Use the official nvidia/RE-USE checkpoint at revision 022e920d727347a64d6c21fbf0628f2a5f37ad78. Keep original weights outside the GGUF output directory. The conversion script verifies shapes and retained F32 tensor bytes against the original safetensors.

python tests/reuse/convert_reuse_gguf.py \
  --source /path/to/RE-USE --output /path/to/RE-USE-GGUF/reuse-f32.gguf \
  --log build/logs/reuse/conversion.log

The runner calls the official SEMamba model and STFT/ISTFT implementations. Run in the qwen3-tts environment with PyTorch, mamba_ssm, causal_conv1d, safetensors, einops, and soundfile installed.

python tests/reuse/run_python_reference.py \
  --reference-root /path/to/RE-USE \
  --audio /path/to/RE-USE/noisy_audio/mic_test2.wav \
  --repeats 4 --out build/logs/reuse/python.wav \
  --log build/logs/reuse/python.log

Ordinary performance runs leave TF32 and model precision at upstream defaults. For separate controlled parity runs, set NVIDIA_TF32_OVERRIDE=0 for both implementations and pass --no-tf32 to Python. Do not report controlled-run timings as ordinary performance. Use --out-format float32 in the CLI when comparing final waveforms. Keep timing measurements sequential.

Compare full-recording restoration at multiple sample rates, stereo channels, and repeated requests. Long-form checks must use the same chunk duration and overlap in both implementations. Python exposes --audio-chunk-duration-sec and --audio-chunk-overlap-sec; the corresponding C++ request options are audio_chunk_duration_sec and audio_chunk_overlap_sec.

test_reuse_scan exercises Mamba-1 state, batch, and strided B/C layouts without model weights. Build it with model tests enabled and select --backend vulkan. This does not replace final-waveform checks or validate Mamba-2 regressions.