environments ai ml automl dnn text gpu

ai-ml-automl-dnn-text-gpu

Overview

An environment used by Azure ML AutoML for training models.

Version: 8

Tags

OS : Ubuntu20.04 Training Preview OpenMpi : 4.1.0 Python : 3.9

View in Studio: https://ml.azure.com/registries/azureml/environments/ai-ml-automl-dnn-text-gpu/version/8

Docker image: mcr.microsoft.com/azureml/curated/ai-ml-automl-dnn-text-gpu:8

Docker build context

Dockerfile

FROM mcr.microsoft.com/aifx/acpt/stable-ubuntu2004-cu121-py310-torch222:biweekly.202410.2

ENV AZUREML_CONDA_ENVIRONMENT_PATH /azureml-envs/azureml-automl-dnn-text-gpu
# Prepend path to AzureML conda environment
ENV PATH $AZUREML_CONDA_ENVIRONMENT_PATH/bin:$PATH

COPY --from=mcr.microsoft.com/azureml/mlflow-ubuntu20.04-py38-cpu-inference:20230306.v3 /var/mlflow_resources/mlflow_score_script.py /var/mlflow_resources/mlflow_score_script.py

ENV MLFLOW_MODEL_FOLDER="mlflow-model"
# ENV AML_APP_ROOT="/var/mlflow_resources"
# ENV AZUREML_ENTRY_SCRIPT="mlflow_score_script.py"

ENV ENABLE_METADATA=true

# begin conda create
# Create conda environment
RUN conda create -p $AZUREML_CONDA_ENVIRONMENT_PATH \
    python=3.9 \
    # begin conda dependencies
    pip=22.1.2 \
    numpy~=1.23.5 \
    scikit-learn=1.5.1 \
    pandas~=1.5.3 \
    setuptools=72.1.0 \
    wheel=0.44.0 \
    scipy=1.10.1 \
    pybind11=2.10.1 \
    # end conda dependencies
    -c conda-forge -c anaconda

# Ensure additional conda and pip install commands apply to our conda env of interest.
SHELL ["conda", "run", "-p", "$AZUREML_CONDA_ENVIRONMENT_PATH", "/bin/bash", "-c"]

# begin pip install
# Install pip dependencies
# Here, we pin sentencepiece since 0.1.98 breaks training. Earlier versions of horovod contain a sev2 vulnerability,
# and earlier versions of tokenizers cause log spam with transformers==4.16.0.
RUN pip install \
                # begin pypi dependencies
                azureml-core==1.58.0 \
                azureml-mlflow==1.58.0.post2 \
                azureml-automl-core==1.58.0 \
                azureml-automl-dnn-nlp==1.58.0 \
                azureml-automl-runtime==1.58.0 \
                azureml-train-automl-client==1.58.0 \
                azureml-train-automl-runtime==1.58.0 \
                azureml-defaults==1.58.0 \
                'azure-identity>=1.16.1' \
                'sentencepiece==0.1.97' \
                'cryptography>=42.0.5'\
                'urllib3>=1.26.18' \
                'requests>=2.31.0' \
                'certifi>=2023.07.22' 
                # end pypi dependencies

# Horovod incompatible with torch>2.0 https://github.com/horovod/horovod/issues/3996
RUN HOROVOD_WITH_PYTORCH=1 pip install --no-cache-dir git+https://github.com/horovod/horovod@3a31d933a13c7c885b8a673f4172b17914ad334d

# Separate updates for fixing vulnerabilities.
# Doing it  separately from pip install above to avoid conflict with other packages
# We should aim for this list to be empty with new and patched releases
# by fixing dependencies in the base packages
RUN pip list && \
    pip install pyarrow==14.0.2 \
                'transformers[sentencepiece,torch]==4.36.2' \
                aiohttp==3.10.2

# end pip install

# Force new version