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.

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

import ast
from pathlib import Path
from types import SimpleNamespace

VISION_PATH = Path(__file__).resolve().parents[1] / "unsloth" / "models" / "vision.py"
TREE = ast.parse(VISION_PATH.read_text(encoding = "utf-8"))


def _module_constant(name):
    for node in TREE.body:
        if isinstance(node, ast.Assign) and any(
            isinstance(t, ast.Name) and t.id == name for t in node.targets
        ):
            return ast.literal_eval(node.value)
    raise AssertionError(f"{name} not found")


def _is_sparse_moe_config():
    for node in TREE.body:
        if isinstance(node, ast.FunctionDef) and node.name == "_is_sparse_moe_config":
            namespace = {}
            exec(
                compile(ast.Module(body = [node], type_ignores = []), str(VISION_PATH), "exec"),
                namespace,
            )
            return namespace["_is_sparse_moe_config"]
    raise AssertionError("_is_sparse_moe_config not found")


GEMMA4_MOE = SimpleNamespace(text_config = SimpleNamespace(num_experts = 128, enable_moe_block = True))
GEMMA4_DENSE = SimpleNamespace(
    text_config = SimpleNamespace(num_experts = None, enable_moe_block = False)
)
QWEN3_5_MOE = SimpleNamespace(text_config = SimpleNamespace(num_experts = 256))
QWEN3_5_DENSE = SimpleNamespace(text_config = SimpleNamespace(num_experts = None))


def test_moe_model_types_pass_the_allowlist():
    supported = _module_constant("VLLM_SUPPORTED_VLM")
    for arch in ("qwen3_5_moe", "gemma4", "gemma4_text"):
        assert arch in supported
    assert len(supported) == len(set(supported))


def test_sparse_moe_is_read_off_the_text_config():
    is_moe = _is_sparse_moe_config()
    assert is_moe(GEMMA4_MOE) is True
    assert is_moe(QWEN3_5_MOE) is True
    assert is_moe(GEMMA4_DENSE) is False
    assert is_moe(QWEN3_5_DENSE) is False


def test_a_config_without_text_config_is_read_directly():
    is_moe = _is_sparse_moe_config()
    assert is_moe(SimpleNamespace(num_experts = 128)) is True
    assert is_moe(SimpleNamespace(num_local_experts = 8)) is True
    assert is_moe(SimpleNamespace(n_routed_experts = 64)) is True
    assert is_moe(SimpleNamespace(enable_moe_block = True)) is True
    assert is_moe(SimpleNamespace()) is False


def test_dense_gemma4_is_refused_only_with_an_old_zoo():
    """Dense Gemma-4 shares the gemma4 model type; an old zoo leaves vLLM profiling its audio encoder."""
    assert set(_module_constant("VLLM_DENSE_NEEDS_ZOO_VLM")) == {"gemma4", "gemma4_text"}
    gate, parents = _gate("needs a newer unsloth_zoo", "_zoo_supports_dense_gemma4_fast_inference")
    # text_only = True sets is_vlm_config False, so the gate must not sit under it.
    assert not any("is_vlm_config" in test for test in parents)
    base = {
        "VLLM_DENSE_NEEDS_ZOO_VLM": _module_constant("VLLM_DENSE_NEEDS_ZOO_VLM"),
        "fast_inference": True,
    }
    old = {**base, "_zoo_supports_dense_gemma4_fast_inference": lambda: False}
    new = {**base, "_zoo_supports_dense_gemma4_fast_inference": lambda: True}
    assert _evaluate(gate, model_types = ["gemma4"], auto_config = GEMMA4_DENSE, **old)
    assert _evaluate(
        gate, model_types = ["gemma4_text", "gemma4"], auto_config = GEMMA4_DENSE.text_config, **old
    )
    assert not _evaluate(gate, model_types = ["gemma4"], auto_config = GEMMA4_DENSE, **new)
    assert not _evaluate(gate, model_types = ["gemma4"], auto_config = GEMMA4_MOE, **old)
    assert not _evaluate(gate, model_types = ["qwen3_5"], auto_config = QWEN3_5_DENSE, **old)
    assert not _evaluate(
        gate, **{**old, "fast_inference": False}, model_types = ["gemma4"], auto_config = GEMMA4_DENSE
    )


def test_the_4bit_refusal_uses_the_same_moe_test():
    source = VISION_PATH.read_text(encoding = "utf-8")
    assert "and _is_sparse_moe_config(model_config)" in source


def _gate(message, condition = ""):
    """The `if` raising `message` (test mentions `condition`) plus its enclosing `if` tests."""

