Merge pull request #2255 from jeromeku/registry-refactor

Registry refactor
2025-04-01 23:16:27 -07:00 · 2025-04-01 23:16:27 -07:00 · fd20192aef
commit fd20192aef
parent 5ae9c4359b 9a14edcd2f
15 changed files with 1306 additions and 0 deletions
--- a/.gitignore
+++ b/.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
--- a/pyproject.toml
+++ b/pyproject.toml
@ -32,6 +32,10 @@ include-package-data = false
 exclude = ["images*", "tests*"]

 [project.optional-dependencies]
+dev = [
+    "pytest",
+]
+
 triton = [
    "triton-windows ; platform_system == 'Windows'",
 ]
--- a/tests/init.py
+++ b/tests/init.py
--- a/tests/test_model_registry.py
+++ b/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)
--- a/unsloth/registry/REGISTRY.md
+++ b/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
--- a/unsloth/registry/init.py
+++ b/unsloth/registry/init.py
@ -0,0 +1,51 @@
+from ._deepseek import register_deepseek_models as _register_deepseek_models
+from ._gemma import register_gemma_models as _register_gemma_models
+from ._llama import register_llama_models as _register_llama_models
+from ._mistral import register_mistral_models as _register_mistral_models
+from ._phi import register_phi_models as _register_phi_models
+from ._qwen import register_qwen_models as _register_qwen_models
+from .registry import MODEL_REGISTRY, ModelInfo, QuantType
+
+_ARE_MODELS_REGISTERED = False
+
+def register_models():
+    global _ARE_MODELS_REGISTERED
+
+    if _ARE_MODELS_REGISTERED:
+        return
+    _register_deepseek_models()
+    _register_gemma_models()
+    _register_llama_models()
+    _register_mistral_models()
+    _register_phi_models()
+    _register_qwen_models()
+
+    _ARE_MODELS_REGISTERED = True
+
+def search_models(org: str = None, base_name: str = None, version: str = None, size: str = None, quant_types: list[QuantType] = None, search_pattern: str = None) -> list[ModelInfo]:
+    """
+    Get model info from the registry.
+
+    See registry.ModelInfo for more fields.
+
+    If search_pattern is provided, the full model path will be matched against the pattern, where the model path is the model_id on huggingface hub.
+
+    """
+    if not _ARE_MODELS_REGISTERED:
+        register_models()
+    
+    model_infos = MODEL_REGISTRY.values()
+    if org:
+        model_infos = [model_info for model_info in model_infos if model_info.org == org]
+    if base_name:
+        model_infos = [model_info for model_info in model_infos if model_info.base_name == base_name]
+    if version:
+        model_infos = [model_info for model_info in model_infos if model_info.version == version]
+    if size:
+        model_infos = [model_info for model_info in model_infos if model_info.size == size]
+    if quant_types:
+        model_infos = [model_info for model_info in model_infos if any(model_info.quant_type == quant_type for quant_type in quant_types)]
+    if search_pattern:
+        model_infos = [model_info for model_info in model_infos if search_pattern in model_info.model_path]
+    
+    return model_infos
--- a/unsloth/registry/_deepseek.py
+++ b/unsloth/registry/_deepseek.py
@ -0,0 +1,179 @@
+from unsloth.registry.registry import ModelInfo, ModelMeta, QuantType, _register_models
+
+_IS_DEEPSEEK_V3_REGISTERED = False
+_IS_DEEPSEEK_V3_0324_REGISTERED = False
+_IS_DEEPSEEK_R1_REGISTERED = False
+_IS_DEEPSEEK_R1_ZERO_REGISTERED = False
+_IS_DEEPSEEK_R1_DISTILL_LLAMA_REGISTERED = False
+_IS_DEEPSEEK_R1_DISTILL_QWEN_REGISTERED = False
+
+class DeepseekV3ModelInfo(ModelInfo):
+    @classmethod
+    def construct_model_name(cls, base_name, version, size, quant_type, instruct_tag):
+        key = f"{base_name}-V{version}"
+        return super().construct_model_name(base_name, version, size, quant_type, instruct_tag, key)
+
+class DeepseekR1ModelInfo(ModelInfo):
+    @classmethod
+    def construct_model_name(cls, base_name, version, size, quant_type, instruct_tag):
+        key = f"{base_name}-{version}" if version else base_name
+        if size:
+            key = f"{key}-{size}B"
+        return super().construct_model_name(base_name, version, size, quant_type, instruct_tag, key)
