Refactor: Introduce Modular Architecture for llama-model.cpp

[devin-ai-integration]

Make sure to read the contributing guidelines before submitting a PR

Summary

This PR introduces the foundational infrastructure for refactoring src/llama-model.cpp (20,498 lines) into a modular architecture. This is Phase 1 only - it creates the scaffolding and patterns but does not yet migrate any existing code.

Link to Devin run: https://app.devin.ai/sessions/a329f05a4cdc4f2eb254bcb7ee3fb2ac
Requested by: Jake Cosme (@jakexcosme)

Changes

New Infrastructure

• Directory structure under src/llama-model/:

  • architectures/ - Will contain one file per architecture (currently has LLAMA example)
  • interfaces/ - Interface definitions (IModelLoader, IGraphBuilder, ArchitectureBuilder)
  • registry/ - Architecture registration system with singleton pattern
  • base/, loading/, memory/ - Empty directories for future modules

• Registry Pattern: ArchitectureRegistry with factory-based lazy instantiation

• Example Implementation: LlamaArchitectureBuilder that demonstrates the pattern (currently just delegates to existing code)

• Documentation: Comprehensive 518-line REFACTORING-GUIDE.md with examples and migration strategy

Build System

• Updated src/CMakeLists.txt to use file(GLOB ...) for automatic inclusion of architecture files
• Updated CODEOWNERS for new directory structure

Backwards Compatibility

✅ All existing APIs unchanged - No modifications to llama_model interface
✅ All 37 tests pass (100% pass rate)
✅ No functional changes - New code is not yet integrated into existing flow

Important Notes for Reviewers

This PR Does NOT:

• Actually refactor any of the 89+ architecture builders from llama-model.cpp
• Integrate the new registry system into the existing code paths
• Change any existing functionality
• Remove any existing code

This PR DOES:

• Create the infrastructure and patterns for future refactoring
• Demonstrate the intended approach with one example (LLAMA)
• Provide comprehensive documentation for contributors
• Ensure the build system can handle the new structure

Key Review Points

1. Architecture Design: Are the interfaces (ArchitectureBuilder, ArchitectureRegistry) well-designed for the intended refactoring?

2. Documentation Quality: Is docs/development/REFACTORING-GUIDE.md clear and comprehensive enough for contributors to follow?

3. Build System: The use of file(GLOB ...) is convenient but requires manual CMake re-runs when files are added. Is this acceptable?

4. Dead Code: The new registry and LLAMA builder are not yet called from anywhere. Is it okay to merge infrastructure before integration?

5. Empty Directories: Several directories (base/, loading/, memory/) are created but empty. Should these be added in later PRs instead?

6. Macro Safety: The REGISTER_ARCHITECTURE macro uses __LINE__ for uniqueness. Review for potential edge cases.

Testing

# Build succeeds
cmake -B build -DLLAMA_FATAL_WARNINGS=ON
cmake --build build --config Release -j $(nproc)

# All tests pass
cd build && ctest -L main --timeout 300
# Result: 100% tests passed, 0 tests failed out of 37

Next Steps (Phase 2)

After this PR is merged, Phase 2 will:

1. Integrate the registry into llama_model::build_graph()
2. Migrate architectures one at a time from the monolithic file
3. Add unit tests for each extracted architecture
4. Gradually deprecate the old switch-statement approach

Checklist

• Directory structure created
• Interfaces defined with documentation
• Registry pattern implemented
• Example architecture (LLAMA) created
• Build system updated
• CODEOWNERS updated
• Comprehensive documentation written
• All existing tests pass
• No breaking changes to public API
• Unit tests for registry (in gitignored directory, will add in Phase 2)
• Integration with existing code (Phase 2)
• Migration of remaining architectures (Phase 2+)

[devin-ai-integration]

🤖 Devin AI Engineer

I'll be helping with this pull request! Here's what you should know:

✅ I will automatically:

• Address comments on this PR. Add '(aside)' to your comment to have me ignore it.
• Look at CI failures and help fix them

Note: I can only respond to comments from users who have write access to this repository.

⚙️ Control Options:

• Disable automatic comment and CI monitoring