[core] Mochi T2V (#9769)

* update

* udpate

* update transformer

* make style

* fix

* add conversion script

* update

* fix

* update

* fix

* update

* fixes

* make style

* update

* update

* update

* init

* update

* update

* add

* up

* up

* up

* update

* mochi transformer

* remove original implementation

* make style

* update inits

* update conversion script

* docs

* Update src/diffusers/pipelines/mochi/pipeline_mochi.py
Co-authored-by: Dhruv Nair <dhruv.nair@gmail.com>

* Update src/diffusers/pipelines/mochi/pipeline_mochi.py
Co-authored-by: Dhruv Nair <dhruv.nair@gmail.com>

* fix docs

* pipeline fixes

* make style

* invert sigmas in scheduler; fix pipeline

* fix pipeline num_frames

* flip proj and gate in swiglu

* make style

* fix

* make style

* fix tests

* latent mean and std fix

* update

* cherry-pick 1069d210e1b9e84a366cdc7a13965626ea258178

* remove additional sigma already handled by flow match scheduler

* fix

* remove hardcoded value

* replace conv1x1 with linear

* Update src/diffusers/pipelines/mochi/pipeline_mochi.py
Co-authored-by: Dhruv Nair <dhruv.nair@gmail.com>

* framewise decoding and conv_cache

* make style

* Apply suggestions from code review

* mochi vae encoder changes

* rebase correctly

* Update scripts/convert_mochi_to_diffusers.py

* fix tests

* fixes

* make style

* update

* make style

* update

* add framewise and tiled encoding

* make style

* make original vae implementation behaviour the default; note: framewise encoding does not work

* remove framewise encoding implementation due to presence of attn layers

* fight test 1

* fight test 2

---------
Co-authored-by: Dhruv Nair <dhruv.nair@gmail.com>
Co-authored-by: yiyixuxu <yixu310@gmail.com>

src/diffusers/schedulers/scheduling_flow_match_euler_discrete.py

@@ -71,6 +71,7 @@ class FlowMatchEulerDiscreteScheduler(SchedulerMixin, ConfigMixin):
         max_shift: Optional[float] = 1.15,
         base_image_seq_len: Optional[int] = 256,
         max_image_seq_len: Optional[int] = 4096,
+        invert_sigmas: bool = False,
     ):
         timesteps = np.linspace(1, num_train_timesteps, num_train_timesteps, dtype=np.float32)[::-1].copy()
         timesteps = torch.from_numpy(timesteps).to(dtype=torch.float32)
@@ -204,9 +205,15 @@ class FlowMatchEulerDiscreteScheduler(SchedulerMixin, ConfigMixin):
         sigmas = torch.from_numpy(sigmas).to(dtype=torch.float32, device=device)
         timesteps = sigmas * self.config.num_train_timesteps
 
-        self.timesteps = timesteps.to(device=device)
-        self.sigmas = torch.cat([sigmas, torch.zeros(1, device=sigmas.device)])
+        if self.config.invert_sigmas:
+            sigmas = 1.0 - sigmas
+            timesteps = sigmas * self.config.num_train_timesteps
+            sigmas = torch.cat([sigmas, torch.ones(1, device=sigmas.device)])
+        else:
+            sigmas = torch.cat([sigmas, torch.zeros(1, device=sigmas.device)])
 
+        self.timesteps = timesteps.to(device=device)
+        self.sigmas = sigmas
         self._step_index = None
         self._begin_index = None
 

src/diffusers/utils/dummy_pt_objects.py

@@ -92,6 +92,21 @@ class AutoencoderKLCogVideoX(metaclass=DummyObject):
         requires_backends(cls, ["torch"])
 
 
+class AutoencoderKLMochi(metaclass=DummyObject):
+    _backends = ["torch"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+    @classmethod
+    def from_config(cls, *args, **kwargs):
+        requires_backends(cls, ["torch"])
+
+    @classmethod
+    def from_pretrained(cls, *args, **kwargs):
+        requires_backends(cls, ["torch"])
+
+
 class AutoencoderKLTemporalDecoder(metaclass=DummyObject):
     _backends = ["torch"]
 
@@ -377,6 +392,21 @@ class LuminaNextDiT2DModel(metaclass=DummyObject):
         requires_backends(cls, ["torch"])
 
 
+class MochiTransformer3DModel(metaclass=DummyObject):
+    _backends = ["torch"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+    @classmethod
+    def from_config(cls, *args, **kwargs):
+        requires_backends(cls, ["torch"])
+
+    @classmethod
+    def from_pretrained(cls, *args, **kwargs):
+        requires_backends(cls, ["torch"])
+
+
 class ModelMixin(metaclass=DummyObject):
     _backends = ["torch"]
 

src/diffusers/utils/dummy_torch_and_transformers_objects.py

@@ -1052,6 +1052,21 @@ class MarigoldNormalsPipeline(metaclass=DummyObject):
         requires_backends(cls, ["torch", "transformers"])
 