+    
+# Deepseek V3 Model Meta
+DeepseekV3Meta = ModelMeta(
+    org="deepseek-ai",
+    base_name="DeepSeek",
+    instruct_tags=[None],
+    model_version="3",
+    model_sizes=[""],
+    model_info_cls=DeepseekV3ModelInfo,
+    is_multimodal=False,
+    quant_types=[QuantType.NONE, QuantType.BF16],
+)
+
+DeepseekV3_0324Meta = ModelMeta(
+    org="deepseek-ai",
+    base_name="DeepSeek",
+    instruct_tags=[None],
+    model_version="3-0324",
+    model_sizes=[""],
+    model_info_cls=DeepseekV3ModelInfo,
+    is_multimodal=False,
+    quant_types=[QuantType.NONE, QuantType.GGUF],
+)
+
+DeepseekR1Meta = ModelMeta(
+    org="deepseek-ai",
+    base_name="DeepSeek-R1",
+    instruct_tags=[None],
+    model_version="",
+    model_sizes=[""],
+    model_info_cls=DeepseekR1ModelInfo,
+    is_multimodal=False,
+    quant_types=[QuantType.NONE, QuantType.BF16, QuantType.GGUF],
+)
+
+DeepseekR1ZeroMeta = ModelMeta(
+    org="deepseek-ai",
+    base_name="DeepSeek-R1",
+    instruct_tags=[None],
+    model_version="Zero",
+    model_sizes=[""],
+    model_info_cls=DeepseekR1ModelInfo,
+    is_multimodal=False,
+    quant_types=[QuantType.NONE, QuantType.GGUF],
+)
+
+DeepseekR1DistillLlamaMeta = ModelMeta(
+    org="deepseek-ai",
+    base_name="DeepSeek-R1-Distill",
+    instruct_tags=[None],
+    model_version="Llama",
+    model_sizes=["8", "70"],
+    model_info_cls=DeepseekR1ModelInfo,
+    is_multimodal=False,
+    quant_types={"8": [QuantType.UNSLOTH, QuantType.GGUF], "70": [QuantType.GGUF]},
+)
+
+# Deepseek R1 Distill Qwen Model Meta
+DeepseekR1DistillQwenMeta = ModelMeta(
+    org="deepseek-ai",
+    base_name="DeepSeek-R1-Distill",
+    instruct_tags=[None],
+    model_version="Qwen",
+    model_sizes=["1.5", "7", "14", "32"],
+    model_info_cls=DeepseekR1ModelInfo,
+    is_multimodal=False,
+    quant_types={
+        "1.5": [QuantType.UNSLOTH, QuantType.BNB, QuantType.GGUF],
+        "7": [QuantType.UNSLOTH, QuantType.BNB],
+        "14": [QuantType.UNSLOTH, QuantType.BNB, QuantType.GGUF],
+        "32": [QuantType.GGUF, QuantType.BNB],
+    },
+)
+        
+def register_deepseek_v3_models(include_original_model: bool = False):
+    global _IS_DEEPSEEK_V3_REGISTERED
+    if _IS_DEEPSEEK_V3_REGISTERED:
+        return
+    _register_models(DeepseekV3Meta, include_original_model=include_original_model)
+    _IS_DEEPSEEK_V3_REGISTERED = True
+
+def register_deepseek_v3_0324_models(include_original_model: bool = False):
+    global _IS_DEEPSEEK_V3_0324_REGISTERED
+    if _IS_DEEPSEEK_V3_0324_REGISTERED:
+        return
+    _register_models(DeepseekV3_0324Meta, include_original_model=include_original_model)
+    _IS_DEEPSEEK_V3_0324_REGISTERED = True
+
+def register_deepseek_r1_models(include_original_model: bool = False):
+    global _IS_DEEPSEEK_R1_REGISTERED
+    if _IS_DEEPSEEK_R1_REGISTERED:
+        return
+    _register_models(DeepseekR1Meta, include_original_model=include_original_model)
+    _IS_DEEPSEEK_R1_REGISTERED = True
+
+def register_deepseek_r1_zero_models(include_original_model: bool = False):
+    global _IS_DEEPSEEK_R1_ZERO_REGISTERED
+    if _IS_DEEPSEEK_R1_ZERO_REGISTERED:
+        return
+    _register_models(DeepseekR1ZeroMeta, include_original_model=include_original_model)
+    _IS_DEEPSEEK_R1_ZERO_REGISTERED = True
+
+def register_deepseek_r1_distill_llama_models(include_original_model: bool = False):
+    global _IS_DEEPSEEK_R1_DISTILL_LLAMA_REGISTERED
+    if _IS_DEEPSEEK_R1_DISTILL_LLAMA_REGISTERED:
+        return
+    _register_models(DeepseekR1DistillLlamaMeta, include_original_model=include_original_model)
+    _IS_DEEPSEEK_R1_DISTILL_LLAMA_REGISTERED = True
+
+def register_deepseek_r1_distill_qwen_models(include_original_model: bool = False):
+    global _IS_DEEPSEEK_R1_DISTILL_QWEN_REGISTERED
+    if _IS_DEEPSEEK_R1_DISTILL_QWEN_REGISTERED:
+        return
+    _register_models(DeepseekR1DistillQwenMeta, include_original_model=include_original_model)
+    _IS_DEEPSEEK_R1_DISTILL_QWEN_REGISTERED = True
+
+def register_deepseek_models(include_original_model: bool = False):
+    register_deepseek_v3_models(include_original_model=include_original_model)
