ROCm not working in repaint: No soundfile backend in process_reference_audio

[shipped-it]

Describe the bug
torchaudio relies on torchcodec, so we can't use an AMD GPU for repainting.

2026-02-21 20:25:02.418 | ERROR    | acestep.core.generation.handler.generate_music:generate_music:183 - [generate_music] Generation failed
Traceback (most recent call last):

  File "/mnt/HeavyData/AI/ACE-Step-1.5/.venv/lib/python3.11/site-packages/torchaudio/_torchcodec.py", line 82, in load_with_torchcodec
    from torchcodec.decoders import AudioDecoder

ModuleNotFoundError: No module named 'torchcodec'


The above exception was the direct cause of the following exception:


Traceback (most recent call last):

  File "/home/user/.local/share/uv/python/cpython-3.11.14-linux-x86_64-gnu/lib/python3.11/threading.py", line 1002, in _bootstrap
    self._bootstrap_inner()
    │    └ <function Thread._bootstrap_inner at 0x7f1710ee5c60>
    └ <WorkerThread(AnyIO worker thread, started daemon 139710781454016)>
  File "/home/user/.local/share/uv/python/cpython-3.11.14-linux-x86_64-gnu/lib/python3.11/threading.py", line 1045, in _bootstrap_inner
    self.run()
    │    └ <function WorkerThread.run at 0x7f11d5706a20>
    └ <WorkerThread(AnyIO worker thread, started daemon 139710781454016)>
  File "/mnt/HeavyData/AI/ACE-Step-1.5/.venv/lib/python3.11/site-packages/anyio/_backends/_asyncio.py", line 986, in run
    result = context.run(func, *args)
             │       │   │      └ (<generator object register_generation_run_handlers.<locals>.generation_wrapper at 0x7f11d8ba1000>,)
             │       │   └ <function run_sync_iterator_async at 0x7f1302f76d40>
             │       └ <method 'run' of '_contextvars.Context' objects>
             └ <_contextvars.Context object at 0x7f11dd1fb280>
  File "/mnt/HeavyData/AI/ACE-Step-1.5/.venv/lib/python3.11/site-packages/gradio/utils.py", line 833, in run_sync_iterator_async
    return next(iterator)
                └ <generator object register_generation_run_handlers.<locals>.generation_wrapper at 0x7f11d8ba1000>
  File "/mnt/HeavyData/AI/ACE-Step-1.5/.venv/lib/python3.11/site-packages/gradio/utils.py", line 1017, in gen_wrapper
    response = next(iterator)
                    └ <generator object register_generation_run_handlers.<locals>.generation_wrapper at 0x7f11d573a260>

  File "/mnt/HeavyData/AI/ACE-Step-1.5/acestep/ui/gradio/events/wiring/generation_run_wiring.py", line 45, in generation_wrapper
    yield from res_h.generate_with_batch_management(dit_handler, llm_handler, *args)
               │     │                              │            │             └ ('', '[Instrumental]', None, '', '', 'unknown', 8, 7, False, '1829544894', '/tmp/gradio/5f72c94c1ec50b94e223adbcbb25840953861...
               │     │                              │            └ <acestep.llm_inference.LLMHandler object at 0x7f13022057d0>
               │     │                              └ <acestep.handler.AceStepHandler object at 0x7f13022f1ad0>
               │     └ <function generate_with_batch_management at 0x7f154d97d1c0>
               └ <module 'acestep.ui.gradio.events.results_handlers' from '/mnt/HeavyData/AI/ACE-Step-1.5/acestep/ui/gradio/events/results_han...

  File "/mnt/HeavyData/AI/ACE-Step-1.5/acestep/ui/gradio/events/results/batch_management.py", line 76, in generate_with_batch_management
    for partial_result in generator:
                          └ <generator object generate_with_progress at 0x7f0ed405d970>

  File "/mnt/HeavyData/AI/ACE-Step-1.5/acestep/ui/gradio/events/results/generation_progress.py", line 159, in generate_with_progress
    result = generate_music(dit_handler, llm_handler, params=gen_params, config=gen_config, progress=progress)
             │              │            │                   │                  │                    └ <gradio.helpers.Progress object at 0x7f154d988dd0>
             │              │            │                   │                  └ GenerationConfig(batch_size=2, allow_lm_batch=True, use_random_seed=False, seeds=[1829544894], lm_batch_chunk_size=8, constra...
             │              │            │                   └ GenerationParams(task_type='text2music', instruction='Fill the audio semantic mask based on the given conditions:', reference...
             │              │            └ <acestep.llm_inference.LLMHandler object at 0x7f13022057d0>
             │              └ <acestep.handler.AceStepHandler object at 0x7f13022f1ad0>
             └ <function generate_music at 0x7f154d959800>

  File "/mnt/HeavyData/AI/ACE-Step-1.5/acestep/inference.py", line 581, in generate_music
    result = dit_handler.generate_music(
             │           └ <function GenerateMusicMixin.generate_music at 0x7f1550d8a660>
             └ <acestep.handler.AceStepHandler object at 0x7f13022f1ad0>

