Registry refactor

.gitignore

@@ -0,0 +1,177 @@
+# Byte-compiled / optimized / DLL files
+__pycache__/
+*.py[cod]
+*$py.class
+
+# C extensions
+*.so
+
+# Distribution / packaging
+.Python
+build/
+develop-eggs/
+dist/
+downloads/
+eggs/
+.eggs/
+lib/
+lib64/
+parts/
+sdist/
+var/
+wheels/
+share/python-wheels/
+*.egg-info/
+.installed.cfg
+*.egg
+MANIFEST
+
+# PyInstaller
+#  Usually these files are written by a python script from a template
+#  before PyInstaller builds the exe, so as to inject date/other infos into it.
+*.manifest
+*.spec
+
+# Installer logs
+pip-log.txt
+pip-delete-this-directory.txt
+
+# Unit test / coverage reports
+htmlcov/
+.tox/
+.nox/
+.coverage
+.coverage.*
+.cache
+nosetests.xml
+coverage.xml
+*.cover
+*.py,cover
+.hypothesis/
+.pytest_cache/
+cover/
+
+# Translations
+*.mo
+*.pot
+
+# Django stuff:
+*.log
+local_settings.py
+db.sqlite3
+db.sqlite3-journal
+
+# Flask stuff:
+instance/
+.webassets-cache
+
+# Scrapy stuff:
+.scrapy
+
+# Sphinx documentation
+docs/_build/
+
+# PyBuilder
+.pybuilder/
+target/
+
+# Jupyter Notebook
+.ipynb_checkpoints
+
+# IPython
+profile_default/
+ipython_config.py
+
+# pyenv
+#   For a library or package, you might want to ignore these files since the code is
+#   intended to run in multiple environments; otherwise, check them in:
+# .python-version
+
+# pipenv
+#   According to pypa/pipenv#598, it is recommended to include Pipfile.lock in version control.
+#   However, in case of collaboration, if having platform-specific dependencies or dependencies
+#   having no cross-platform support, pipenv may install dependencies that don't work, or not
+#   install all needed dependencies.
+#Pipfile.lock
+
+# UV
+#   Similar to Pipfile.lock, it is generally recommended to include uv.lock in version control.
+#   This is especially recommended for binary packages to ensure reproducibility, and is more
+#   commonly ignored for libraries.
+#uv.lock
+
+# poetry
+#   Similar to Pipfile.lock, it is generally recommended to include poetry.lock in version control.
+#   This is especially recommended for binary packages to ensure reproducibility, and is more
+#   commonly ignored for libraries.
+#   https://python-poetry.org/docs/basic-usage/#commit-your-poetrylock-file-to-version-control
+#poetry.lock
+
+# pdm
+#   Similar to Pipfile.lock, it is generally recommended to include pdm.lock in version control.
+#pdm.lock
+#   pdm stores project-wide configurations in .pdm.toml, but it is recommended to not include it
+#   in version control.
+#   https://pdm.fming.dev/latest/usage/project/#working-with-version-control
+.pdm.toml
+.pdm-python
+.pdm-build/
+
+# PEP 582; used by e.g. github.com/David-OConnor/pyflow and github.com/pdm-project/pdm
+__pypackages__/
+
+# Celery stuff
+celerybeat-schedule
+celerybeat.pid
+
+# SageMath parsed files
+*.sage.py
+
+# Environments
+.env
+.venv
+env/
+venv/
+ENV/
+env.bak/
+venv.bak/
+
+# Spyder project settings
+.spyderproject
+.spyproject
+
+# Rope project settings
+.ropeproject
+
+# mkdocs documentation
+/site
+
+# mypy
+.mypy_cache/
+.dmypy.json
+dmypy.json
+
+# Pyre type checker
+.pyre/
+
+# pytype static type analyzer
+.pytype/
+
+# Cython debug symbols
+cython_debug/
+
+# PyCharm
+#  JetBrains specific template is maintained in a separate JetBrains.gitignore that can
+#  be found at https://github.com/github/gitignore/blob/main/Global/JetBrains.gitignore
+#  and can be added to the global gitignore or merged into this file.  For a more nuclear
+#  option (not recommended) you can uncomment the following to ignore the entire idea folder.
+#.idea/
+
+# Ruff stuff:
+.ruff_cache/
+
+# PyPI configuration file
+.pypirc
+
+# unsloth compiled cache
+unsloth_compiled_cache

pyproject.toml

@@ -32,6 +32,10 @@ include-package-data = false
 exclude = ["images*", "tests*"]
 
