Recover on CPU after CUDA inference failure

[Slavenin]

Problem

A CUDA execution provider can accept a session at startup but fail only when the first model kernel runs.

This happens with CUDA provider builds that do not contain a kernel image for a newer GPU architecture. For example, RTX 50 / Blackwell GPUs (SM 120) can fail with:

CUDA error cudaErrorNoKernelImageForDevice:
no kernel image is available for execution on the device

Before this change, LinuxBroadcast logged the inference error and dropped every subsequent frame, leaving the virtual camera without usable output.

Changes

• Retain the selected ONNX model bytes in the feeder.

• When GPU inference fails, rebuild the segmenter with the CPU execution provider.

• Retry the frame that triggered the GPU error on CPU, so the virtual camera continues immediately.

• Publish the live backend state from the feeder to the Pipeline and GUI.

• Show a persistent GPU unavailable · CPU footer badge after recovery.

  • The original technical error is available in the badge hover text.

• Log an explicit warning explaining that GPU acceleration was disabled and CPU processing is now active.

• Document the RTX 50 / Blackwell limitation and the requirement for a custom ONNX Runtime CUDA provider built with CMAKE_CUDA_ARCHITECTURES=120 (or native).

The fallback is intentionally not limited to Blackwell. Session creation alone does not prove that a CUDA provider can execute every model kernel, so any delayed GPU inference failure is recovered the same way.

Validation

• cargo test --workspace

  • 65 tests passed
  • 1 test ignored

• Manually reproduced on an RTX 5070:

  • CUDA session initialization succeeded.
  • The first RVM inference failed with cudaErrorNoKernelImageForDevice.
  • LinuxBroadcast logged the fallback warning.
  • Processing continued successfully on CPU.

Example log after this change:

[WARN lb_pipeline::lazy] GPU acceleration was disabled after an inference error;
LinuxBroadcast is now using CPU. ort run rvm: ... cudaErrorNoKernelImageForDevice ...
To restore GPU acceleration, install an ONNX Runtime CUDA provider built for this GPU.

Issue #1

[Pedrojok01]

Good catch, thank you for the fix!
LGTM