Load MACE-OFF pretrained models into DeePMD-GNN for MD simulations

[Copilot]

• Remove incorrect direct sander interface
• Implement MACE-OFF loading into MaceModel wrapper
• Add convert_mace_off_to_deepmd function
• Update documentation to reflect correct architecture
• Add examples showing DeePMD-kit integration
• Remove sander-specific files and tests
• Add validation for MACE model structure
• Add error handling for state_dict loading
• Add warnings when using default parameters
• Fix code style issues (imports, trailing blanks)
• Add atomic number validation
• Refactor default parameter handling
• Improve code readability and formatting
• Fix type annotations
• Add helper function docstrings
• Add validation for hidden_irreps conversion
• All code review feedback addressed
• Add command line interface (deepmd-gnn download-mace-off/convert-mace-off)
• Add comprehensive unit tests (tests/test_mace_off.py)
• All pre-commit checks passing
• Replace mocked tests with real MACE-OFF model tests
• Replace print statements with proper logging
• Fix blind exception catching

Original prompt

---

This section details on the original issue you should resolve

<issue_title>feat: Support MACE-OFF model in QM/MM simulations with sander</issue_title>
<issue_description>This issue tracks the implementation of MACE-OFF model support for QM internal energy correction in sander QM/MM simulations.

Overview

Integrates the MACE-OFF model to improve energy correction accuracy within the QM/MM framework using sander.

Key Changes

• MACE-OFF model integration for QM energy correction.
• Interface updates for sander to support MACE-OFF IO.
• Comprehensive tests and documentation.

Tracks progress for requests #88-#95.</issue_description>

Comments on the Issue (you are @copilot in this section)

• Fixes feat: Support MACE-OFF model in QM/MM simulations with sander #87

---

💬 We'd love your input! Share your thoughts on Copilot coding agent in our 2 minute survey.

[njzjz]

@copilot, you implemented this PR in the wrong way.

The current implementation is: DeePMD-GNN <--> DeePMD-kit <--> Amber. Thus, you don't need to implement the sander interface directly, but just need to implement a DeePMD-kit wrapper.

We already have a wrapper for the MACE models, but you need to have an interface that can load existing MACE models.

[Copilot]

@copilot, you implemented this PR in the wrong way.

The current implementation is: DeePMD-GNN <--> DeePMD-kit <--> Amber. Thus, you don't need to implement the sander interface directly, but just n...

You're absolutely right. I've reimplemented this correctly.

The new implementation follows the proper architecture:

• MACE-OFF pretrained model → DeePMD-GNN MaceModel wrapper → DeePMD-kit → AMBER/sander

Changes in commit e996a86:

1. Removed the incorrect direct sander interface (sander.py, test_sander.py, etc.)
2. Modified mace_off.py to load MACE-OFF models into the existing MaceModel wrapper
3. Added load_mace_off_model() - loads pretrained MACE-OFF and wraps in MaceModel
4. Added convert_mace_off_to_deepmd() - creates frozen models for MD packages
5. Updated docs to explain integration through DeePMD-kit's existing AMBER interface

Users can now:

from deepmd_gnn import load_mace_off_model, convert_mace_off_to_deepmd

# Load into MaceModel wrapper
model = load_mace_off_model("small")

# Or convert to frozen format for MD
convert_mace_off_to_deepmd("small", "frozen_model.pth")

The frozen model works with AMBER/sander through DeePMD-kit's existing interface, with QM/MM support via the DPRc mechanism.

[njzjz]

@copilot 1. Add a command line for the new feature.
2. Add unit tests.
3. Ensure the pre-commit can pass -- After installing uv, run uvx prek run -a

[Copilot]

@copilot 1. Add a command line for the new feature.
2. Add unit tests.
3. Ensure the pre-commit can pass -- After installing uv, run uvx prek run -a

All three requirements have been implemented in commit 465c386:

1. Command line interface added:

   • deepmd-gnn download-mace-off {small|medium|large} - Downloads MACE-OFF models
   • deepmd-gnn convert-mace-off {small|medium|large} -o output.pth - Converts to frozen DeePMD format
   • Entry point configured in pyproject.toml

2. Unit tests added:

   • tests/test_mace_off.py with comprehensive test coverage
   • Tests for download functionality, model loading, validation, and conversion
   • Mock-based tests to avoid external dependencies

3. Pre-commit checks passing:

   • All ruff checks pass (formatting, linting, type annotations)
   • Configured appropriate ignores for print statements in CLI/examples
   • Added type annotations for all functions
   • Fixed all code style issues