+    register_deepseek_v3_0324_models(include_original_model=include_original_model)
+    register_deepseek_r1_models(include_original_model=include_original_model)
+    register_deepseek_r1_zero_models(include_original_model=include_original_model)
+    register_deepseek_r1_distill_llama_models(include_original_model=include_original_model)
+    register_deepseek_r1_distill_qwen_models(include_original_model=include_original_model)
+
+def _list_deepseek_r1_distill_models():
+    from unsloth.utils.hf_hub import ModelInfo as HfModelInfo
+    from unsloth.utils.hf_hub import list_models
+    models: list[HfModelInfo] = list_models(author="unsloth", search="Distill", limit=1000)
+    distill_models = []
+    for model in models:
+        model_id = model.id
+        model_name = model_id.split("/")[-1]
+        # parse out only the version
+        version = model_name.removeprefix("DeepSeek-R1-Distill-")
+        distill_models.append(version)
+
+    return distill_models
+
+
+register_deepseek_models(include_original_model=True)
+
+if __name__ == "__main__":
+    from unsloth.registry.registry import MODEL_REGISTRY, _check_model_info
+    MODEL_REGISTRY.clear()
+    
+    register_deepseek_models(include_original_model=True)
+    
+    for model_id, model_info in MODEL_REGISTRY.items():
+        model_info = _check_model_info(model_id)
+        if model_info is None:
+            print(f"\u2718 {model_id}")
+        else:
+            print(f"\u2713 {model_id}")
+    # distill_models = _list_deepseek_r1_distill_models()
+    # for model in sorted(distill_models):
+    #     if "qwen" in model.lower():
+    #         print(model)
--- a/unsloth/registry/_gemma.py
+++ b/unsloth/registry/_gemma.py
@ -0,0 +1,66 @@
+from unsloth.registry.registry import ModelInfo, ModelMeta, QuantType, _register_models
+
+_IS_GEMMA_3_BASE_REGISTERED = False
+_IS_GEMMA_3_INSTRUCT_REGISTERED = False
+
+class GemmaModelInfo(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)
+
+# Gemma3 Base Model Meta
+GemmaMeta3Base = ModelMeta(
+    org="google",
+    base_name="gemma",
+    instruct_tags=["pt"],  # pt = base
+    model_version="3",
+    model_sizes=["1", "4", "12", "27"],
+    model_info_cls=GemmaModelInfo,
+    is_multimodal=True,
+    quant_types=[QuantType.NONE, QuantType.BNB, QuantType.UNSLOTH],
+)
+
+# Gemma3 Instruct Model Meta
+GemmaMeta3Instruct = ModelMeta(
+    org="google",
+    base_name="gemma",
+    instruct_tags=["it"],  # it = instruction tuned
+    model_version="3",
+    model_sizes=["1", "4", "12", "27"],
+    model_info_cls=GemmaModelInfo,
+    is_multimodal=True,
+    quant_types=[QuantType.NONE, QuantType.BNB, QuantType.UNSLOTH, QuantType.GGUF],
+)
+
+def register_gemma_3_base_models(include_original_model: bool = False):
+    global _IS_GEMMA_3_BASE_REGISTERED
+    if _IS_GEMMA_3_BASE_REGISTERED:
+        return
+    _register_models(GemmaMeta3Base, include_original_model=include_original_model)
+    _IS_GEMMA_3_BASE_REGISTERED = True
+
+def register_gemma_3_instruct_models(include_original_model: bool = False):
+    global _IS_GEMMA_3_INSTRUCT_REGISTERED
+    if _IS_GEMMA_3_INSTRUCT_REGISTERED:
+        return
+    _register_models(GemmaMeta3Instruct, include_original_model=include_original_model)
+    _IS_GEMMA_3_INSTRUCT_REGISTERED = True
+
+def register_gemma_models(include_original_model: bool = False):
+    register_gemma_3_base_models(include_original_model=include_original_model)
+    register_gemma_3_instruct_models(include_original_model=include_original_model)
+
+
+if __name__ == "__main__":
+    from unsloth.registry.registry import MODEL_REGISTRY, _check_model_info
+    MODEL_REGISTRY.clear()
+    
+    register_gemma_models(include_original_model=True)
+    
+    for model_id, model_info in MODEL_REGISTRY.items():
+        model_info = _check_model_info(model_id)
+        if model_info is None:
+            print(f"\u2718 {model_id}")
+        else:
+            print(f"\u2713 {model_id}")
--- a/unsloth/registry/_llama.py
+++ b/unsloth/registry/_llama.py
@ -0,0 +1,113 @@
+from unsloth.registry.registry import ModelInfo, ModelMeta, QuantType, _register_models
