Skip to content

Local UI to run and train LLMs and diffusion models. Supports GGUF, MLX, Qwen3.8, DeepSeek-V4.1, MiniMax-H3, Gemma 4, FLUX and more.

13.3 KB
Raw Blame
# SPDX-License-Identifier: AGPL-3.0-only
# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved.

import __future__
import ast
import inspect
import fnmatch
import os
import sys
from pathlib import Path
from types import SimpleNamespace

import pytest


class DispatchReached(Exception):
    pass


def _mistral_format_names(tree):
    """Names loader.py imports from .mistral_format, and its own *mistral_format* helpers."""
    names = set()
    for node in tree.body:
        if isinstance(node, ast.ImportFrom) and node.module == "mistral_format":
            names.update(alias.asname or alias.name for alias in node.names)
        elif isinstance(node, ast.FunctionDef) and "mistral_format" in node.name:
            names.add(node.name)
    return names


def _tiling_names(tree):
    """Names loader.py imports from ..tiling (tiled MLP), which these precision checks never reach."""
    return {
        alias.asname or alias.name
        for node in tree.body
        if isinstance(node, ast.ImportFrom) and node.module == "tiling" and node.level == 2
        for alias in node.names
    }


def _loader_utils_functions(tree, path, method):
    """Real definitions of the .loader_utils functions loader.py imports and `method` calls."""
    imported = {
        alias.asname or alias.name: alias.name
        for node in tree.body
        if isinstance(node, ast.ImportFrom) and node.module == "loader_utils"
        for alias in node.names
    }
    used = {n.id for n in ast.walk(method) if isinstance(n, ast.Name)}
    source = ast.parse((path.parent / "loader_utils.py").read_text(encoding = "utf-8"))
    defs = {n.name: n for n in source.body if isinstance(n, ast.FunctionDef)}
    return [
        defs[imported[name]] for name in sorted(used & imported.keys()) if imported[name] in defs
    ]


def _module_definitions(tree):
    """loader.py's own top-level functions and literal constants, by name."""
    defined = {}
    for node in tree.body:
        if isinstance(node, ast.FunctionDef):
            defined[node.name] = node
        elif (
            isinstance(node, ast.Assign)
            and len(node.targets) == 1
            and isinstance(node.targets[0], ast.Name)
        ):
            try:
                ast.literal_eval(node.value)
            except ValueError:
                continue
            defined[node.targets[0].id] = node
    return defined


def _closure(roots, defined):
    """Every definition in `defined` the roots reach by name, transitively, in source order."""
    wanted, pending = set(), list(roots)
    while pending:
        for node in ast.walk(pending.pop()):
            if isinstance(node, ast.Name) and node.id in defined and node.id not in wanted:
                wanted.add(node.id)
                pending.append(defined[node.id])
    return [node for name, node in defined.items() if name in wanted]


@pytest.fixture
def loader():
    path = Path(__file__).resolve().parents[1] / "unsloth/models/loader.py"
    tree = ast.parse(path.read_text(encoding = "utf-8"))
    cls = next(
        n for n in tree.body if isinstance(n, ast.ClassDef) and n.name == "FastLanguageModel"
    )
    method = next(
        n for n in cls.body if isinstance(n, ast.FunctionDef) and n.name == "from_pretrained"
    )
    method.decorator_list = []
    captured = {"config_calls": 0}

    def dispatch(**kwargs):
        captured["dispatch"] = kwargs
        raise DispatchReached()

    def config(*args, **kwargs):
        captured["config_calls"] += 1
        return SimpleNamespace(model_type = "llama", rope_scaling = None)

    def no_adapter(*args, **kwargs):
        raise ValueError("No adapter")