> File "/mnt/HeavyData/AI/ACE-Step-1.5/acestep/core/generation/handler/generate_music.py", line 111, in generate_music
    refer_audios, processed_src_audio, audio_error = self._prepare_reference_and_source_audio(
                                                     │    └ <function GenerateMusicRequestMixin._prepare_reference_and_source_audio at 0x7f1550dc44a0>
                                                     └ <acestep.handler.AceStepHandler object at 0x7f13022f1ad0>

  File "/mnt/HeavyData/AI/ACE-Step-1.5/acestep/core/generation/handler/generate_music_request.py", line 97, in _prepare_reference_and_source_audio
    processed_ref_audio = self.process_reference_audio(reference_audio)
                          │    │                       └ '/tmp/gradio/5f72c94c1ec50b94e223adbcbb25840953861f08e9c9d9df9f21286dcdb0d65c/flight.wav'
                          │    └ <function IoAudioMixin.process_reference_audio at 0x7f1550dc76a0>
                          └ <acestep.handler.AceStepHandler object at 0x7f13022f1ad0>

  File "/mnt/HeavyData/AI/ACE-Step-1.5/acestep/core/generation/handler/io_audio.py", line 78, in process_reference_audio
    audio, sr = torchaudio.load(audio_file)
                │          │    └ '/tmp/gradio/5f72c94c1ec50b94e223adbcbb25840953861f08e9c9d9df9f21286dcdb0d65c/flight.wav'
                │          └ <function load at 0x7f1550b35ee0>
                └ <module 'torchaudio' from '/mnt/HeavyData/AI/ACE-Step-1.5/.venv/lib/python3.11/site-packages/torchaudio/__init__.py'>

  File "/mnt/HeavyData/AI/ACE-Step-1.5/.venv/lib/python3.11/site-packages/torchaudio/__init__.py", line 86, in load
    return load_with_torchcodec(
           └ <function load_with_torchcodec at 0x7f1550bb8720>
  File "/mnt/HeavyData/AI/ACE-Step-1.5/.venv/lib/python3.11/site-packages/torchaudio/_torchcodec.py", line 84, in load_with_torchcodec
    raise ImportError(

[coderabbitai]

⚠️ Possible Duplicate Issue(s)

• torchcodec needs to be built from source for ROCM (AMD) and probably Intel #462

---

📝 Issue Planner

_{Check the box below or use the @coderabbitai plan command to generate an implementation plan and prompts that you can use with your favorite coding assistant.}

• Create Plan

---

🧪 Issue enrichment is currently in open beta.

You can configure auto-planning by selecting labels in the issue_enrichment configuration.

To disable automatic issue enrichment, add the following to your .coderabbit.yaml:

issue_enrichment:
  auto_enrich:
    enabled: false

💬 Have feedback or questions? Drop into our discord!

[pkrasicki]

Reference audio, remix, repaint - none of those work. And I couldn't train a lora either. The dataset preprocessing step seems to require having this library as well.

[shipped-it]

#462 seems to imply that torchcodec is buildable for ROCM, then we need to reference manually? Still it would be best for this project to switch to soundfile when torchcodec is not available (like they say in some documentation file I read)

[pkrasicki]

I don't know, but meta-pytorch/torchcodec#444 seems to suggest that torchcodec is Nvidia only. Maybe we could use it on the CPU somehow.

[shipped-it]

I don't know, but meta-pytorch/torchcodec#444 seems to suggest that torchcodec is Nvidia only.

True, but soundfile should be used like they say it here

ACE-Step-1.5/docs/en/INSTALL.md

Line 406 in c0a32a2

> **Note:** `torchcodec` is not available for AMD ROCm GPUs due to CUDA-specific dependencies. ACE-Step automatically uses `soundfile` as a fallback for audio I/O, which provides full functionality on ROCm platforms.

It seems that this half LLM generated documentation piece hallucinated the soundfile fallback. When I read the source code, many functions rely exclusively on torchcodec.

[BeForU]

I changed io_audio.py temporarily. It seems something can run, but it's not clear.
path : ACE-Step-1.5\acestep\core\generation\handler\io_audio.py

process_reference_audio()

from :
import torchaudio
audio, sr = torchaudio.load(audio_file)

to :
import soundfile as sf
audio_np, sr = sf.read(audio_file, dtype="float32", always_2d=True)
audio = torch.from_numpy(audio_np).transpose(0, 1).contiguous()

process_src_audio()

from :
import torchaudio
audio, sr = torchaudio.load(audio_file)
return self._normalize_audio_to_stereo_48k(audio, sr)

to :
import soundfile as sf
audio_np, sr = sf.read(audio_file, dtype="float32", always_2d=True)
audio = torch.from_numpy(audio_np).transpose(0, 1).contiguous()
return self._normalize_audio_to_stereo_48k(audio, sr)

[pkrasicki]

There is a PR that was recently merged, which does this, but for Intel GPUs: #746. Maybe it would be helpful to see what the author did.

[shipped-it]

I changed io_audio.py temporarily. It seems something can run, but it's not clear.

@BeForU Were you able to run a repainting task? What do you mean by "it's not clear"?

[fisheryv]

I fixed this problem in PR #1242.
I implemented a dual-path compatibility scheme with torchcodec as the priority and soundfile/ffmpeg as fallbacks, to address the issue where XPU and ROCm do not support torchcodec, or torchcodec incompatibility prevents saving generated audio.