Forum Discussion

Pitchbendphil's avatar
Pitchbendphil
Honored Guest
27 days ago

AttributeError: MusicGen object has no attribute 'from_pretrained' despite correct Audiocraft/Torch

Hello everyone,

I'm encountering a very persistent issue while trying to get audiocraft.models.MusicGen to work on a Hugging Face Space equipped with a T4 GPU. Despite (I believe) having configured all dependencies and the environment correctly, every attempt to load the model using MusicGen.from_pretrained("facebook/musicgen-small") results in the following error in my Streamlit app:

Generated code

A critical error occurred while loading the AI model.
Error details: AttributeError: type object 'MusicGen' has no attribute 'from_pretrained'
Note: The error 'has no attribute from_pretrained' often indicates a problem with the Audiocraft/Torch installation or version.

This error occurs even though the MusicGen class itself can be imported successfully (from audiocraft.models import MusicGen does not raise an error). It's only the call to .from_pretrained() that fails. This setup was previously working on a different platform (Google Cloud VM).

I've already tried numerous approaches to isolate and fix the problem, including:

  • Using a detailed Dockerfile with an isolated Python Virtual Environment (venv).

  • Precisely pinning all relevant library versions in requirements.txt.

  • Ensuring correct user permissions and paths within the Docker container.

  • Resolving PermissionError issues related to Streamlit and Hugging Face cache directories.

  • Testing various versions of torch, torchaudio, torchvision, audiocraft, numpy, accelerate, and librosa.

The current configuration, which still leads to this error, is as follows:

 

The build process completes without errors. The pip install commands all report success. The container starts, Streamlit is accessible, but the call to MusicGen.from_pretrained() results in the aforementioned AttributeError.

I also tested an earlier version of the code using importlib to load the class dynamically, which resulted in the exact same AttributeError. This suggests that the MusicGen object available to Python is somehow "corrupted" or incomplete.

I would be grateful for any help or ideas as to what might be causing this or what further debugging steps I could take. It feels like something very subtle in the build process or the runtime environment of Hugging Face Spaces is preventing the correct initialization of audiocraft.

Thank you in advance!

2 Replies

  • Pitchbendphil's avatar
    Pitchbendphil
    Honored Guest
    27 days ago

    1. Dockerfile:

    Generated dockerfile

    # 1. Base Image
FROM nvidia/cuda:12.1.1-cudnn8-runtime-ubuntu22.04

# 2. Environment Variables
ENV DEBIAN_FRONTEND=noninteractive
ENV PYTHONUNBUFFERED=1
ENV STREAMLIT_HOME="/home/user/.streamlit"
ENV HF_HOME="/home/user/.cache/huggingface"
ENV TRANSFORMERS_CACHE="/home/user/.cache/huggingface/hub"
ENV HF_DATASETS_CACHE="/home/user/.cache/huggingface/datasets"

# 3. Install system dependencies as root
RUN apt-get update && \
    apt-get install -y --no-install-recommends \
    python3.10-venv \
    python3-pip \
    git \
    git-lfs \
    cmake \
    ffmpeg \
    && rm -rf /var/lib/apt/lists/*

# 4. Create an isolated virtual environment (venv)
RUN python3 -m venv /opt/venv
ENV PATH="/opt/venv/bin:${PATH}"

# 5. Install Python dependencies INTO the venv
WORKDIR /app

# 5.1. Upgrade pip
RUN pip install --upgrade pip

# 5.2. FIRST copy requirements.txt and install critical packages
COPY requirements.txt .
# Explicit installation of critical packages to ensure order
RUN pip install --no-cache-dir torch==2.0.1 torchvision==0.15.2 torchaudio==2.0.2 --extra-index-url https://download.pytorch.org/whl/cu118
RUN pip install --no-cache-dir xformers==0.0.20
# Install the rest of the packages from requirements.txt
RUN pip install --no-cache-dir -r requirements.txt 

# 5.3. THEN copy the rest of the app code and assets
COPY . . 

# 6. Create user and set permissions for /app AND cache directories
RUN useradd -m -u 1000 user && \
    mkdir -p /home/user/.streamlit /home/user/.cache/huggingface/hub /home/user/.cache/huggingface/datasets && \
    chown -R user:user /app /home/user/.streamlit /home/user/.cache
USER user

# 7. Expose port and define the final start command
EXPOSE 7860
CMD ["streamlit", "run", "app.py", "--server.port=7860"]
     
    content_copydownload
    Use code with caution.Dockerfile
     

    2. requirements.txt:

    Generated txt

    --extra-index-url https://download.pytorch.org/whl/cu118
torch==2.0.1
torchaudio==2.0.2
torchvision==0.15.2
numpy==1.24.4
audiocraft==1.0.0
xformers==0.0.20
transformers>=4.31.0,<4.32.0 
diffusers==0.20.0
accelerate>=0.20.3 
streamlit
scipy
pydub
librosa==0.9.2 
xmltodict
soundfile
     
    content_copydownload
    Use code with caution.Txt
     

    3. app.py (minimal version to reproduce the error):

    Generated python

    import streamlit as st

st.title("Audiocraft/MusicGen Import & Load Test")

try:
    st.write("Attempting to import audiocraft.models.MusicGen...")
    from audiocraft.models import MusicGen
    st.success("Class 'MusicGen' imported successfully!")
    
    st.write("Checking if 'from_pretrained' method exists on the MusicGen object...")
    if hasattr(MusicGen, 'from_pretrained'):
        st.success("Method 'MusicGen.from_pretrained()' appears to exist on the class object.")
        
        st.write("Now attempting to load the model using MusicGen.from_pretrained('facebook/musicgen-small')...")
        # The following call triggers the AttributeError
        model = MusicGen.from_pretrained("facebook/musicgen-small")
        st.success("MODEL LOADED AND INITIALIZED SUCCESSFULLY! THE ISSUE IS RESOLVED!")
        st.balloons()
    else:
        st.error("ERROR: Method 'MusicGen.from_pretrained()' does NOT exist on the imported MusicGen class object.")
        
except ImportError as ie:
    st.error(f"ImportError occurred: {type(ie).__name__}: {ie}")
except AttributeError as ae:
    # This is the error currently being triggered
    st.error(f"AttributeError occurred: {type(ae).__name__}: {ae}")
except Exception as e:
    st.error(f"An unexpected error occurred: {type(e).__name__}: {e}")

st.markdown("---")
st.subheader("Environment Information (as known):")
st.json({
    "Platform": "Hugging Face Spaces",
    "Hardware": "T4 GPU (small)",
    "Dockerfile Base Image": "nvidia/cuda:12.1.1-cudnn8-runtime-ubuntu22.04",
    "Python Version in venv": "3.10 (as provided by base image and used for venv)",
    "Key Library Versions (from requirements.txt)": {
        "torch": "2.0.1 (cu118)",
        "audiocraft": "1.0.0",
        "numpy": "1.24.4"
    }
})
     
    content_copydownload
    Use code with caution.Python
     

    The build process completes without errors. The pip install commands all report success. The container starts, Streamlit is accessible, but the call to MusicGen.from_pretrained() results in the aforementioned AttributeError.

    • Pitchbendphil's avatar
      Pitchbendphil
      Honored Guest
      27 days ago

      this code above is one of many versions that did not work. I hope anyone can help this this! Thanks in advance