 
+class MochiPipeline(metaclass=DummyObject):
+    _backends = ["torch", "transformers"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch", "transformers"])
+
+    @classmethod
+    def from_config(cls, *args, **kwargs):
+        requires_backends(cls, ["torch", "transformers"])
+
+    @classmethod
+    def from_pretrained(cls, *args, **kwargs):
+        requires_backends(cls, ["torch", "transformers"])
+
+
 class MusicLDMPipeline(metaclass=DummyObject):
     _backends = ["torch", "transformers"]
 

tests/models/transformers/test_models_transformer_mochi.py

+# coding=utf-8
+# Copyright 2024 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 unittest
+
+import torch
+
+from diffusers import MochiTransformer3DModel
+from diffusers.utils.testing_utils import enable_full_determinism, torch_device
+
+from ..test_modeling_common import ModelTesterMixin
+
+
+enable_full_determinism()
+
+
+class MochiTransformerTests(ModelTesterMixin, unittest.TestCase):
+    model_class = MochiTransformer3DModel
+    main_input_name = "hidden_states"
+    uses_custom_attn_processor = True
+
+    @property
+    def dummy_input(self):
+        batch_size = 2
+        num_channels = 4
+        num_frames = 2
+        height = 16
+        width = 16
+        embedding_dim = 16
+        sequence_length = 16
+
+        hidden_states = torch.randn((batch_size, num_channels, num_frames, height, width)).to(torch_device)
+        encoder_hidden_states = torch.randn((batch_size, sequence_length, embedding_dim)).to(torch_device)
+        encoder_attention_mask = torch.ones((batch_size, sequence_length)).bool().to(torch_device)
+        timestep = torch.randint(0, 1000, size=(batch_size,)).to(torch_device)
+
+        return {
+            "hidden_states": hidden_states,
+            "encoder_hidden_states": encoder_hidden_states,
+            "timestep": timestep,
+            "encoder_attention_mask": encoder_attention_mask,
+        }
+
+    @property
+    def input_shape(self):
+        return (4, 2, 16, 16)
+
+    @property
+    def output_shape(self):
+        return (4, 2, 16, 16)
+
+    def prepare_init_args_and_inputs_for_common(self):
+        init_dict = {
+            "patch_size": 2,
+            "num_attention_heads": 2,
+            "attention_head_dim": 8,
+            "num_layers": 2,
+            "pooled_projection_dim": 16,
+            "in_channels": 4,
+            "out_channels": None,
+            "qk_norm": "rms_norm",
+            "text_embed_dim": 16,
+            "time_embed_dim": 4,
+            "activation_fn": "swiglu",
+            "max_sequence_length": 16,
+        }
+        inputs_dict = self.dummy_input
+        return init_dict, inputs_dict
+
+    def test_gradient_checkpointing_is_applied(self):
+        expected_set = {"MochiTransformer3DModel"}
+        super().test_gradient_checkpointing_is_applied(expected_set=expected_set)

