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.

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

"""Fused LoRA kernels are skipped when lora_dropout or bias is set, and the summary must say why.

The fused kernels now run lora_dropout > 0 themselves, so the dropout clause only appears under
the UNSLOTH_FUSED_LORA_DROPOUT=0 kill switch, which these tests set.
"""

from __future__ import annotations

import ast
import inspect
import itertools
import textwrap

import pytest
import unsloth  # noqa: F401

from real_accelerator import has_real_accelerator
from unsloth.kernels import fast_lora
from unsloth.models import llama as llama_module
from unsloth.models.llama import _fused_lora_skip_reason


@pytest.fixture(autouse = True)
def _fused_dropout_off(monkeypatch):
    # lora_dropout > 0 only closes the fused gate under the UNSLOTH_FUSED_LORA_DROPOUT=0 kill switch.
    monkeypatch.setattr(fast_lora, "FUSED_LORA_DROPOUT", False)


def test_no_reason_when_the_fused_kernels_were_installed():
    assert _fused_lora_skip_reason(0, "none") == ""
    assert _fused_lora_skip_reason(0.0, "none") == ""


def test_lora_dropout_is_named_with_its_value():
    reason = _fused_lora_skip_reason(0.1, "none")
    assert "lora_dropout = 0.1" in reason
    assert "bias =" not in reason
    assert "Training is unaffected." in reason


def test_bias_is_named_with_its_value():
    reason = _fused_lora_skip_reason(0, "all")
    assert "bias = 'all'" in reason
    assert "lora_dropout" not in reason


def test_both_reasons_are_joined():
    reason = _fused_lora_skip_reason(0.05, "lora_only")
    assert "lora_dropout = 0.05" in reason
    assert "bias = 'lora_only'" in reason
    assert " and " in reason


def test_reason_appends_cleanly_to_the_summary_sentence():
    """The summary line ends in a full stop and the reason is appended straight onto it,
    so the reason must start with a space and must not start a new sentence mid word."""
    reason = _fused_lora_skip_reason(0.1, "none")
    assert reason.startswith(" ")
    assert reason.endswith(".")
    assert "  " not in ("...0 MLP layers." + reason)


@pytest.mark.parametrize(
    "lora_dropout,bias,float32_base,fsdp",
    list(
        itertools.product(
            [0, 0.0, 0.1, 0.5], ["none", "all", "lora_only"], [False, True], [False, True]
        )
    ),
)
def test_reason_is_non_empty_exactly_when_the_fused_gate_is_closed(
    lora_dropout, bias, float32_base, fsdp
):
    """Parity with the fused-kernel gate in patch_peft_model."""
    fused_installed = lora_dropout == 0 and bias == "none" and not float32_base and not fsdp
    reason = _fused_lora_skip_reason(lora_dropout, bias, float32_base, fsdp = fsdp)
    assert bool(reason) is not fused_installed


@pytest.mark.parametrize("lora_dropout", [0, 0.1, 0.5])
def test_lora_dropout_keeps_the_fused_gate_open_by_default(monkeypatch, lora_dropout):
    monkeypatch.setattr(fast_lora, "FUSED_LORA_DROPOUT", True)
    assert _fused_lora_skip_reason(lora_dropout, "none") == ""


def _gate_tests(source: str) -> list[str]:
    """Every `if` test in patch_peft_model that reads both lora_dropout and bias."""
    tree = ast.parse(textwrap.dedent(source))
    out = []
    for node in ast.walk(tree):
        if not isinstance(node, ast.If):
            continue
        names = {n.id for n in ast.walk(node.test) if isinstance(n, ast.Name)}
        if {"lora_dropout", "bias"} <= names:
            out.append(ast.unparse(node.test))
    return out


def test_patch_peft_model_still_gates_the_fused_kernels_on_the_same_values():
    """Every gate condition needs a matching _fused_lora_skip_reason clause."""
    source = inspect.getsource(llama_module.FastLlamaModel.patch_peft_model)
    assert _gate_tests(source) == [
        "(lora_dropout == 0 or _fast_lora.FUSED_LORA_DROPOUT) and bias == 'none' and (not float32_base) and (not fused_lora_declined_for_fsdp)"
    ]


def test_the_summary_call_carries_the_reason():
    source = inspect.getsource(llama_module.FastLlamaModel.patch_peft_model)
    flat = "".join(source.split())
    assert (
        "unfused_reason=_fused_lora_skip_reason(lora_dropout,bias,float32_base,fsdp=fused_lora_declined_for_fsdp)"
        in flat
    )
    assert "MLP layers.{unfused_reason}" in source


# has_real_accelerator(), not torch.cuda.is_available(): the zoo CUDA spoof patches the
# latter True process-wide before skipif runs.
@pytest.mark.gpu
@pytest.mark.skipif(
    not has_real_accelerator(),
    reason = "loads a real checkpoint through FastLanguageModel; needs an accelerator",
)
def test_summary_reason_is_logged_for_a_real_model():
    """unsloth#2076 end to end: the reason has to reach the user's console."""
    import logging

    from unsloth import FastLanguageModel

    records = []

    class _Capture(logging.Handler):
        def emit(self, record):
            records.append(record.getMessage())

    model, _tokenizer = FastLanguageModel.from_pretrained(
        "unsloth/Llama-3.2-1B-Instruct",
        max_seq_length = 512,
        load_in_4bit = True,
    )
    # warning_once (patched onto logging.Logger by transformers) dedupes process wide.
    getattr(llama_module.logger.warning_once, "cache_clear", lambda: None)()

    # The logger llama.py writes to, so transformers disabling propagation does not matter.
    handler = _Capture()
    target = llama_module.logger
    target.addHandler(handler)
    previous_level = target.level
    target.setLevel(logging.WARNING)
    try:
        model = FastLanguageModel.get_peft_model(
            model,
            r = 8,
            lora_alpha = 16,
            lora_dropout = 0.1,
            target_modules = ["q_proj", "k_proj", "v_proj", "o_proj"],
            random_state = 0,
        )
    finally:
        target.removeHandler(handler)
        target.setLevel(previous_level)

    summary = [m for m in records if "MLP layers." in m]
    assert summary, records
    assert "lora_dropout = 0.1" in summary[-1], summary[-1]