 [project.optional-dependencies]
+dev = [
+    "pytest",
+]
+
 triton = [
     "triton-windows ; platform_system == 'Windows'",
 ]

tests/test_model_registry.py

@@ -0,0 +1,91 @@
+"""
+
+Test model registration methods
+Checks that model registration methods work for respective models as well as all models
+The check is performed
+- by registering the models
+- checking that the instantiated models can be found on huggingface hub by querying for the model id
+
+"""
+
+from dataclasses import dataclass
+
+import pytest
+from huggingface_hub import ModelInfo as HfModelInfo
+
+from unsloth.registry import register_models, search_models
+from unsloth.registry._deepseek import register_deepseek_models
+from unsloth.registry._gemma import register_gemma_models
+from unsloth.registry._llama import register_llama_models
+from unsloth.registry._mistral import register_mistral_models
+from unsloth.registry._phi import register_phi_models
+from unsloth.registry._qwen import register_qwen_models
+from unsloth.registry.registry import MODEL_REGISTRY, QUANT_TAG_MAP, QuantType
+from unsloth.utils.hf_hub import get_model_info
+
+MODEL_NAMES = [
+    "llama",
+    "qwen",
+    "mistral",
+    "phi",
+    "gemma",
+    "deepseek",
+]
+MODEL_REGISTRATION_METHODS = [
+    register_llama_models,
+    register_qwen_models,
+    register_mistral_models,
+    register_phi_models,
+    register_gemma_models,
+    register_deepseek_models,
+]
+
+
+@dataclass
+class ModelTestParam:
+    name: str
+    register_models: callable
+
+
+def _test_model_uploaded(model_ids: list[str]):
+    missing_models = []
+    for _id in model_ids:
+        model_info: HfModelInfo = get_model_info(_id)
+        if not model_info:
+            missing_models.append(_id)
+
+    return missing_models
+
+
+TestParams = [
+    ModelTestParam(name, models)
+    for name, models in zip(MODEL_NAMES, MODEL_REGISTRATION_METHODS)
+]
+
+
+# Test that model registration methods register respective models
+@pytest.mark.parametrize("model_test_param", TestParams, ids=lambda param: param.name)
+def test_model_registration(model_test_param: ModelTestParam):
+    MODEL_REGISTRY.clear()
+    registration_method = model_test_param.register_models
+    registration_method()
+    registered_models = MODEL_REGISTRY.keys()
+    missing_models = _test_model_uploaded(registered_models)
+    assert not missing_models, (
+        f"{model_test_param.name} missing following models: {missing_models}"
+    )
+
+
+def test_all_model_registration():
+    register_models()
+    registered_models = MODEL_REGISTRY.keys()
+    missing_models = _test_model_uploaded(registered_models)
+    assert not missing_models, f"Missing following models: {missing_models}"
+
+def test_quant_type():
+    # Test that the quant_type is correctly set for model paths
+    # NOTE: for models registered under org="unsloth" with QuantType.NONE aliases QuantType.UNSLOTH
+    dynamic_quant_models = search_models(quant_types=[QuantType.UNSLOTH])
+    assert all(m.quant_type == QuantType.UNSLOTH for m in dynamic_quant_models)
+    quant_tag = QUANT_TAG_MAP[QuantType.UNSLOTH]
+    assert all(quant_tag in m.model_path for m in dynamic_quant_models)

