Many problems here with CUDA/docker on debian (dependency-problems + first initialiazation + transcriptions)

Start with docker-compose.yml, docker-compose.build.blackwell.yml, and Dockerfile.cuda.12.9, then trace the WhisperX paths and dependency errors shown in the report. Reproduce and separate the volume-permission, GPU out-of-memory, and lilcom installation failures; done means each failure has a confirmed cause and a tested fix or documented limitation.

Hello there,
The project seems really good, but I encounter some problems, no matter what I try.

Environment

I'm running a Debian 13 + docker (docker's repo) + nvidia open drivers + cuda toolkit + container toolit.
Hardware : RTX 5070Ti + a lot of RAM (512GB).
Note: cuda upgraded to 13.1 recently (following the repo's releases), when debian upgraded to 13.3, kernel revision and so on. nvidia drivers upgraded 590.48.x

docker images used

As there is no prebuilt blackwell image, I built it as specified in the doc+yml.
I also tried the prebuilt CUDA:v1.2.0 in case. I won't describe the problems in this case, because of dependency-problems when transcripting, problems I can admit because of old libs for a blackwell environment/hardware.

Problem 1) DB iniatialization & permissions problems (for both images by the way)

As stated in the doc, some permission problems might appear with the 2 volumes on the host. The best solution I found is to force the entrypoint to /app/scriberr in the docker-compose.yml for 2nd and future starts. Some problem with the ubuntu/appuser and the OS user mapping?

Problem 2) dependency problem for parakeet (both images)

As mentionned in bug 384, parakeet won't build. Not a major problem as there are other models available.

Problem 3) Transcription failed - focus on blackwell locally built model, based on docker-compose.build.blackwell.yml and Dockerfile.cuda.12.9 - but I use 13.1, following nvidia's official releases).

With whisper / GPU : no matter what the model size and standard settings, it ends up with :
Fails with :
scriberr-local-blackwell | time=11:23:43 level=ERROR msg="Job processing failed" worker_id=0 job_id=1a76f472-88fc-4b95-a1c3-daa7169b0c9a error="single-track processing failed: transcription failed: WhisperX execution failed: exit status 1\nLogs:\n/1370 for more details.\n\n warnings.warn(\n2026-01-16 11:23:41 - whisperx.asr - INFO - Detected language: fr (0.22) in first 30s of audio\nTraceback (most recent call last):\n File \"/root/.local/share/uv/python/cpython-3.10.19-linux-x86_64-gnu/lib/python3.10/runpy.py\", line 196, in _run_module_as_main\n return _run_code(code, main_globals, None,\n File \"/root/.local/share/uv/python/cpython-3.10.19-linux-x86_64-gnu/lib/python3.10/runpy.py\", line 86, in _run_code\n exec(code, run_globals)\n File \"/app/whisperx-env/WhisperX/whisperx/__main__.py\", line 102, in <module>\n cli()\n File \"/app/whisperx-env/WhisperX/whisperx/__main__.py\", line 98, in cli\n transcribe_task(args, parser)\n File \"/app/whisperx-env/WhisperX/whisperx/transcribe.py\", line 150, in transcribe_task\n result: TranscriptionResult = model.transcribe(\n File \"/app/whisperx-env/WhisperX/whisperx/asr.py\", line 261, in transcribe\n for idx, out in enumerate(self.__call__(data(audio, vad_segments), batch_size=batch_size, num_workers=num_workers)):\n File \"/app/whisperx-env/WhisperX/.venv/lib/python3.10/site-packages/transformers/pipelines/pt_utils.py\", line 126, in __next__\n item = next(self.iterator)\n File \"/app/whisperx-env/WhisperX/.venv/lib/python3.10/site-packages/transformers/pipelines/pt_utils.py\", line 127, in __next__\n processed = self.infer(item, **self.params)\n File \"/app/whisperx-env/WhisperX/.venv/lib/python3.10/site-packages/transformers/pipelines/base.py\", line 1374, in forward\n model_outputs = self._forward(model_inputs, **forward_params)\n File \"/app/whisperx-env/WhisperX/whisperx/asr.py\", line 165, in _forward\n outputs = self.model.generate_segment_batched(model_inputs['inputs'], self.tokenizer, self.options)\n File \"/app/whisperx-env/WhisperX/whisperx/asr.py\", line 60, in generate_segment_batched\n encoder_output = self.encode(features)\n File \"/app/whisperx-env/WhisperX/whisperx/asr.py\", line 99, in encode\n return self.model.encode(features, to_cpu=to_cpu)\nRuntimeError: CUDA failed with error out of memory\n"
Memory problem must be a consequence of something else, considering my hardware.