    def walk(node, parents):
        for child in ast.iter_child_nodes(node):
            if (
                isinstance(child, ast.If)
                and condition in ast.unparse(child.test)
                and any(
                    isinstance(stmt, ast.Raise) and message in ast.unparse(stmt)
                    for stmt in child.body
                )
            ):
                return child, parents
            found = walk(
                child, parents + [ast.unparse(child.test)] if isinstance(child, ast.If) else parents
            )
            if found:
                return found
        return None

    found = walk(TREE, [])
    assert found, message
    return found


def _evaluate(node, **names):
    namespace = {
        "VLLM_ZOO_MOE_VLM": _module_constant("VLLM_ZOO_MOE_VLM"),
        "_is_sparse_moe_config": _is_sparse_moe_config(),
        **names,
    }
    return eval(compile(ast.Expression(node.test), str(VISION_PATH), "eval"), namespace)


def test_the_zoo_gate_covers_text_only_moe_loads():
    """text_only = True sets is_vlm_config False, so the zoo gate must not sit under it."""
    gate, parents = _gate("needs a newer unsloth_zoo", "_zoo_supports_moe_fast_inference")
    assert not any("is_vlm_config" in test for test in parents)
    old_zoo = lambda: False
    for model_types, config in (
        (["qwen3_5_moe"], QWEN3_5_MOE.text_config),
        (["gemma4_text", "gemma4"], GEMMA4_MOE.text_config),
        (["qwen3_5_moe"], QWEN3_5_MOE),
    ):
        refused = _evaluate(
            gate,
            fast_inference = True,
            model_types = model_types,
            auto_config = config,
            _zoo_supports_moe_fast_inference = old_zoo,
        )
        assert refused, model_types
    assert not _evaluate(
        gate,
        fast_inference = True,
        model_types = ["gemma4_text"],
        auto_config = GEMMA4_DENSE,
        _zoo_supports_moe_fast_inference = old_zoo,
    )
    assert not _evaluate(
        gate,
        fast_inference = True,
        model_types = ["gemma4"],
        auto_config = GEMMA4_MOE,
        _zoo_supports_moe_fast_inference = lambda: True,
    )
    assert not _evaluate(
        gate,
        fast_inference = False,
        model_types = ["gemma4"],
        auto_config = GEMMA4_MOE,
        _zoo_supports_moe_fast_inference = old_zoo,
    )


def test_bnb_moe_loads_are_refused_without_load_in_4bit():
    gate, _ = _gate("does not support bitsandbytes weights")

    def quant_type(method):
        return lambda config: method

    def refused(
        load_in_4bit,
        model_name,
        method,
        config = GEMMA4_MOE,
        model_types = ("gemma4",),
        load_in_8bit = False,
    ):
        return bool(
            _evaluate(
                gate,
                load_in_4bit = load_in_4bit,
                load_in_8bit = load_in_8bit,
                model_name = model_name,
                get_quant_type = quant_type(method),
                model_config = config,
                model_types = list(model_types),
            )
        )

    assert refused(True, "google/gemma-4-26B-A4B-it", None)
    assert refused(False, "unsloth/gemma-4-26B-A4B-it-bnb-4bit", None)
    assert refused(False, "someone/gemma-4-26b-moe-4bit", "bitsandbytes")
    assert refused(False, "Qwen/Qwen3.6-35B-A3B", "bitsandbytes", QWEN3_5_MOE, ("qwen3_5_moe",))
    # load_vllm has no 8-bit mode, so the experts would silently load in 16-bit.
    assert refused(
        False, "Qwen/Qwen3.6-35B-A3B", None, QWEN3_5_MOE, ("qwen3_5_moe",), load_in_8bit = True
    )
    assert not refused(False, "google/gemma-4-26B-A4B-it", None)
    assert not refused(False, "Qwen/Qwen3.6-35B-A3B-FP8", "fp8", QWEN3_5_MOE, ("qwen3_5_moe",))
    assert not refused(False, "unsloth/gemma-4-E2B-it-bnb-4bit", "bitsandbytes", GEMMA4_DENSE)


def test_bnb_refusal_covers_text_sparse_moe():
    """Qwen3-MoE 4-bit crashed in vLLM init ("output with shape [8, 1] ... [8, 768]"); vLLM serves no LoRA on bnb experts."""
    gate, _ = _gate("does not support bitsandbytes weights")

