[Bug]: Dereverb models fail to load in latest version (works fine in 0.37.1)

Describe the bug

I was testing some dereverb models in the latest version of python-audio-separator, and found that a few of them no longer work, even though they work fine in version 0.37.1.

Affected models:

dereverb_super_big_mbr_ep_346.ckpt
dereverb_echo_mbr_fused.ckpt
dereverb_big_mbr_ep_362.ckpt

Have you searched for existing issues? 🔎

• I have searched and found no existing issues.

Screenshots or Videos

No response

Logs

--- Testing model [208]: dereverb_super_big_mbr_ep_346.ckpt ---
[2025-10-23 17:05:41,492: INFO/ForkPoolWorker-1] Separator version 0.39.1 instantiating with output_dir: download/dereverb_super_big_mbr_ep_346.ckpt_1b3aae8e4c08405590dc69e655c3ba3c, output_format: MP3
[2025-10-23 17:05:41,492: INFO/ForkPoolWorker-1] Using model directory from model_file_dir parameter: model/
[2025-10-23 17:05:41,492: INFO/ForkPoolWorker-1] Operating System: Linux #85~22.04.1-Ubuntu SMP PREEMPT_DYNAMIC Fri Sep 19 16:18:59 UTC 2
[2025-10-23 17:05:41,492: INFO/ForkPoolWorker-1] System: Linux Node: kumar Release: 6.8.0-85-generic Machine: x86_64 Proc: x86_64
[2025-10-23 17:05:41,492: INFO/ForkPoolWorker-1] Python Version: 3.13.8
[2025-10-23 17:05:41,492: INFO/ForkPoolWorker-1] PyTorch Version: 2.9.0+cu128
[2025-10-23 17:05:41,529: INFO/ForkPoolWorker-1] FFmpeg installed: ffmpeg version 4.4.2-0ubuntu0.22.04.1+esm10 Copyright (c) 2000-2021 the FFmpeg developers
[2025-10-23 17:05:41,530: INFO/ForkPoolWorker-1] ONNX Runtime GPU package installed with version: 1.23.0
[2025-10-23 17:05:41,530: INFO/ForkPoolWorker-1] CUDA is available in Torch, setting Torch device to CUDA
[2025-10-23 17:05:41,530: INFO/ForkPoolWorker-1] ONNXruntime has CUDAExecutionProvider available, enabling acceleration
[2025-10-23 17:05:41,530: INFO/ForkPoolWorker-1] Loading model dereverb_super_big_mbr_ep_346.ckpt...

[2025-10-23 17:08:50,164: WARNING/ForkPoolWorker-1] Could not detect model type from config or path model/dereverb_super_big_mbr_ep_346.ckpt, defaulting to bs_roformer
[2025-10-23 17:08:50,247: ERROR/ForkPoolWorker-1] Failed to instantiate Roformer model: the number of freqs in the bands must equal 1025 based on the STFT settings, but got 126. This may be due to missing parameters or configuration validation failures.
[2025-10-23 17:08:50,248: WARNING/ForkPoolWorker-1] ❌ Failed on dereverb_super_big_mbr_ep_346.ckpt: Failed to instantiate Roformer model: the number of freqs in the bands must equal 1025 based on the STFT settings, but got 126. This may be due to missing parameters or configuration validation failures.

System Info

Operating System: Ubuntu 22.04
Python version: 3.13
python-audio-separator:  0.39.1

Additional Information

Other models (non-dereverb ones) tested are working normally.

Reverting to version 0.37.1 fixes the issue completely (models load and separate correctly).

Reproduce loading dereverb_super_big_mbr_ep_346.ckpt, dereverb_echo_mbr_fused.ckpt, and dereverb_big_mbr_ep_362.ckpt in version 0.39.1, then compare the model-loading behavior with 0.37.1. Use the logged Roformer configuration error as the starting point; done means all three models load and separate correctly in the latest version.