** With whisper / CPU (in case there's something wrong with my nvidia setup - which I doubt)
Transcription is working 👌 **

** With nvidia canary :**
Fails because of language:
scriberr-local-blackwell | time=11:37:39 level=ERROR msg="Job processing failed" worker_id=0 job_id=1a76f472-88fc-4b95-a1c3-daa7169b0c9a error="single-track processing failed: transcription failed: Canary execution failed: exit status 2\nLogs:\n\x96\x88████▏| 330M/360M [00:01<00:00, 293MB/s]\r 99%|█████████▉| 358M/360M [00:01<00:00, 293MB/s]\r100%|██████████| 360M/360M [00:01<00:00, 288MB/s]\n2026-01-16 11:30:32 - whisperx.transcribe - INFO - New language found (fr)! Previous was (en), loading new alignment model for new language...\nDownloading: \"https://download.pytorch.org/torchaudio/models/wav2vec2_voxpopuli_base_10k_asr_fr.pt\" to /root/.cache/torch/hub/checkpoints/wav2vec2_voxpopuli_base_10k_asr_fr.pt\n\r 0%| | 0.00/360M [00:00<?, ?B/s]\r 6%|▌ | 21.5M/360M [00:00<00:01, 225MB/s]\r 14%|█▎ | 49.5M/360M [00:00<00:01, 265MB/s]\r 22%|██▏ | 77.5M/360M [00:00<00:01, 278MB/s]\r 29%|██▉ | 106M/360M [00:00<00:00, 284MB/s] \r 37%|███▋ | 134M/360M [00:00<00:00, 288MB/s]\r 45%|████▍ | 162M/360M [00:00<00:00, 290MB/s]\r 53%|█████▎ | 190M/360M [00:00<00:00, 291MB/s]\r 60%|██████ | 218M/360M [00:00<00:00, 292MB/s]\r 68%|██████▊ | 246M/360M [00:00<00:00, 293MB/s]\r 76%|███████▌ | 274M/360M [00:01<00:00, 293MB/s]\r 84%|████████▍ | 302M/360M [00:01<00:00, 293MB/s]\r 92%|█████████▏| 330M/360M [00:01<00:00, 293MB/s]\r 99%|█████████▉| 358M/360M [00:01<00:00, 293MB/s]\r100%|██████████| 360M/360M [00:01<00:00, 288MB/s]\n2026-01-16 11:30:34 - whisperx.transcribe - INFO - Performing alignment...\nerror: Distribution 'lilcom==1.8.2 @ registry+https://pypi.org/simple' can't be installed because it doesn't have a source distribution or wheel for the current platform\n\nhint: You're on Linux ('manylinux_2_39_x86_64'), but 'lilcom' (v1.8.2) only has wheels for the following platforms: 'manylinux_2_17_aarch64', 'manylinux2014_aarch64', 'macosx_10_13_universal2', 'win_amd64'; consider adding \"sys_platform == 'linux' and platform_machine == 'x86_64'\" to 'tool.uv.required-environments' to ensure uv resolves to a version with compatible wheels\n

Regards. Tell me if I can do some tests for this specific hardware.