mlenvdoctor 0.1.5

Diagnose & fix ML environments for LLM fine-tuning

ML Environment Doctor

Diagnose ML environment issues, export shareable reports, and generate safe starting fixes.

ML Environment Doctor is a Python CLI for inspecting machine learning environments used for local development, CI jobs, and GPU containers. It focuses on practical readiness checks for PyTorch, TensorFlow/Keras, and JAX/Flax workflows, then turns those results into readable reports and automatable outputs.

Repository: https://github.com/Dheena731/Ml-env-doctor

What It Does Today

• Diagnoses Python runtime compatibility
• Checks NVIDIA driver visibility with nvidia-smi
• Validates PyTorch version, CUDA visibility, and a basic CUDA tensor execution path
• Validates TensorFlow import, GPU enumeration, and a small tensor execution path
• Validates JAX backend visibility, Flax presence, and a small JAX array execution path
• Checks common ML training libraries like transformers, peft, trl, datasets, and accelerate
• Optionally checks GPU memory, disk space, Docker GPU support, and Hugging Face connectivity
• Exports JSON, CSV, and HTML reports
• Generates requirements files, Conda environment files, and Dockerfiles
• Provides a safe fix workflow with planning, dry-run, apply, and verification modes

What It Does Not Claim Yet

• It does not fully repair every environment problem automatically
• It does not maintain a large model registry
• It does not replace framework-native setup guides for CUDA, cuDNN, or TPU drivers
• Its Docker generation is configurable, but still template-based rather than a full build system

Quick Start

pip install mlenvdoctor

mlenvdoctor diagnose
mlenvdoctor diagnose --full
mlenvdoctor diagnose --json -
mlenvdoctor report
mlenvdoctor fix --plan
mlenvdoctor fix --dry-run
mlenvdoctor fix --apply --yes
mlenvdoctor fix --verify
mlenvdoctor dockerize tinyllama --stack minimal

Main Commands

Diagnose

Detailed evidence and export command.

mlenvdoctor diagnose
mlenvdoctor diagnose --full
mlenvdoctor diagnose --json diagnostics.json
mlenvdoctor diagnose --json -
mlenvdoctor diagnose --csv diagnostics.csv --html diagnostics.html

diagnose returns stable exit codes:

• 0: no warnings or critical issues
• 1: warnings present
• 2: critical issues present

Doctor

Action-oriented triage command. It is intentionally different from diagnose:

• doctor tells you what failed, the likely cause, the best next fix, and how to verify it
• diagnose shows the full evidence and supports exports

Compact CI-friendly output:

mlenvdoctor doctor --ci
mlenvdoctor doctor --ci --full
mlenvdoctor doctor --guided

Report

Save a shareable JSON and HTML bundle:

mlenvdoctor report
mlenvdoctor report --quick --output-dir artifacts/mlenvdoctor

Fix

The fix workflow is intentionally explicit:

mlenvdoctor fix --plan
mlenvdoctor fix --dry-run
mlenvdoctor fix --apply
mlenvdoctor fix --apply --yes
mlenvdoctor fix --verify
mlenvdoctor fix --rollback
mlenvdoctor fix --venv --apply --yes
mlenvdoctor fix --conda
mlenvdoctor fix --stack llm-training --dry-run

Current fix behavior:

• Plans file-generation and environment actions from detected issues
• Labels each planned action by risk level
• Supports dry-run mode before changes
• Can create a virtual environment
• Can generate either requirements-mlenvdoctor.txt or environment-mlenvdoctor.yml
• Can install requirements when --apply is used
• Supports explicit verification via mlenvdoctor fix --verify
• Re-runs diagnostics for verification after a successful apply

Stack

mlenvdoctor stack llm-training
mlenvdoctor stack llm-training --output requirements-llm-training.txt

Dockerize

mlenvdoctor dockerize tinyllama
mlenvdoctor dockerize mistral-7b --stack llm-training
mlenvdoctor dockerize --service --base-image nvidia/cuda:12.4.0-runtime-ubuntu22.04
mlenvdoctor dockerize gpt2 --python-version 3.10 -o Dockerfile.gpt2

Current Docker generation supports:

• model-aware defaults for tinyllama, gpt2, and mistral-7b
• stack selection
• base image override
• Python version selection
• training and FastAPI service profiles

MCP

mlenvdoctor mcp serve

MCP now exposes a JSON-lines interface over stdin/stdout.

Core integration tools:

• diagnose
• get_fixes
• doctor_summary
• report_bundle
• diagnose_environment (v1 alias)
• get_fix_plan (v1 alias)
• verify_fix (v1 alias)
• export_report (v1 alias)
• list_tools
• tool_schema
• health

Request format:

{"tool":"list_tools","arguments":{}}

{"tool":"tool_schema","arguments":{"name":"report_bundle"}}

Compatibility policy:

• MCP schemas evolve additively.
• v1 alias names are kept for migration safety.
• Use tool_schema to discover canonical tool names.

Output Schema

JSON exports include:

• issue name, status, severity, and fix command
• check ID and category
• recommendation, likely cause, and verify steps
• confidence
• evidence and metadata blocks
• runtime context (platform, detected backend, NVIDIA tooling visibility)
• summary counts and exit code

That makes the tool more useful for CI parsing, dashboards, and future integrations.

Development

git clone https://github.com/Dheena731/Ml-env-doctor.git
cd Ml-env-doctor
python -m venv .venv
source .venv/bin/activate
pip install -e ".[dev]"
python -m black src tests
python -m ruff check src tests
python -m pytest
python -m mypy src

CI/CD and Release Flow

• CI runs formatting (black --check), lint (ruff), tests (pytest), and package build checks (python -m build, twine check) on every push/PR.
• The matrix tests Linux (3.8-3.11), macOS (3.11), and Windows (3.11).
• A tagged GitHub release triggers the publish job to upload built artifacts to PyPI.
• If CI fails locally vs. GitHub, run these exact commands first:

python -m black src tests
python -m ruff check src tests
python -m pytest

Test Strategy

The repository uses two test layers:

• direct Python and Typer CliRunner tests for most command behavior
• a smaller set of module-level tests for exports, validators, diagnostics, and file generation

This keeps the main test suite independent from whether mlenvdoctor was installed as a shell command in the current environment.

Repository Notes

• CONTRIBUTING.md covers local setup and contribution flow
• IMPROVEMENTS.md tracks active improvement themes
• IMPROVEMENTS_ROADMAP.md outlines future milestones
• docs/PROJECT_OVERVIEW.md explains the current project architecture and product shape
• docs/FAILURE_TAXONOMY.md defines prioritized mismatch codes used by diagnostics and MCP
• docs/KILLER_PRODUCT_PLAN.md is the strategic roadmap for turning the project into a category-leading product
• docs/PHASED_EXECUTION_PLAN.md breaks the strategy into execution phases and current build focus
• docs/CASE_STUDY_TEMPLATE.md helps publish real failure-to-fix stories for user onboarding
• docker/README.md documents Docker-specific workflows

License

This project is licensed under the MIT License. See LICENSE.