+
+_IS_LLAMA_3_1_REGISTERED = False
+_IS_LLAMA_3_2_REGISTERED = False
+_IS_LLAMA_3_2_VISION_REGISTERED = False
+
+
+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)
+
+
+class LlamaVisionModelInfo(ModelInfo):
+    @classmethod
+    def construct_model_name(cls, base_name, version, size, quant_type, instruct_tag):
+        key = f"{base_name}-{version}-{size}B-Vision"
+        return super().construct_model_name(base_name, version, size, quant_type, instruct_tag, key)
+
+
+# Llama 3.1
+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],
+)
+
+# Llama 3.2 Base Models
+LlamaMeta_3_2_Base = ModelMeta(
+    org="meta-llama",
+    base_name="Llama",
+    instruct_tags=[None],
+    model_version="3.2",
+    model_sizes=["1", "3"],
+    model_info_cls=LlamaModelInfo,
+    is_multimodal=False,
+    quant_types=[QuantType.NONE, QuantType.BNB, QuantType.UNSLOTH],
+)
+
+# Llama 3.2 Instruction Tuned Models
+LlamaMeta_3_2_Instruct = ModelMeta(
+    org="meta-llama",
+    base_name="Llama",
+    instruct_tags=["Instruct"],
+    model_version="3.2",
+    model_sizes=["1", "3"],
+    model_info_cls=LlamaModelInfo,
+    is_multimodal=False,
+    quant_types=[QuantType.NONE, QuantType.BNB, QuantType.UNSLOTH, QuantType.GGUF],
+)
+
+# Llama 3.2 Vision
+LlamaMeta_3_2_Vision = ModelMeta(
+    org="meta-llama",
+    base_name="Llama",
+    instruct_tags=[None, "Instruct"],
+    model_version="3.2",
+    model_sizes=["11", "90"],
+    model_info_cls=LlamaVisionModelInfo,
+    is_multimodal=True,
+    quant_types={
+        "11": [QuantType.NONE, QuantType.BNB, QuantType.UNSLOTH],
+        "90": [QuantType.NONE],
+    },
+)
+
+
+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
+
+def register_llama_3_2_models(include_original_model: bool = False):
+    global _IS_LLAMA_3_2_REGISTERED
+    if _IS_LLAMA_3_2_REGISTERED:
+        return
+    _register_models(LlamaMeta_3_2_Base, include_original_model=include_original_model)
+    _register_models(LlamaMeta_3_2_Instruct, include_original_model=include_original_model)
+    _IS_LLAMA_3_2_REGISTERED = True
+
+def register_llama_3_2_vision_models(include_original_model: bool = False):
+    global _IS_LLAMA_3_2_VISION_REGISTERED
+    if _IS_LLAMA_3_2_VISION_REGISTERED:
+        return
+    _register_models(LlamaMeta_3_2_Vision, include_original_model=include_original_model)
+    _IS_LLAMA_3_2_VISION_REGISTERED = True
+
+
+def register_llama_models(include_original_model: bool = False):
+    register_llama_3_1_models(include_original_model=include_original_model)
+    register_llama_3_2_models(include_original_model=include_original_model)
+    register_llama_3_2_vision_models(include_original_model=include_original_model)
+
+if __name__ == "__main__":
+    from unsloth.registry.registry import MODEL_REGISTRY, _check_model_info
+    MODEL_REGISTRY.clear()
+
+    register_llama_models(include_original_model=True)
+
+    for model_id, model_info in MODEL_REGISTRY.items():
+        model_info = _check_model_info(model_id)
+        if model_info is None:
+            print(f"\u2718 {model_id}")
+        else:
+            print(f"\u2713 {model_id}")
--- a/unsloth/registry/_mistral.py
+++ b/unsloth/registry/_mistral.py
@ -0,0 +1,70 @@
+import copy
+
+from unsloth.registry.registry import ModelInfo, ModelMeta, QuantType, _register_models
+
+_IS_MISTRAL_SMALL_REGISTERED = False
+
+_MISTRAL_SMALL_03_25_VERSION = "2503"
+_MISTRAL_SMALL_01_25_VERSION = "2501"
+_MISTRAL_SMALL_09_24_VERSION = "2409" # Not uploaded to unsloth
+
+class MistralSmallModelInfo(ModelInfo):
+    @classmethod
+    def construct_model_name(cls, base_name, version, size, quant_type, instruct_tag):
+        if version == _MISTRAL_SMALL_03_25_VERSION:
+            key = f"{base_name}-3.1-{size}B-{instruct_tag}"
+        else:
+            key = f"{base_name}-{size}B-{instruct_tag}"
+        key += f"-{version}"
+        key = cls.append_quant_type(key, quant_type)
+        
+        return key
+
+
+MistralSmall_2503_Base_Meta = ModelMeta(
+    org="mistralai",
+    base_name="Mistral-Small",
+    instruct_tags=["Base"],
+    model_version=_MISTRAL_SMALL_03_25_VERSION,