    env = dict(
        os = os,
        DEFAULT_DEVICE_MAP = "sequential",
        OFFLOAD_EMBEDDING_AUTO = "auto",
        torch = SimpleNamespace(
            float16 = "float16", bfloat16 = "bfloat16", float32 = "float32", dtype = type(None)
        ),
        _requested_float32 = lambda dtype: False,
        hf_login = lambda token: token,
        requested_device_map = lambda device: device,
        deepspeed_zero3_device_map = lambda device: device,
        is_automatic_device_map = lambda device: isinstance(device, str),
        prepare_device_map = lambda: ("sequential", False),
        ALLOW_BITSANDBYTES = True,
        ALLOW_PREQUANTIZED_MODELS = True,
        USE_MODELSCOPE = False,
        logger = SimpleNamespace(warning_once = lambda *args, **kwargs: None),
        SUPPORTS_LLAMA32 = True,
        get_model_name = lambda name, **kwargs: name,
        _revision_for_resolved_repo = lambda revision, *args: revision,
        AutoConfig = SimpleNamespace(from_pretrained = config),
        PeftConfig = SimpleNamespace(from_pretrained = no_adapter),
        # No load scope here, so every probe reaches its loader, as _memo_config_probe does outside one.
        _memo_config_probe = lambda loader, name, **kwargs: loader(name, **kwargs),
        get_transformers_model_type = lambda *args, **kwargs: ["llama"],
        FastLlamaModel = SimpleNamespace(from_pretrained = dispatch),
        apply_unsloth_gradient_checkpointing = lambda value, *args: value,
        _resolve_checkpoint_tokenizer_name = lambda *args: None,
        patch_compiling_bitsandbytes = lambda: None,
        _get_dtype = lambda dtype: dtype,
        _revision_for_tokenizer_repo = lambda *args: None,
        _raise_if_modeling_ignores_config = lambda *args: None,
        raise_if_fast_inference_under_fsdp2 = lambda fast_inference: None,
    )
    # No fixture is Mistral-format: stub each loader.py helper deciding that, new ones included.
    for name in _mistral_format_names(tree):
        env.setdefault(name, RuntimeError if name[0].isupper() else (lambda *args, **kwargs: False))
    for name in _tiling_names(tree):
        env.setdefault(name, lambda *args, **kwargs: None)
    # Real loader_utils helpers (e.g. the fp8 dequantize check from #14516) where no stub is given.
    env.setdefault("inspect", inspect)
    real = [n for n in _loader_utils_functions(tree, path, method) if n.name not in env]
    flags = __future__.annotations.compiler_flag  # annotations such as Union stay unevaluated
    source = str(path.parent / "loader_utils.py")
    exec(compile(ast.Module(body = real, type_ignores = []), source, "exec", flags = flags), env)
    # Every loader.py helper from_pretrained reaches and no stub above replaces, new ones included
    # (a fixed list missed _agree_fast_inference_across_ranks from #15390).
    defined = {name: node for name, node in _module_definitions(tree).items() if name not in env}
    helpers = _closure([method], defined)
    for node in helpers:
        if isinstance(node, ast.FunctionDef):
            node.decorator_list = []
    exec(compile(ast.Module(body = [*helpers, method], type_ignores = []), str(path), "exec"), env)
    return env, captured


@pytest.mark.parametrize(
    "kwargs",
    [
        {"load_in_16bit": True},
        {"load_in_4bit": True, "load_in_16bit": True},
        {"load_in_16bit": True, "quantization_config": {"load_in_4bit": True}},
        {"load_in_fp8": True},
    ],
)
def test_conflicts_fail_before_model_loading(loader, kwargs):
    env, captured = loader
    with pytest.raises(RuntimeError, match = "Can only load in"):
        env["from_pretrained"](**kwargs)
    assert "dispatch" not in captured


@pytest.mark.parametrize(
    "kwargs, allow_bnb, expected_4bit",
    [
        ({}, True, True),
        ({"load_in_4bit": False, "load_in_16bit": True}, True, False),
        ({"model_name": "example/model-bf16", "load_in_16bit": True}, True, False),
        ({"load_in_16bit": True}, False, False),
        ({"load_in_16bit": True, "quantization_config": {"quant_method": "awq"}}, True, False),
    ],
)
def test_valid_precision_and_existing_overrides(loader, kwargs, allow_bnb, expected_4bit):
    env, captured = loader
    env["ALLOW_BITSANDBYTES"] = allow_bnb
    with pytest.raises(DispatchReached):
        env["from_pretrained"](**kwargs)
    assert captured["dispatch"]["load_in_4bit"] is expected_4bit
    if "quantization_config" in kwargs:
        assert captured["dispatch"]["quantization_config"] == kwargs["quantization_config"]


@pytest.mark.parametrize("base_name, expected", [("owner/base-bf16", False), ("owner/base", None)])
def test_adapter_base_precision_is_resolved_before_validation(loader, base_name, expected):
    env, captured = loader

    def config(name, **kwargs):
        if name == "owner/adapter":
            raise ValueError("Adapter has no model config")
        return SimpleNamespace(model_type = "llama", rope_scaling = None)

    env["AutoConfig"] = SimpleNamespace(from_pretrained = config)
    env["PeftConfig"] = SimpleNamespace(
        from_pretrained = lambda *a, **k: SimpleNamespace(base_model_name_or_path = base_name)
    )
    if expected is None:
        with pytest.raises(RuntimeError, match = "Can only load in"):
            env["from_pretrained"](model_name = "owner/adapter", load_in_16bit = True)
        assert "dispatch" not in captured
    else:
        with pytest.raises(DispatchReached):
            env["from_pretrained"](model_name = "owner/adapter", load_in_16bit = True)
        assert captured["dispatch"]["model_name"] == base_name
        assert captured["dispatch"]["load_in_4bit"] is expected


