Docker support

[tuxx]

Hiya,

I got a working Dockerfile that builds the container with GPU support.
Also added github actions to publish the container when a new tag has been pushed.

The resulting docker container URL would be: ghcr.io/haoheliu/AudioLDM/audioldm:latest

What needs to be done

What i can do

• Update README
  • How to use the container
  • Explaining that you need nvidia-container-toolkit for running this with docker
• Fix the build
  • Add webapp support to container: https://github.com/tuxx/AudioLDM/actions/runs/13754503222

What repo maintainer should do

• Make sure your repository has "Read and write permissions" enabled for workflows (in repository settings under Actions → General) -- (In my forked repo this setting was already enabled)
• Tag releases like v1.0.0 to build new containers automatically

Running it locally

docker build -t audioldm:gpu .

docker run --gpus all -v $(pwd)/output:/app/output audioldm:gpu --text "spaceship shooting 1 bullet" -dur 2.5 --save_path /app/output

Output (Removed some python futurewarnings, etc for readability):

Load AudioLDM: %s audioldm-m-full
Downloading the main structure of audioldm-m-full into /root/.cache/audioldm
Weights downloaded in: /root/.cache/audioldm/audioldm-m-full.ckpt Size: 4571683377
100% |########################################################################|
DiffusionWrapper has 415.95 M params.
  WeightNorm.apply(module, name, dim)
  warnings.warn(
  fft_window = librosa.util.pad_center(fft_window, n_fft)
  return _VF.meshgrid(tensors, **kwargs)  # type: ignore[attr-defined]
  return torch.load(checkpoint_file, map_location="cpu")
Some weights of the model checkpoint at roberta-base were not used when initializing RobertaModel: ['lm_head.layer_norm.weight', 'lm_head.layer_norm.bias', 'lm_head.bias', 'lm_head.dense.bias', 'lm_head.dense.weight', 'lm_head.decoder.weight']
- This IS expected if you are initializing RobertaModel from the checkpoint of a model trained on another task or with another architecture (e.g. initializing a BertForSequenceClassification model from a BertForPreTraining model).
- This IS NOT expected if you are initializing RobertaModel from the checkpoint of a model that you expect to be exactly identical (initializing a BertForSequenceClassification model from a BertForSequenceClassification model).
  checkpoint = torch.load(resume_from_checkpoint, map_location=device)
Generate audio using text spaceship shooting 1 bullet
DDIM Sampler: 100%|██████████| 200/200 [00:11<00:00, 17.64it/s]
  warnings.warn(                                                                                                                           
  Save audio to /app/output/generation/10_03_2025_00_17_42_spaceship shooting 1 bullet_0.wav

[tuxx]

Examples with the container from my fork:

Running in CLI mode:
docker run --gpus all -v $(pwd)/output:/app/output ghcr.io/tuxx/audioldm/audioldm:latest --text "spaceship shooting 1 bullet" -dur 2.5 --save_path /app/output

Running webapp:
docker run --gpus all -p 7860:7860 ghcr.io/tuxx/audioldm/audioldm:latest webapp

[tuxx]

Waiting on build to fix this:

$ docker run --gpus all -v $(pwd)/output:/app/output ghcr.io/tuxx/audioldm/audioldm:latest --text "spaceship shooting 1 bullet" -dur 2.5 --save_path /app/output

Traceback (most recent call last):
  File "/usr/local/bin/audioldm", line 3, in <module>
    from audioldm import text_to_audio, style_transfer, build_model, save_wave, get_time, round_up_duration, get_duration
  File "/usr/local/lib/python3.8/dist-packages/audioldm/__init__.py", line 1, in <module>
    from .ldm import LatentDiffusion
  File "/usr/local/lib/python3.8/dist-packages/audioldm/ldm.py", line 6, in <module>
    from audioldm.utils import default, instantiate_from_config, save_wave
  File "/usr/local/lib/python3.8/dist-packages/audioldm/utils.py", line 6, in <module>
    import soundfile as sf
ModuleNotFoundError: No module named 'soundfile'
exit status 1

[tuxx]

Should work now :)

[tuxx]

Got one more issue, when portforwarding the webapp port, i can not access it outside the container. Maybe i should run the app.py on 0.0.0.0 or something? Not sure yet.

[ElanHasson]

Hi @tuxx, I tried the docker image build locally and get a "This page isn’t working right now
127.0.0.1 didn’t send any data.
ERR_EMPTY_RESPONSE" error.

docker run -it --gpus all -p 7860:7860 docker.io/library/audioldm:gpu webapp

[tuxx]

Hi @tuxx, I tried the docker image build locally and get a "This page isn’t working right now 127.0.0.1 didn’t send any data. ERR_EMPTY_RESPONSE" error.

docker run -it --gpus all -p 7860:7860 docker.io/library/audioldm:gpu webapp

I think the app should be running on 0.0.0.0 to make this work. Havent been working on this tho 😅