+    model_sizes=["24"],
+    model_info_cls=MistralSmallModelInfo,
+    is_multimodal=False,
+    quant_types=[QuantType.NONE, QuantType.UNSLOTH, QuantType.BNB],
+)
+
+MistralSmall_2503_Instruct_Meta = copy.deepcopy(MistralSmall_2503_Base_Meta)
+MistralSmall_2503_Instruct_Meta.instruct_tags = ["Instruct"]
+MistralSmall_2503_Instruct_Meta.quant_types = [QuantType.NONE, QuantType.UNSLOTH, QuantType.BNB, QuantType.GGUF]
+
+MistralSmall_2501_Base_Meta = copy.deepcopy(MistralSmall_2503_Base_Meta)
+MistralSmall_2501_Base_Meta.model_version = _MISTRAL_SMALL_01_25_VERSION
+
+MistralSmall_2501_Instruct_Meta = copy.deepcopy(MistralSmall_2503_Instruct_Meta)
+MistralSmall_2501_Instruct_Meta.model_version = _MISTRAL_SMALL_01_25_VERSION
+
+def register_mistral_small_models(include_original_model: bool = False):
+    global _IS_MISTRAL_SMALL_REGISTERED
+    if _IS_MISTRAL_SMALL_REGISTERED:
+        return
+    _register_models(MistralSmall_2503_Base_Meta, include_original_model=include_original_model)
+    _register_models(MistralSmall_2503_Instruct_Meta, include_original_model=include_original_model)
+    _register_models(MistralSmall_2501_Base_Meta, include_original_model=include_original_model)
+    _register_models(MistralSmall_2501_Instruct_Meta, include_original_model=include_original_model)
+
+    _IS_MISTRAL_SMALL_REGISTERED = True
+
+def register_mistral_models(include_original_model: bool = False):
+    register_mistral_small_models(include_original_model=include_original_model)
+
+if __name__ == "__main__":
+    from unsloth.registry.registry import MODEL_REGISTRY, _check_model_info
+    MODEL_REGISTRY.clear()
+    
+    register_mistral_models(include_original_model=True)
+    
+    for model_id, model_info in MODEL_REGISTRY.items():
+        model_info = _check_model_info(model_id)
+        if model_info is None:
+            print(f"\u2718 {model_id}")
+        else:
+            print(f"\u2713 {model_id}")    
--- a/unsloth/registry/_phi.py
+++ b/unsloth/registry/_phi.py
@ -0,0 +1,65 @@
+from unsloth.registry.registry import ModelInfo, ModelMeta, QuantType, _register_models
+
+_IS_PHI_4_REGISTERED = False
+_IS_PHI_4_INSTRUCT_REGISTERED = False
+
+class PhiModelInfo(ModelInfo):
+    @classmethod
+    def construct_model_name(cls, base_name, version, size, quant_type, instruct_tag):
+        key = f"{base_name}-{version}"
+        return super().construct_model_name(base_name, version, size, quant_type, instruct_tag, key)
+
+# Phi Model Meta
+PhiMeta4 = ModelMeta(
+    org="microsoft",
+    base_name="phi",
+    instruct_tags=[None],
+    model_version="4",
+    model_sizes=["1"],  # Assuming only one size
+    model_info_cls=PhiModelInfo,
+    is_multimodal=False,
+    quant_types=[QuantType.NONE, QuantType.BNB, QuantType.UNSLOTH],
+)
+
+# Phi Instruct Model Meta
+PhiInstructMeta4 = ModelMeta(
+    org="microsoft",
+    base_name="phi",
+    instruct_tags=["mini-instruct"],
+    model_version="4",
+    model_sizes=["1"],  # Assuming only one size
+    model_info_cls=PhiModelInfo,
+    is_multimodal=False,
+    quant_types=[QuantType.NONE, QuantType.BNB, QuantType.UNSLOTH, QuantType.GGUF],
+)
+
+def register_phi_4_models(include_original_model: bool = False):
+    global _IS_PHI_4_REGISTERED
+    if _IS_PHI_4_REGISTERED:
+        return
+    _register_models(PhiMeta4, include_original_model=include_original_model)
+    _IS_PHI_4_REGISTERED = True
+
+def register_phi_4_instruct_models(include_original_model: bool = False):
+    global _IS_PHI_4_INSTRUCT_REGISTERED
+    if _IS_PHI_4_INSTRUCT_REGISTERED:
+        return
+    _register_models(PhiInstructMeta4, include_original_model=include_original_model)
+    _IS_PHI_4_INSTRUCT_REGISTERED = True
+
+def register_phi_models(include_original_model: bool = False):
+    register_phi_4_models(include_original_model=include_original_model)
+    register_phi_4_instruct_models(include_original_model=include_original_model)
+
+if __name__ == "__main__":
+    from unsloth.registry.registry import MODEL_REGISTRY, _check_model_info
+    MODEL_REGISTRY.clear()
+    
+    register_phi_models(include_original_model=True)