@pytest.fixture
def modelscope_snapshot(monkeypatch, tmp_path):
    cache = tmp_path / "modelscope-cache"
    cache.mkdir()
    downloaded = []
    calls = []

    def snapshot_download(name, allow_file_pattern = None):
        calls.append((name, allow_file_pattern))
        if name.startswith("hf-only/"):
            raise ValueError(f"{name} is not on ModelScope")
        snapshot = cache / name.replace("/", "--")
        snapshot.mkdir(exist_ok = True)
        for filename in (
            "config.json",
            "adapter_config.json",
            "configuration_custom.py",
            "model.safetensors",
        ):
            if allow_file_pattern is None or any(
                fnmatch.fnmatch(filename, pattern) for pattern in allow_file_pattern
            ):
                (snapshot / filename).write_text("test fixture")
                downloaded.append(filename)
        return str(snapshot)

    monkeypatch.setitem(
        sys.modules, "modelscope", SimpleNamespace(snapshot_download = snapshot_download)
    )
    return cache, downloaded, calls


def use_modelscope_adapter(env, cache, base_name):
    def config(name, **kwargs):
        if name == str(cache / "owner--model"):
            raise ValueError("Adapter has no model config")
        return SimpleNamespace(model_type = "llama", rope_scaling = None)

    env["AutoConfig"] = SimpleNamespace(from_pretrained = config)
    env["PeftConfig"] = SimpleNamespace(
        from_pretrained = lambda *a, **k: SimpleNamespace(base_model_name_or_path = base_name)
    )


@pytest.mark.parametrize(
    "kwargs, adapter_base",
    [
        ({"load_in_16bit": True}, None),
        ({"load_in_4bit": True, "load_in_16bit": True}, None),
        ({"load_in_16bit": True, "quantization_config": {"load_in_4bit": True}}, None),
        ({"load_in_16bit": True}, "owner/base"),
    ],
)
def test_modelscope_conflicts_do_not_download_weights(
    loader, modelscope_snapshot, kwargs, adapter_base
):
    env, captured = loader
    cache, downloaded, calls = modelscope_snapshot
    env["USE_MODELSCOPE"] = True
    if adapter_base:
        use_modelscope_adapter(env, cache, adapter_base)

    with pytest.raises(RuntimeError, match = "Can only load in"):
        env["from_pretrained"](model_name = "owner/model", **kwargs)

    assert "dispatch" not in captured
    assert "model.safetensors" not in downloaded
    assert "config.json" in downloaded
    assert "adapter_config.json" in downloaded
    assert "configuration_custom.py" in downloaded
    assert len(calls) == 1


@pytest.mark.parametrize(
    "kwargs, adapter_base, expected_4bit",
    [
        ({}, None, True),
        ({"load_in_4bit": False, "load_in_16bit": True}, None, False),
        ({"load_in_16bit": True}, "owner/base-bf16", False),
    ],
)
def test_modelscope_valid_loads_download_weights(
    loader, modelscope_snapshot, kwargs, adapter_base, expected_4bit
):
    env, captured = loader
    cache, downloaded, calls = modelscope_snapshot
    env["USE_MODELSCOPE"] = True
    if adapter_base:
        use_modelscope_adapter(env, cache, adapter_base)

    with pytest.raises(DispatchReached):
        env["from_pretrained"](model_name = "owner/model", **kwargs)

    assert "model.safetensors" in downloaded
    # An adapter's base is fetched from ModelScope as well (#3726).
    loaded = adapter_base or "owner/model"
    assert captured["dispatch"]["model_name"] == str(cache / loaded.replace("/", "--"))
    assert captured["dispatch"]["load_in_4bit"] is expected_4bit
    assert calls[-1] == ("owner/model", None)
    assert (loaded, None) in calls


def test_modelscope_adapter_base_missing_there_loads_from_the_hub(loader, modelscope_snapshot):
    """#3726: an adapter on ModelScope whose base is only on the Hub still loads that base by repo id."""
    env, captured = loader
    cache, downloaded, calls = modelscope_snapshot
    env["USE_MODELSCOPE"] = True
    use_modelscope_adapter(env, cache, "hf-only/base")

    with pytest.raises(DispatchReached):
        env["from_pretrained"](model_name = "owner/model", load_in_4bit = False, load_in_16bit = True)

    assert captured["dispatch"]["model_name"] == "hf-only/base"
    assert ("hf-only/base", None) in calls