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.

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

"""Processor selection for VLMs registering only AutoModel/AutoModelForCausalLM, using stub classes."""

import ast
import importlib.util
import types
from pathlib import Path

import pytest

transformers = pytest.importorskip("transformers")

VISION = Path(__file__).parents[1] / "unsloth" / "models" / "vision.py"
if not VISION.exists():  # copied out of the repo: read the tree on PYTHONPATH
    VISION = (
        Path(importlib.util.find_spec("unsloth").submodule_search_locations[0])
        / "models"
        / "vision.py"
    )
_SELECTION_NAMES = {
    "is_vlm",
    "is_whisper",
    "needs_processor",
    "is_vlm_config",
    "auto_processor",
    "try_repo_processor",
}


def _from_pretrained():
    tree = ast.parse(VISION.read_text(encoding = "utf-8"))
    cls = next(n for n in tree.body if isinstance(n, ast.ClassDef) and n.name == "FastBaseModel")
    return next(
        n for n in cls.body if isinstance(n, ast.FunctionDef) and n.name == "from_pretrained"
    )


def _selection_nodes():
    nodes = []
    for node in _from_pretrained().body:
        if isinstance(node, ast.Assign) and any(
            getattr(t, "id", None) in _SELECTION_NAMES for t in node.targets
        ):
            nodes.append(node)
        elif isinstance(node, ast.FunctionDef) and node.name == "_acquire_processor":
            nodes.append(node)
    return nodes


class _Tokenizer:
    padding_side = "left"


class _Processor:
    def __init__(self):
        self.tokenizer = _Tokenizer()
        self.image_processor = object()


def _stub(result, calls, name):
    def from_pretrained(path, **kwargs):
        calls.append((name, kwargs))
        if isinstance(result, Exception):
            raise result
        return result() if isinstance(result, type) else result

    return type(name, (), {"from_pretrained": staticmethod(from_pretrained)})


def _select(
    auto_processor_result,
    *,
    has_vision_config = True,
    text_only = False,
    auto_model = transformers.AutoModelForCausalLM,
):
    calls = []
    fake_processor = _stub(auto_processor_result, calls, "AutoProcessor")
    fake_tokenizer = _stub(_Tokenizer, calls, "AutoTokenizer")
    config = types.SimpleNamespace()
    if has_vision_config:
        config.vision_config = types.SimpleNamespace()
    namespace = {
        "auto_model": auto_model,
        "auto_config": config,
        "text_only": text_only,
        "whisper_language": None,
        "whisper_task": None,
        "AutoModelForVision2Seq": getattr(
            transformers, "AutoModelForVision2Seq", transformers.AutoModelForImageTextToText
        ),
        "AutoModelForImageTextToText": transformers.AutoModelForImageTextToText,
        "AutoModelForSpeechSeq2Seq": transformers.AutoModelForSpeechSeq2Seq,
        "_multimodal_auto_classes": lambda: (),
        "_audio_lm_class": lambda *args, **kwargs: None,
        "_revision": None,
        "local_files_only": False,
        "_config_maps_audio_auto_class": lambda config: False,
        "_from_pretrained_left_padded": lambda loader, name, **kw: loader.from_pretrained(
            name, padding_side = "left", **kw
        ),
        "AutoProcessor": fake_processor,
        "AutoTokenizer": fake_tokenizer,
        "tokenizer_name": "org/repo",
        "token": None,
        "trust_remote_code": True,
        "kwargs": {},
        "_tokenizer_revision": "abc123",
        "model_type_arch": "remote_vlm",
        "get_auto_processor": lambda *a, **k: None,
        "_construct_vlm_processor_fallback": lambda *a, **k: (None, None),
        "_is_offline_related_error": lambda e: False,
    }
    module = ast.Module(body = _selection_nodes(), type_ignores = [])
    ast.fix_missing_locations(module)
    exec(compile(module, str(VISION), "exec"), namespace)
    tok, err = namespace["_acquire_processor"](False)
    return tok, err, calls


def test_repo_autoprocessor_is_used_for_automodel_only_vlm():
    """Nemotron-3-Nano-Omni."""
    tok, err, calls = _select(_Processor)
    assert isinstance(tok, _Processor), f"got {type(tok).__name__}, calls={calls}"
    assert hasattr(tok, "image_processor")
    name, kwargs = calls[0]
    assert name == "AutoProcessor"
    assert kwargs["padding_side"] == "left"
    assert kwargs["revision"] == "abc123"
    assert kwargs["trust_remote_code"] is True


def test_missing_autoprocessor_falls_back_to_tokenizer():
    """DeepSeek-OCR."""
    tok, err, calls = _select(ValueError("Unrecognized processing class"))
    assert isinstance(tok, _Tokenizer)
    assert err is None
    assert [c[0] for c in calls][-1] == "AutoTokenizer"


def test_autoprocessor_that_resolves_to_a_tokenizer_is_not_kept():
    """Nemotron-Nano-VL."""
    tok, err, calls = _select(_Tokenizer)
    assert isinstance(tok, _Tokenizer)
    assert [c[0] for c in calls][-1] == "AutoTokenizer"


def test_automodel_only_vlm_uses_repo_autoprocessor():
    """DeepSeek-OCR style auto_map with only AutoModel."""
    tok, err, calls = _select(_Processor, auto_model = transformers.AutoModel)
    assert isinstance(tok, _Processor)


def test_sequence_classification_vlm_keeps_tokenizer():
    """Gemma 3 with num_labels: native multimodal config, text classification head."""
    tok, err, calls = _select(
        _Processor, auto_model = transformers.AutoModelForSequenceClassification
    )
    assert isinstance(tok, _Tokenizer)
    assert [c[0] for c in calls] == ["AutoTokenizer"]


def test_text_model_never_tries_autoprocessor():
    tok, err, calls = _select(_Processor, has_vision_config = False)
    assert isinstance(tok, _Tokenizer)
    assert [c[0] for c in calls] == ["AutoTokenizer"]


def test_text_only_vlm_never_tries_autoprocessor():
    tok, err, calls = _select(_Processor, text_only = True)
    assert isinstance(tok, _Tokenizer)
    assert [c[0] for c in calls] == ["AutoTokenizer"]