+    
+    for model_id, model_info in MODEL_REGISTRY.items():
+        model_info = _check_model_info(model_id)
+        if model_info is None:
+            print(f"\u2718 {model_id}")
+        else:
+            print(f"\u2713 {model_id}") 
--- a/unsloth/registry/_qwen.py
+++ b/unsloth/registry/_qwen.py
@ -0,0 +1,117 @@
+from unsloth.registry.registry import ModelInfo, ModelMeta, QuantType, _register_models
+
+_IS_QWEN_2_5_REGISTERED = False
+_IS_QWEN_2_5_VL_REGISTERED = False
+_IS_QWEN_QWQ_REGISTERED = False
+class QwenModelInfo(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)
+
+
+class QwenVLModelInfo(ModelInfo):
+    @classmethod
+    def construct_model_name(cls, base_name, version, size, quant_type, instruct_tag):
+        key = f"{base_name}{version}-VL-{size}B"
+        return super().construct_model_name(base_name, version, size, quant_type, instruct_tag, key)
+
+class QwenQwQModelInfo(ModelInfo):
+    @classmethod
+    def construct_model_name(cls, base_name, version, size, quant_type, instruct_tag):
+        key = f"{base_name}-{size}B"
+        return super().construct_model_name(base_name, version, size, quant_type, instruct_tag, key)
+    
+class QwenQVQPreviewModelInfo(ModelInfo):
+    @classmethod
+    def construct_model_name(cls, base_name, version, size, quant_type, instruct_tag):
+        key = f"{base_name}-{size}B-Preview"
+        return super().construct_model_name(base_name, version, size, quant_type, instruct_tag, key)
+    
+# Qwen2.5 Model Meta
+Qwen_2_5_Meta = ModelMeta(
+    org="Qwen",
+    base_name="Qwen",
+    instruct_tags=[None, "Instruct"],
+    model_version="2.5",
+    model_sizes=["3", "7"],
+    model_info_cls=QwenModelInfo,
+    is_multimodal=False,
+    quant_types=[QuantType.NONE, QuantType.BNB, QuantType.UNSLOTH],
+)
+
+# Qwen2.5 VL Model Meta
+Qwen_2_5_VLMeta = ModelMeta(
+    org="Qwen",
+    base_name="Qwen",
+    instruct_tags=["Instruct"],  # No base, only instruction tuned
+    model_version="2.5",
+    model_sizes=["3", "7", "32", "72"],
+    model_info_cls=QwenVLModelInfo,
+    is_multimodal=True,
+    quant_types=[QuantType.NONE, QuantType.BNB, QuantType.UNSLOTH],
+)
+
+# Qwen QwQ Model Meta
+QwenQwQMeta = ModelMeta(
+    org="Qwen",
+    base_name="QwQ",
+    instruct_tags=[None],
+    model_version="",
+    model_sizes=["32"],
+    model_info_cls=QwenQwQModelInfo,
+    is_multimodal=False,
+    quant_types=[QuantType.NONE, QuantType.BNB, QuantType.UNSLOTH, QuantType.GGUF],
+)
+
+# Qwen QVQ Preview Model Meta
+QwenQVQPreviewMeta = ModelMeta(
+    org="Qwen",
+    base_name="QVQ",
+    instruct_tags=[None],
+    model_version="",
+    model_sizes=["72"],
+    model_info_cls=QwenQVQPreviewModelInfo,
+    is_multimodal=True,
+    quant_types=[QuantType.NONE, QuantType.BNB],
+)
+
+def register_qwen_2_5_models(include_original_model: bool = False):
+    global _IS_QWEN_2_5_REGISTERED
+    if _IS_QWEN_2_5_REGISTERED:
+        return
+    _register_models(Qwen_2_5_Meta, include_original_model=include_original_model)
+    _IS_QWEN_2_5_REGISTERED = True
+
+def register_qwen_2_5_vl_models(include_original_model: bool = False):
+    global _IS_QWEN_2_5_VL_REGISTERED
+    if _IS_QWEN_2_5_VL_REGISTERED:
+        return
+    _register_models(Qwen_2_5_VLMeta, include_original_model=include_original_model)
+    _IS_QWEN_2_5_VL_REGISTERED = True
+
+def register_qwen_qwq_models(include_original_model: bool = False):
+    global _IS_QWEN_QWQ_REGISTERED
+    if _IS_QWEN_QWQ_REGISTERED:
+        return
+    _register_models(QwenQwQMeta, include_original_model=include_original_model)
+    _register_models(QwenQVQPreviewMeta, include_original_model=include_original_model)
+    _IS_QWEN_QWQ_REGISTERED = True
+
+def register_qwen_models(include_original_model: bool = False):
+    register_qwen_2_5_models(include_original_model=include_original_model)
+    register_qwen_2_5_vl_models(include_original_model=include_original_model)
+    register_qwen_qwq_models(include_original_model=include_original_model)
+
+if __name__ == "__main__":
+    from unsloth.registry.registry import MODEL_REGISTRY, _check_model_info
+    MODEL_REGISTRY.clear()
