diffusers/tests/models/transformers/test_models_transformer_chroma.py

# coding=utf-8
# Copyright 2026 HuggingFace Inc.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

import torch

from diffusers import ChromaTransformer2DModel
from diffusers.utils.torch_utils import randn_tensor

from ...testing_utils import enable_full_determinism, torch_device
from ..testing_utils import (
    AttentionTesterMixin,
    BaseModelTesterConfig,
    LoraHotSwappingForModelTesterMixin,
    LoraTesterMixin,
    ModelTesterMixin,
    SingleFileTesterMixin,
    TrainingTesterMixin,
)


enable_full_determinism()


class ChromaTransformerTesterConfig(BaseModelTesterConfig):
    @property
    def model_class(self):
        return ChromaTransformer2DModel

    @property
    def main_input_name(self) -> str:
        return "hidden_states"

    @property
    def model_split_percents(self) -> list:
        return [0.8, 0.7, 0.7]

    @property
    def output_shape(self) -> tuple:
        return (16, 4)

    @property
    def input_shape(self) -> tuple:
        return (16, 4)

    @property
    def generator(self):
        return torch.Generator("cpu").manual_seed(0)

    def get_init_dict(self) -> dict:
        return {
            "patch_size": 1,
            "in_channels": 4,
            "num_layers": 1,
            "num_single_layers": 1,
            "attention_head_dim": 16,
            "num_attention_heads": 2,
            "joint_attention_dim": 32,
            "axes_dims_rope": [4, 4, 8],
            "approximator_num_channels": 8,
            "approximator_hidden_dim": 16,
            "approximator_layers": 1,
        }

    def get_dummy_inputs(self, batch_size: int = 1) -> dict[str, torch.Tensor]:
        num_latent_channels = 4
        num_image_channels = 3
        height = width = 4
        sequence_length = 48
        embedding_dim = 32

        return {
            "hidden_states": randn_tensor(
                (batch_size, height * width, num_latent_channels), generator=self.generator, device=torch_device
            ),
            "encoder_hidden_states": randn_tensor(
                (batch_size, sequence_length, embedding_dim), generator=self.generator, device=torch_device
            ),
            "img_ids": randn_tensor(
                (height * width, num_image_channels), generator=self.generator, device=torch_device
            ),
            "txt_ids": randn_tensor(
                (sequence_length, num_image_channels), generator=self.generator, device=torch_device
            ),
            "timestep": torch.tensor([1.0]).to(torch_device).expand(batch_size),
        }


class TestChromaTransformer(ChromaTransformerTesterConfig, ModelTesterMixin):
    def test_deprecated_inputs_img_txt_ids_3d(self):
        init_dict = self.get_init_dict()
        inputs_dict = self.get_dummy_inputs()

        model = self.model_class(**init_dict)
        model.to(torch_device)
        model.eval()

        with torch.no_grad():
            output_1 = model(**inputs_dict).to_tuple()[0]

        text_ids_3d = inputs_dict["txt_ids"].unsqueeze(0)
        image_ids_3d = inputs_dict["img_ids"].unsqueeze(0)

        assert text_ids_3d.ndim == 3
        assert image_ids_3d.ndim == 3

        inputs_dict["txt_ids"] = text_ids_3d
        inputs_dict["img_ids"] = image_ids_3d

        with torch.no_grad():
            output_2 = model(**inputs_dict).to_tuple()[0]

        assert output_1.shape == output_2.shape
        assert torch.allclose(output_1, output_2, atol=1e-5), (
            "output with deprecated inputs (img_ids and txt_ids as 3d torch tensors) "
            "are not equal as them as 2d inputs"
        )


class TestChromaTransformerAttention(ChromaTransformerTesterConfig, AttentionTesterMixin):
    """Attention processor tests for Chroma Transformer."""


class TestChromaTransformerTraining(ChromaTransformerTesterConfig, TrainingTesterMixin):
    def test_gradient_checkpointing_is_applied(self):
        expected_set = {"ChromaTransformer2DModel"}
        super().test_gradient_checkpointing_is_applied(expected_set=expected_set)


class TestChromaTransformerLoRA(ChromaTransformerTesterConfig, LoraTesterMixin):
    pass


class TestChromaTransformerLoRAHotSwap(ChromaTransformerTesterConfig, LoraHotSwappingForModelTesterMixin):
    @property
    def different_shapes_for_compilation(self):
        return [(4, 4), (4, 8), (8, 8)]

    def get_dummy_inputs(self, height: int = 4, width: int = 4) -> dict[str, torch.Tensor]:
        batch_size = 1
        num_latent_channels = 4
        num_image_channels = 3
        sequence_length = 24
        embedding_dim = 32

        return {
            "hidden_states": randn_tensor((batch_size, height * width, num_latent_channels), device=torch_device),
            "encoder_hidden_states": randn_tensor((batch_size, sequence_length, embedding_dim), device=torch_device),
            "img_ids": randn_tensor((height * width, num_image_channels), device=torch_device),
            "txt_ids": randn_tensor((sequence_length, num_image_channels), device=torch_device),
            "timestep": torch.tensor([1.0]).to(torch_device).expand(batch_size),
        }


class TestChromaTransformer2DSingleFile(ChromaTransformerTesterConfig, SingleFileTesterMixin):
    @property
    def ckpt_path(self):
        return "https://huggingface.co/lodestones/Chroma1-HD/blob/main/Chroma1-HD.safetensors"

    @property
    def pretrained_model_name_or_path(self):
        return "lodestones/Chroma1-HD"

    @property
    def pretrained_model_kwargs(self):
        return {"subfolder": "transformer"}