tests/pipelines/mochi/test_mochi.py

+# Copyright 2024 The HuggingFace Team.
+#
+# 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 gc
+import inspect
+import unittest
+
+import numpy as np
+import torch
+from transformers import AutoTokenizer, T5EncoderModel
+
+from diffusers import AutoencoderKLMochi, FlowMatchEulerDiscreteScheduler, MochiPipeline, MochiTransformer3DModel
+from diffusers.utils.testing_utils import (
+    enable_full_determinism,
+    numpy_cosine_similarity_distance,
+    require_torch_gpu,
+    slow,
+    torch_device,
+)
+
+from ..pipeline_params import TEXT_TO_IMAGE_BATCH_PARAMS, TEXT_TO_IMAGE_IMAGE_PARAMS, TEXT_TO_IMAGE_PARAMS
+from ..test_pipelines_common import PipelineTesterMixin, to_np
+
+
+enable_full_determinism()
+
+
+class MochiPipelineFastTests(PipelineTesterMixin, unittest.TestCase):
+    pipeline_class = MochiPipeline
+    params = TEXT_TO_IMAGE_PARAMS - {"cross_attention_kwargs"}
+    batch_params = TEXT_TO_IMAGE_BATCH_PARAMS
+    image_params = TEXT_TO_IMAGE_IMAGE_PARAMS
+    image_latents_params = TEXT_TO_IMAGE_IMAGE_PARAMS
+    required_optional_params = frozenset(
+        [
+            "num_inference_steps",
+            "generator",
+            "latents",
+            "return_dict",
+            "callback_on_step_end",
+            "callback_on_step_end_tensor_inputs",
+        ]
+    )
+    test_xformers_attention = False
+
+    def get_dummy_components(self):
+        torch.manual_seed(0)
+        transformer = MochiTransformer3DModel(
+            patch_size=2,
+            num_attention_heads=2,
+            attention_head_dim=8,
+            num_layers=2,
+            pooled_projection_dim=16,
+            in_channels=12,
+            out_channels=None,
+            qk_norm="rms_norm",
+            text_embed_dim=32,
+            time_embed_dim=4,
+            activation_fn="swiglu",
+            max_sequence_length=16,
+        )
+        transformer.pos_frequencies.data = transformer.pos_frequencies.new_full(transformer.pos_frequencies.shape, 0)
+
+        torch.manual_seed(0)
+        vae = AutoencoderKLMochi(
+            latent_channels=12,
+            out_channels=3,
+            encoder_block_out_channels=(32, 32, 32, 32),
+            decoder_block_out_channels=(32, 32, 32, 32),
+            layers_per_block=(1, 1, 1, 1, 1),
+        )
+
+        torch.manual_seed(0)
+        scheduler = FlowMatchEulerDiscreteScheduler()
+        text_encoder = T5EncoderModel.from_pretrained("hf-internal-testing/tiny-random-t5")
+        tokenizer = AutoTokenizer.from_pretrained("hf-internal-testing/tiny-random-t5")
+
+        components = {
+            "transformer": transformer,
+            "vae": vae,
+            "scheduler": scheduler,
+            "text_encoder": text_encoder,
+            "tokenizer": tokenizer,
+        }
+        return components
+
+    def get_dummy_inputs(self, device, seed=0):
+        if str(device).startswith("mps"):
+            generator = torch.manual_seed(seed)
+        else:
+            generator = torch.Generator(device=device).manual_seed(seed)
+        inputs = {
+            "prompt": "dance monkey",
+            "negative_prompt": "",
+            "generator": generator,
+            "num_inference_steps": 2,
+            "guidance_scale": 4.5,
+            "height": 16,
+            "width": 16,
+            # 6 * k + 1 is the recommendation
+            "num_frames": 7,
+            "max_sequence_length": 16,
+            "output_type": "pt",
+        }
+        return inputs
+
+    def test_inference(self):
+        device = "cpu"
+
+        components = self.get_dummy_components()
+        pipe = self.pipeline_class(**components)
+        pipe.to(device)
+        pipe.set_progress_bar_config(disable=None)
+
+        inputs = self.get_dummy_inputs(device)
+        video = pipe(**inputs).frames
+        generated_video = video[0]
+
+        self.assertEqual(generated_video.shape, (7, 3, 16, 16))
+        expected_video = torch.randn(7, 3, 16, 16)
+        max_diff = np.abs(generated_video - expected_video).max()
+        self.assertLessEqual(max_diff, 1e10)
+
+    def test_callback_inputs(self):
+        sig = inspect.signature(self.pipeline_class.__call__)
+        has_callback_tensor_inputs = "callback_on_step_end_tensor_inputs" in sig.parameters
+        has_callback_step_end = "callback_on_step_end" in sig.parameters
+
+        if not (has_callback_tensor_inputs and has_callback_step_end):
+            return
+
+        components = self.get_dummy_components()
+        pipe = self.pipeline_class(**components)
+        pipe = pipe.to(torch_device)
+        pipe.set_progress_bar_config(disable=None)
+        self.assertTrue(
+            hasattr(pipe, "_callback_tensor_inputs"),
+            f" {self.pipeline_class} should have `_callback_tensor_inputs` that defines a list of tensor variables its callback function can use as inputs",
+        )
+
+        def callback_inputs_subset(pipe, i, t, callback_kwargs):
+            # iterate over callback args
+            for tensor_name, tensor_value in callback_kwargs.items():
+                # check that we're only passing in allowed tensor inputs
+                assert tensor_name in pipe._callback_tensor_inputs
+
+            return callback_kwargs
+
