN; NV_ESC_RM_MAP_MEMORY for the usermode doorbell BAR. Pure funcs, no main. license_tier: ORIGINAL