CUDA initialization failiure on latest release drivers

NVIDIA Open GPU Kernel Modules Version

570.86.16 [Edit: Still failing on 570.124.04]

Please confirm this issue does not happen with the proprietary driver (of the same version). This issue tracker is only for bugs specific to the open kernel driver.

• I confirm that this does not happen with the proprietary driver package.

Operating System and Version

NixOS 25.05 (Warbler)

Kernel Release

Linux 6.13.4 #1-NixOS SMP PREEMPT_DYNAMIC Fri Feb 21 13:11:21 UTC 2025 x86_64 GNU/Linux

Please confirm you are running a stable release kernel (e.g. not a -rc). We do not accept bug reports for unreleased kernels.

• I am running on a stable kernel release.

Hardware: GPU

NVIDIA GeForce RTX 3080 Laptop GPU

Describe the bug

The specific issue seems to have first been described here: https://github.com/NVIDIA/open-gpu-kernel-modules/discussions/780

When attempting to initialize NVDEC/CUDA in any capacity:

cu->cuInit(0) failed -> CUDA_ERROR_NOT_INITIALIZED: initialization

or

        53.376880366 [2624-2624] ../src/direct/nv-driver.c: 295            init_nvdriver Initing nvdriver...
        53.376904963 [2624-2624] ../src/direct/nv-driver.c: 313            init_nvdriver NVIDIA kernel driver version: 570.86.16, major version: 570, minor version: 86
        53.376909963 [2624-2624] ../src/direct/nv-driver.c: 320            init_nvdriver Got dev info: 100 1 2 6
        53.386246461 [2624-2624] ../src/direct/direct-export-buf.c:  27       findGPUIndexFromFd CUDA ERROR 'initialization error' (3)

        53.386254391 [2624-2624] ../src/vabackend.c:2252       __vaDriverInit_1_0 CUDA ERROR 'initialization error' (3)

is returned.

This does not occur on proprietary kernel.

Setting "uvm_disable_hmm=1" on open kernel, as described in the above link, SOMETIMES worksaround/solves the issue.

To Reproduce

This occurred when using any program trying to use CUDA/NVDEC.

I ran into it while using:

ffmpeg -y -vsync 0 -hwaccel cuda -hwaccel_output_format cuda -i input.mp4 -c:a copy -c:v h264_nvenc -b:v 5M output.mp4

which would return

[AVHWDeviceContext @ 0x359865c0] cu->cuInit(0) failed -> CUDA_ERROR_NOT_INITIALIZED: initialization error.

Similar initialization error would occur when invoking "vainfo" [through nvidia-vaapi-driver]:

        53.376880366 [2624-2624] ../src/direct/nv-driver.c: 295            init_nvdriver Initing nvdriver...
        53.376904963 [2624-2624] ../src/direct/nv-driver.c: 313            init_nvdriver NVIDIA kernel driver version: 570.86.16, major version: 570, minor version: 86
        53.376909963 [2624-2624] ../src/direct/nv-driver.c: 320            init_nvdriver Got dev info: 100 1 2 6
        53.386246461 [2624-2624] ../src/direct/direct-export-buf.c:  27       findGPUIndexFromFd CUDA ERROR 'initialization error' (3)

        53.386254391 [2624-2624] ../src/vabackend.c:2252       __vaDriverInit_1_0 CUDA ERROR 'initialization error' (3)

Bug Incidence

Always

nvidia-bug-report.log.gz

nvidia-bug-report-BUG.log.gz

nvidia-bug-report-SUCCESSFUL-EX.log.gz

nvidia-bug-report-FAILURE-WITH-UVM-DISABLE-HMM.log.gz

Success generated on open kernel with uvm_disable_hmm=1, which also still sometimes fails to fix

More Info

Start with the CUDA/NVDEC reproducer, the attached nvidia-bug-report logs, and the linked discussion about uvm_disable_hmm=1; compare the failing and successful open-driver runs with the proprietary-driver behavior. Done means CUDA and NVDEC initialize reliably on the open kernel driver without requiring the workaround.