+    
+    register_qwen_models(include_original_model=True)
+    
+    for model_id, model_info in MODEL_REGISTRY.items():
+        model_info = _check_model_info(model_id)
+        if model_info is None:
+            print(f"\u2718 {model_id}")
+        else:
+            print(f"\u2713 {model_id}")
--- a/unsloth/registry/registry.py
+++ b/unsloth/registry/registry.py
@ -0,0 +1,185 @@
+import warnings
+from dataclasses import dataclass, field
+from enum import Enum
+
+
+class QuantType(Enum):
+    BNB = "bnb"
+    UNSLOTH = "unsloth" # dynamic 4-bit quantization
+    GGUF = "GGUF"
+    NONE = "none"
+    BF16 = "bf16" # only for Deepseek V3
+
+# Tags for Hugging Face model paths
+BNB_QUANTIZED_TAG = "bnb-4bit"
+UNSLOTH_DYNAMIC_QUANT_TAG = "unsloth" + "-" + BNB_QUANTIZED_TAG
+GGUF_TAG = "GGUF"
+BF16_TAG = "bf16"
+
+QUANT_TAG_MAP = {
+    QuantType.BNB: BNB_QUANTIZED_TAG,
+    QuantType.UNSLOTH: UNSLOTH_DYNAMIC_QUANT_TAG,
+    QuantType.GGUF: GGUF_TAG,
+    QuantType.NONE: None,
+    QuantType.BF16: BF16_TAG,
+} 
+
+# NOTE: models registered with org="unsloth" and QUANT_TYPE.NONE are aliases of QUANT_TYPE.UNSLOTH
+@dataclass
+class ModelInfo:
+    org: str
+    base_name: str
+    version: str
+    size: int
+    name: str = None  # full model name, constructed from base_name, version, and size unless provided
+    is_multimodal: bool = False
+    instruct_tag: str = None
+    quant_type: QuantType = None
+    description: str = None
+
+    def __post_init__(self):
+        self.name = self.name or self.construct_model_name(
+            self.base_name,
+            self.version,
+            self.size,
+            self.quant_type,
+            self.instruct_tag,
+        )
+
+    @staticmethod
+    def append_instruct_tag(key: str, instruct_tag: str = None):
+        if instruct_tag:
+            key = "-".join([key, instruct_tag])
+        return key
+
+    @staticmethod
+    def append_quant_type(
+        key: str, quant_type: QuantType = None
+    ):
+        if quant_type != QuantType.NONE:
+            key = "-".join([key, QUANT_TAG_MAP[quant_type]])
+        return key
+
+    @classmethod
+    def construct_model_name(cls, base_name, version, size, quant_type, instruct_tag, key=""):
+        key = cls.append_instruct_tag(key, instruct_tag)
+        key = cls.append_quant_type(key, quant_type)
+        return key
+
+    @property
+    def model_path(
+        self,
+    ) -> str:
+        return f"{self.org}/{self.name}"
+
+
+@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
+
+
+MODEL_REGISTRY: dict[str, ModelInfo] = {}
+
+
+def register_model(
+    model_info_cls: ModelInfo,
+    org: str,
+    base_name: str,
+    version: str,
+    size: int,
+    instruct_tag: str = None,
+    quant_type: QuantType = None,
+    is_multimodal: bool = False,
+    name: str = None,
+):
+    name = name or model_info_cls.construct_model_name(
+        base_name=base_name,
+        version=version,
+        size=size,
+        quant_type=quant_type,
+        instruct_tag=instruct_tag,
+    )
+    key = f"{org}/{name}"
+
+    if key in MODEL_REGISTRY:
+        raise ValueError(f"Model {key} already registered, current keys: {MODEL_REGISTRY.keys()}")
+
+    MODEL_REGISTRY[key] = model_info_cls(
+        org=org,
+        base_name=base_name,
+        version=version,
+        size=size,
+        is_multimodal=is_multimodal,
+        instruct_tag=instruct_tag,
+        quant_type=quant_type,
+        name=name,
+    )
+
+
+def _check_model_info(model_id: str, properties: list[str] = ["lastModified"]):
+    from huggingface_hub import HfApi
+    from huggingface_hub import ModelInfo as HfModelInfo
+    from huggingface_hub.utils import RepositoryNotFoundError
+
+    api = HfApi()
+
+    try:
+        model_info: HfModelInfo = api.model_info(model_id, expand=properties)
+    except Exception as e:
+        if isinstance(e, RepositoryNotFoundError):
+            warnings.warn(f"{model_id} not found on Hugging Face")
+            model_info = None
+        else:
+            raise e
+    return model_info
+
+
+def _register_models(model_meta: ModelMeta, include_original_model: bool = False):