    def refused(
        config,
        model_types,
        load_in_4bit = True,
        method = None,
    ):
        return bool(
            _evaluate(
                gate,
                load_in_4bit = load_in_4bit,
                load_in_8bit = False,
                model_name = "unsloth/Qwen3-30B-A3B-Instruct-2507",
                get_quant_type = lambda config: method,
                model_config = config,
                model_types = list(model_types),
            )
        )

    qwen3_moe = SimpleNamespace(num_experts = 128)
    mixtral = SimpleNamespace(num_local_experts = 8)
    deepseek = SimpleNamespace(n_routed_experts = 64)
    for config, types in (
        (qwen3_moe, ("qwen3_moe",)),
        (mixtral, ("mixtral",)),
        (deepseek, ("deepseek_v3",)),
    ):
        assert refused(config, types), types
        assert not refused(config, types, load_in_4bit = False), types
        assert refused(config, types, load_in_4bit = False, method = "bitsandbytes"), types
    assert not refused(SimpleNamespace(num_experts = None), ("qwen3",))
    assert not refused(SimpleNamespace(num_experts = 0), ("qwen3",))


def _zoo_probe():
    for node in TREE.body:
        if (
            isinstance(node, ast.FunctionDef)
            and node.name == "_zoo_supports_dense_gemma4_fast_inference"
        ):
            namespace = {}
            exec(
                compile(ast.Module(body = [node], type_ignores = []), str(VISION_PATH), "exec"),
                namespace,
            )
            return namespace["_zoo_supports_dense_gemma4_fast_inference"]
    raise AssertionError("_zoo_supports_dense_gemma4_fast_inference not found")


def test_dense_gemma4_zoo_probe_reads_the_clamp_bound_templates(monkeypatch):
    import sys
    import types

    probe = _zoo_probe()
    bound = "model.vision_tower.encoder.layers.{kk}.self_attn.q_proj.input_min"

    def install(vision_layers):
        module = types.ModuleType("unsloth_zoo.empty_model")
        if vision_layers is not None:
            module.get_model_layer_config = lambda: {"vision_layers": set(vision_layers)}
        package = sys.modules.get("unsloth_zoo") or types.ModuleType("unsloth_zoo")
        monkeypatch.setitem(sys.modules, "unsloth_zoo", package)
        monkeypatch.setitem(sys.modules, "unsloth_zoo.empty_model", module)
        monkeypatch.setattr(package, "empty_model", module, raising = False)

    install({bound, "model.vision_tower.encoder.layers.{kk}.self_attn.q_proj"})
    assert probe() is True
    install({"model.vision_tower.encoder.layers.{kk}.self_attn.q_proj"})
    assert probe() is False
    install(None)
    assert probe() is False


def _new_vlm_probe():
    for node in TREE.body:
        if (
            isinstance(node, ast.FunctionDef)
            and node.name == "_zoo_supports_new_vlm_fast_inference"
        ):
            namespace = {}
            exec(
                compile(ast.Module(body = [node], type_ignores = []), str(VISION_PATH), "exec"),
                namespace,
            )
            return namespace["_zoo_supports_new_vlm_fast_inference"]
    raise AssertionError("_zoo_supports_new_vlm_fast_inference not found")


def test_new_vlm_zoo_gate_only_probes_unified_and_muse(monkeypatch):
    import sys
    import types

    probe = _new_vlm_probe()
    old_zoo = types.ModuleType("unsloth_zoo.empty_model")  # neither probed helper exists
    package = sys.modules.get("unsloth_zoo") or types.ModuleType("unsloth_zoo")
    monkeypatch.setitem(sys.modules, "unsloth_zoo", package)
    monkeypatch.setitem(sys.modules, "unsloth_zoo.empty_model", old_zoo)
    monkeypatch.setattr(package, "empty_model", old_zoo, raising = False)
    # An older zoo must not refuse models it already rebuilds.
    for model_types in (["qwen2_5_vl"], ["gemma4"], ["qwen3_5_moe"], ["llama"]):
        assert probe(model_types) is True, model_types
    for model_types in (["gemma4_unified"], ["gemma4_unified_text"], ["muse_glimmer_text"]):
        assert probe(model_types) is False, model_types


def test_new_vlm_zoo_gate_covers_text_only_loads():
    gate, parents = _gate("needs a newer unsloth_zoo", "_zoo_supports_new_vlm_fast_inference")
    assert not any("is_vlm_config" in test for test in parents)