Local AI Development Environment Setup

Setting up an AI development environment on a local machine

A properly configured local environment is the foundation of an AI developer's productivity. Key requirements include: isolation of dependencies between projects, convenient GPU support (if available), integration with cloud resources for model training, and code and data versioning.

Conda as a basis for isolation

# Установка Miniconda
wget https://repo.anaconda.com/miniconda/Miniconda3-latest-Linux-x86_64.sh
bash Miniconda3-latest-Linux-x86_64.sh

# Базовое окружение для AI-разработки
conda create -n ai-dev python=3.11
conda activate ai-dev

# Core ML stack
pip install torch torchvision --index-url https://download.pytorch.org/whl/cu121
pip install transformers datasets peft accelerate
pip install scikit-learn xgboost lightgbm catboost
pip install pandas numpy scipy matplotlib seaborn plotly
pip install mlflow dvc[s3] great_expectations
pip install jupyter jupyterlab ipywidgets
pip install pytest black isort mypy pre-commit

VS Code Configuration for AI Development

.vscode/settings.json:

{
  "python.defaultInterpreterPath": "~/miniconda3/envs/ai-dev/bin/python",
  "python.formatting.provider": "black",
  "editor.formatOnSave": true,
  "jupyter.notebookFileRoot": "${workspaceFolder}",
  "python.testing.pytestEnabled": true,
  "files.exclude": {
    "**/__pycache__": true,
    "**/*.pyc": true,
    ".dvc": false
  }
}

GPU profiling locally

# PyTorch Profiler для анализа GPU-операций
with torch.profiler.profile(
    activities=[
        torch.profiler.ProfilerActivity.CPU,
        torch.profiler.ProfilerActivity.CUDA,
    ],
    record_shapes=True,
    profile_memory=True,
) as prof:
    output = model(input_tensor)

print(prof.key_averages().table(sort_by="cuda_time_total", row_limit=10))
# Экспорт в Chrome trace
prof.export_chrome_trace("trace.json")

Cloud development for resource-intensive tasks

For local development without a powerful GPU: VS Code Remote SSH connects to a cloud GPU instance. Code is stored locally (Git), data is in S3/GCS, and computation is in the cloud. This is cheaper than maintaining a local GPU and more convenient than working in the browser-based Jupyter.

Key: ~/.ssh/config with aliases for GPU instances + automatic mounting of the remote filesystem via SSHFS for transparent work with remote files.