unsloth/registry/REGISTRY.md

@@ -0,0 +1,110 @@
+## Model Registry
+
+### Structure
+```
+unsloth
+    -registry
+        __init__.py
+        registry.py
+        _llama.py
+        _mistral.py
+        _phi.py
+        ...
+```
+
+Each model is registered in a separate file within the `registry` module (e.g. `registry/_llama.py`).
+
+Within each model registration file, a high-level `ModelMeta` is created for each model version, with the following structure:
+```python
+@dataclass
+class ModelMeta:
+    org: str
+    base_name: str
+    model_version: str
+    model_info_cls: type[ModelInfo]
+    model_sizes: list[str] = field(default_factory=list)
+    instruct_tags: list[str] = field(default_factory=list)
+    quant_types: list[QuantType] | dict[str, list[QuantType]] = field(default_factory=list)
+    is_multimodal: bool = False
+```
+
+Each model then instantiates a global `ModelMeta` for its specific model version, defining how the model path (e.g. `unsloth/Llama-3.1-8B-Instruct`) is constructed since each model type has a different naming convention.
+```python
+LlamaMeta_3_1 = ModelMeta(
+    org="meta-llama",
+    base_name="Llama",
+    instruct_tags=[None, "Instruct"],
+    model_version="3.1",
+    model_sizes=["8"],
+    model_info_cls=LlamaModelInfo,
+    is_multimodal=False,
+    quant_types=[QuantType.NONE, QuantType.BNB, QuantType.UNSLOTH],
+)
+```
+
+`LlamaModelInfo` is a subclass of `ModelInfo` that defines the model path for each model size and quant type.
+```python
+class LlamaModelInfo(ModelInfo):
+    @classmethod
+    def construct_model_name(cls, base_name, version, size, quant_type, instruct_tag):
+        key = f"{base_name}-{version}-{size}B"
+        return super().construct_model_name(base_name, version, size, quant_type, instruct_tag, key)
+```
+
+Once these constructs are defined, the model is registered by writing a register_xx_models function.
+```python
+def register_llama_3_1_models(include_original_model: bool = False):
+    global _IS_LLAMA_3_1_REGISTERED
+    if _IS_LLAMA_3_1_REGISTERED:
+        return
+    _register_models(LlamaMeta_3_1, include_original_model=include_original_model)
+    _IS_LLAMA_3_1_REGISTERED = True
+```
+
+`_register_models` is a helper function that registers the model with the registry.  The global `_IS_XX_REGISTERED` is used to prevent duplicate registration.
+
+Once a model is registered, registry.registry.MODEL_REGISTRY is updated with the model info and can be searched with `registry.search_models`.
+
+### Tests
+
+The `tests/test_model_registry.py` file contains tests for the model registry.
+
+Also, each model registration file is an executable module that checks that all registered models are available on `huggingface_hub`.
+```python
+python unsloth.registry._llama.py
+```
+
+Prints the following (abridged) output:
+```bash
+✓ unsloth/Llama-3.1-8B
+✓ unsloth/Llama-3.1-8B-bnb-4bit
+✓ unsloth/Llama-3.1-8B-unsloth-bnb-4bit
+✓ meta-llama/Llama-3.1-8B
+✓ unsloth/Llama-3.1-8B-Instruct
+✓ unsloth/Llama-3.1-8B-Instruct-bnb-4bit
+✓ unsloth/Llama-3.1-8B-Instruct-unsloth-bnb-4bit
+✓ meta-llama/Llama-3.1-8B-Instruct
+✓ unsloth/Llama-3.2-1B
+✓ unsloth/Llama-3.2-1B-bnb-4bit
+✓ unsloth/Llama-3.2-1B-unsloth-bnb-4bit
+✓ meta-llama/Llama-3.2-1B
+...
+```
+
+### TODO
+- Model Collections
+    - [x] Gemma3
+    - [ ] Llama3.1
+    - [x] Llama3.2
+    - [x] MistralSmall
+    - [x] Qwen2.5
+    - [x] Qwen2.5-VL
+    - [ ] Qwen2.5 Coder
+    - [x] QwenQwQ-32B
+    - [x] Deepseek v3
+    - [x] Deepseek R1
+    - [x] Phi-4
+    - [ ] Unsloth 4-bit Dynamic Quants
+    - [ ] Vision/multimodal models
+- Sync model uploads with registry
+- Add utility methods for tracking model stats