+        def callback_inputs_all(pipe, i, t, callback_kwargs):
+            for tensor_name in pipe._callback_tensor_inputs:
+                assert tensor_name in callback_kwargs
+
+            # iterate over callback args
+            for tensor_name, tensor_value in callback_kwargs.items():
+                # check that we're only passing in allowed tensor inputs
+                assert tensor_name in pipe._callback_tensor_inputs
+
+            return callback_kwargs
+
+        inputs = self.get_dummy_inputs(torch_device)
+
+        # Test passing in a subset
+        inputs["callback_on_step_end"] = callback_inputs_subset
+        inputs["callback_on_step_end_tensor_inputs"] = ["latents"]
+        output = pipe(**inputs)[0]
+
+        # Test passing in a everything
+        inputs["callback_on_step_end"] = callback_inputs_all
+        inputs["callback_on_step_end_tensor_inputs"] = pipe._callback_tensor_inputs
+        output = pipe(**inputs)[0]
+
+        def callback_inputs_change_tensor(pipe, i, t, callback_kwargs):
+            is_last = i == (pipe.num_timesteps - 1)
+            if is_last:
+                callback_kwargs["latents"] = torch.zeros_like(callback_kwargs["latents"])
+            return callback_kwargs
+
+        inputs["callback_on_step_end"] = callback_inputs_change_tensor
+        inputs["callback_on_step_end_tensor_inputs"] = pipe._callback_tensor_inputs
+        output = pipe(**inputs)[0]
+        assert output.abs().sum() < 1e10
+
+    def test_inference_batch_single_identical(self):
+        self._test_inference_batch_single_identical(batch_size=3, expected_max_diff=1e-3)
+
+    def test_attention_slicing_forward_pass(
+        self, test_max_difference=True, test_mean_pixel_difference=True, expected_max_diff=1e-3
+    ):
+        if not self.test_attention_slicing:
+            return
+
+        components = self.get_dummy_components()
+        pipe = self.pipeline_class(**components)
+        for component in pipe.components.values():
+            if hasattr(component, "set_default_attn_processor"):
+                component.set_default_attn_processor()
+        pipe.to(torch_device)
+        pipe.set_progress_bar_config(disable=None)
+
+        generator_device = "cpu"
+        inputs = self.get_dummy_inputs(generator_device)
+        output_without_slicing = pipe(**inputs)[0]
+
+        pipe.enable_attention_slicing(slice_size=1)
+        inputs = self.get_dummy_inputs(generator_device)
+        output_with_slicing1 = pipe(**inputs)[0]
+
+        pipe.enable_attention_slicing(slice_size=2)
+        inputs = self.get_dummy_inputs(generator_device)
+        output_with_slicing2 = pipe(**inputs)[0]
+
+        if test_max_difference:
+            max_diff1 = np.abs(to_np(output_with_slicing1) - to_np(output_without_slicing)).max()
+            max_diff2 = np.abs(to_np(output_with_slicing2) - to_np(output_without_slicing)).max()
+            self.assertLess(
+                max(max_diff1, max_diff2),
+                expected_max_diff,
+                "Attention slicing should not affect the inference results",
+            )
+
+    def test_vae_tiling(self, expected_diff_max: float = 0.2):
+        generator_device = "cpu"
+        components = self.get_dummy_components()
+
+        pipe = self.pipeline_class(**components)
+        pipe.to("cpu")
+        pipe.set_progress_bar_config(disable=None)
+
+        # Without tiling
+        inputs = self.get_dummy_inputs(generator_device)
+        inputs["height"] = inputs["width"] = 128
+        output_without_tiling = pipe(**inputs)[0]
+
+        # With tiling
+        pipe.vae.enable_tiling(
+            tile_sample_min_height=96,
+            tile_sample_min_width=96,
+            tile_sample_stride_height=64,
+            tile_sample_stride_width=64,
+        )
+        inputs = self.get_dummy_inputs(generator_device)
+        inputs["height"] = inputs["width"] = 128
+        output_with_tiling = pipe(**inputs)[0]
+
+        self.assertLess(
+            (to_np(output_without_tiling) - to_np(output_with_tiling)).max(),
+            expected_diff_max,
+            "VAE tiling should not affect the inference results",
+        )
+
+
+@slow
+@require_torch_gpu
+class MochiPipelineIntegrationTests(unittest.TestCase):
+    prompt = "A painting of a squirrel eating a burger."
+
+    def setUp(self):
+        super().setUp()
+        gc.collect()
+        torch.cuda.empty_cache()
+
+    def tearDown(self):
+        super().tearDown()
+        gc.collect()
+        torch.cuda.empty_cache()
+
+    def test_cogvideox(self):
+        generator = torch.Generator("cpu").manual_seed(0)
+
+        pipe = MochiPipeline.from_pretrained("genmo/mochi-1-preview", torch_dtype=torch.float16)
+        pipe.enable_model_cpu_offload()
+        prompt = self.prompt
+
+        videos = pipe(
+            prompt=prompt,
+            height=480,
+            width=848,
+            num_frames=19,
+            generator=generator,
+            num_inference_steps=2,
+            output_type="pt",
+        ).frames
+
+        video = videos[0]
+        expected_video = torch.randn(1, 16, 480, 848, 3).numpy()
+
+        max_diff = numpy_cosine_similarity_distance(video, expected_video)
+        assert max_diff < 1e-3, f"Max diff is too high. got {video}"