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.

6.9 KB
Raw Blame
# SPDX-License-Identifier: AGPL-3.0-only
# Copyright 2023-present Daniel Han-Chen & the Unsloth team. All rights reserved.

"""Remote decoder loops that bypass the checkpoint function must still checkpoint when enabled."""

import sys

import pytest
import torch
import torch.nn as nn

transformers = pytest.importorskip("transformers")
from transformers import PretrainedConfig, PreTrainedModel

from unsloth.models.loader_utils import install_remote_gradient_checkpointing

REMOTE = "transformers_modules.some_repo.modeling_x"


class RemoteLayer(nn.Module):
    calls = 0

    def __init__(self, hidden):
        super().__init__()
        self.up = nn.Linear(hidden, 4 * hidden)
        self.down = nn.Linear(4 * hidden, hidden)

    def forward(
        self,
        hidden_states,
        attention_mask = None,
        past_key_value = None,
        use_cache = False,
    ):
        type(self).calls += 1
        return (hidden_states + self.down(torch.relu(self.up(hidden_states))),)


class RemoteBackbone(PreTrainedModel):
    config_class = PretrainedConfig
    supports_gradient_checkpointing = True

    def __init__(
        self,
        config,
        layer_cls = RemoteLayer,
    ):
        super().__init__(config)
        self.layers = nn.ModuleList(layer_cls(16) for _ in range(3))
        self.gradient_checkpointing = False

    def forward(
        self,
        hidden_states,
        past_key_values = None,
    ):
        for layer in self.layers:
            hidden_states = layer(
                hidden_states, attention_mask = None, past_key_value = past_key_values
            )[0]
        return hidden_states


for _cls in (RemoteLayer, RemoteBackbone):
    _cls.__module__ = REMOTE
# transformers reads the defining module's source for some class checks.
sys.modules.setdefault(REMOTE, sys.modules[__name__])


def _build(layer_cls = RemoteLayer):
    torch.manual_seed(0)
    return RemoteBackbone(PretrainedConfig(), layer_cls = layer_cls)


def _step(model, **kwargs):
    model.train()
    x = torch.randn(2, 8, 16, requires_grad = True)
    before = type(model.layers[0]).calls
    model(x, **kwargs).square().sum().backward()
    grads = [p.grad.clone() for p in model.parameters()]
    model.zero_grad()
    return type(model.layers[0]).calls - before, grads


def test_the_flag_really_checkpoints_a_remote_loop():
    reference = _build()
    install_remote_gradient_checkpointing(reference, verbose = False)
    calls_off, grads_off = _step(reference)
    assert calls_off == 3

    model = _build()
    install_remote_gradient_checkpointing(model, verbose = False)
    assert RemoteLayer.__dict__["forward"]._unsloth_manual_checkpoint
    model.gradient_checkpointing_enable(gradient_checkpointing_kwargs = {"use_reentrant": True})
    calls_on, grads_on = _step(model)
    assert calls_on == 6  # every layer recomputed in the backward
    assert all(torch.allclose(a, b) for a, b in zip(grads_off, grads_on))


def test_the_installed_checkpoint_function_is_used():
    model = _build()
    install_remote_gradient_checkpointing(model, verbose = False)
    model.gradient_checkpointing_enable()
    seen = []
    original = model._gradient_checkpointing_func

    def spy(function, *args):
        seen.append(len(args))
        return original(function, *args)

    for module in model.modules():
        if hasattr(module, "gradient_checkpointing"):
            module._gradient_checkpointing_func = spy
    _step(model)
    assert seen == [1, 1, 1]


def test_a_cache_is_never_recomputed():
    model = _build()
    install_remote_gradient_checkpointing(model, verbose = False)
    model.gradient_checkpointing_enable()
    calls, _ = _step(model, past_key_values = object())
    assert calls == 3


def test_a_loop_that_already_checkpoints_is_left_alone():
    class CheckpointingBackbone(RemoteBackbone):
        def forward(
            self,
            hidden_states,
            past_key_values = None,
        ):
            for layer in self.layers:
                hidden_states = self._gradient_checkpointing_func(layer.__call__, hidden_states)[0]
            return hidden_states

    CheckpointingBackbone.__module__ = REMOTE
    torch.manual_seed(0)
    model = CheckpointingBackbone(PretrainedConfig())
    assert install_remote_gradient_checkpointing(model, verbose = False) == []


def test_native_models_are_never_wrapped():
    class NativeLayer(nn.Module):
        def __init__(self, hidden):
            super().__init__()
            self.up = nn.Linear(hidden, hidden)

        def forward(self, hidden_states, **kwargs):
            return (self.up(hidden_states),)

    class NativeBackbone(RemoteBackbone):
        pass

    native = "transformers.models.x.modeling_x"
    NativeLayer.__module__ = NativeBackbone.__module__ = native
    sys.modules.setdefault(native, sys.modules[__name__])
    model = NativeBackbone(PretrainedConfig(), layer_cls = NativeLayer)
    assert install_remote_gradient_checkpointing(model, verbose = False) == []
    assert not hasattr(NativeLayer.forward, "_unsloth_manual_checkpoint")


class ResidualLayer(RemoteLayer):
    """Kimi-K3's attention-residual layers: a second tensor that needs a gradient is carried
    from layer to layer as a keyword."""

    def forward(
        self,
        hidden_states,
        attention_mask = None,
        past_key_value = None,
        block_residual = None,
    ):
        type(self).calls += 1
        out = hidden_states + self.down(torch.relu(self.up(hidden_states + block_residual)))
        return out, block_residual + out


class ResidualBackbone(RemoteBackbone):
    def forward(
        self,
        hidden_states,
        past_key_values = None,
    ):
        block_residual = hidden_states * 0.5
        for layer in self.layers:
            hidden_states, block_residual = layer(
                hidden_states,
                attention_mask = None,
                past_key_value = past_key_values,
                block_residual = block_residual,
            )
        return hidden_states + block_residual


for _cls in (ResidualLayer, ResidualBackbone):
    _cls.__module__ = REMOTE


@pytest.mark.parametrize("use_reentrant", [True, False])
def test_a_keyword_tensor_carried_across_layers_is_recomputed_once(use_reentrant):
    def build():
        torch.manual_seed(0)
        return ResidualBackbone(PretrainedConfig(), layer_cls = ResidualLayer)

    reference = build()
    install_remote_gradient_checkpointing(reference, verbose = False)
    calls_off, grads_off = _step(reference)
    assert calls_off == 3

    model = build()
    install_remote_gradient_checkpointing(model, verbose = False)
    model.gradient_checkpointing_enable(
        gradient_checkpointing_kwargs = {"use_reentrant": use_reentrant}
    )
    calls_on, grads_on = _step(model)
    # A closure-held block_residual tied each recompute to earlier layers' graphs.
    assert calls_on == 6
    assert all(torch.allclose(a, b, atol = 1e-6) for a, b in zip(grads_off, grads_on))