+    org = model_meta.org
+    base_name = model_meta.base_name
+    instruct_tags = model_meta.instruct_tags
+    model_version = model_meta.model_version
+    model_sizes = model_meta.model_sizes
+    is_multimodal = model_meta.is_multimodal
+    quant_types = model_meta.quant_types
+    model_info_cls = model_meta.model_info_cls
+
+    for size in model_sizes:
+        for instruct_tag in instruct_tags:
+            # Handle quant types per model size
+            if isinstance(quant_types, dict):
+                _quant_types = quant_types[size]
+            else:
+                _quant_types = quant_types
+            for quant_type in _quant_types:
+                # NOTE: models registered with org="unsloth" and QUANT_TYPE.NONE are aliases of QUANT_TYPE.UNSLOTH
+                _org = "unsloth" # unsloth models -- these are all quantized versions of the original model
+                register_model(
+                    model_info_cls=model_info_cls,
+                    org=_org,
+                    base_name=base_name,
+                    version=model_version,
+                    size=size,
+                    instruct_tag=instruct_tag,
+                    quant_type=quant_type,
+                    is_multimodal=is_multimodal,
+                )
+            # include original model from releasing organization
+            if include_original_model:
+                register_model(
+                    model_info_cls=model_info_cls,
+                    org=org,
+                    base_name=base_name,
+                    version=model_version,
+                    size=size,
+                    instruct_tag=instruct_tag,
+                    quant_type=QuantType.NONE,
+                    is_multimodal=is_multimodal,
+                )
--- a/unsloth/utils/init.py
+++ b/unsloth/utils/init.py
--- a/unsloth/utils/hf_hub.py
+++ b/unsloth/utils/hf_hub.py
@ -0,0 +1,78 @@
+from huggingface_hub import HfApi, ModelInfo
+
+_HFAPI: HfApi = None
+
+POPULARITY_PROPERTIES = [
+    "downloads",
+    "downloadsAllTime",
+    "trendingScore",
+    "likes",
+]
+THOUSAND = 1000
+MILLION = 1000000
+BILLION = 1000000000
+
+
+def formatted_int(value: int) -> str:
+    if value < THOUSAND:
+        return str(value)
+    elif value < MILLION:
+        return f"{float(value) / 1000:,.1f}K"
+    elif value < BILLION:
+        return f"{float(value) // 1000000:,.1f}M"
+
+
+def get_model_info(
+    model_id: str, properties: list[str] = ["safetensors", "lastModified"]
+) -> ModelInfo:
+    """
+    Get the model info for a specific model.
+
+    properties: list[str] = See https://huggingface.co/docs/huggingface_hub/api-ref/hf_hub/hf_api/model_info
+    Default properties: ["safetensors", "lastModified"], only retrieves minimal information.
+    Set to None to retrieve the full model information.
+    """
+    global _HFAPI
+    if _HFAPI is None:
+        _HFAPI = HfApi()
+    try:
+        model_info: ModelInfo = _HFAPI.model_info(model_id, expand=properties)
+    except Exception as e:
+        print(f"Error getting model info for {model_id}: {e}")
+        model_info = None
+    return model_info
+
+
+def list_models(
+    properties: list[str] = None,
+    full: bool = False,
+    sort: str = "downloads",
+    author: str = "unsloth",
+    search: str = None,
+    limit: int = 10,
+) -> list[ModelInfo]:
+    """
+    Retrieve model information from the Hugging Face Hub.
+
+    properties: list[str] = See https://huggingface.co/docs/huggingface_hub/api-ref/hf_hub/hf_api/list_models
+    full: bool = Whether to retrieve the full model information, if True properties will be ignored.
+    sort: str = The sort order.
+    author: str = The author of the model.
+    search: str = The search query for filtering models.
+
+    """
+    global _HFAPI
+    if _HFAPI is None:
+        _HFAPI = HfApi()
+    if full:
+        properties = None
+
+    models: list[ModelInfo] = _HFAPI.list_models(
+        author=author,
+        search=search,
+        sort=sort,
+        limit=limit,
+        expand=properties,
+        full=full,
+    )
+    return models