|
| 1 | +# Copyright 2024 HuggingFace Inc. |
| 2 | +# |
| 3 | +# Licensed under the Apache License, Version 2.0 (the "License"); |
| 4 | +# you may not use this file except in compliance with the License. |
| 5 | +# You may obtain a copy of the License at |
| 6 | +# |
| 7 | +# http://www.apache.org/licenses/LICENSE-2.0 |
| 8 | +# |
| 9 | +# Unless required by applicable law or agreed to in writing, software |
| 10 | +# distributed under the License is distributed on an "AS IS" BASIS, |
| 11 | +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. |
| 12 | +# See the License for the specific language governing permissions and |
| 13 | +# limitations under the License. |
| 14 | + |
| 15 | +import unittest |
| 16 | + |
| 17 | +import torch |
| 18 | + |
| 19 | +from diffusers import HunyuanVideoFramepackTransformer3DModel |
| 20 | +from diffusers.utils.testing_utils import ( |
| 21 | + enable_full_determinism, |
| 22 | + torch_device, |
| 23 | +) |
| 24 | + |
| 25 | +from ..test_modeling_common import ModelTesterMixin |
| 26 | + |
| 27 | + |
| 28 | +enable_full_determinism() |
| 29 | + |
| 30 | + |
| 31 | +class HunyuanVideoTransformer3DTests(ModelTesterMixin, unittest.TestCase): |
| 32 | + model_class = HunyuanVideoFramepackTransformer3DModel |
| 33 | + main_input_name = "hidden_states" |
| 34 | + uses_custom_attn_processor = True |
| 35 | + model_split_percents = [0.5, 0.7, 0.9] |
| 36 | + |
| 37 | + @property |
| 38 | + def dummy_input(self): |
| 39 | + batch_size = 1 |
| 40 | + num_channels = 4 |
| 41 | + num_frames = 3 |
| 42 | + height = 4 |
| 43 | + width = 4 |
| 44 | + text_encoder_embedding_dim = 16 |
| 45 | + image_encoder_embedding_dim = 16 |
| 46 | + pooled_projection_dim = 8 |
| 47 | + sequence_length = 12 |
| 48 | + |
| 49 | + hidden_states = torch.randn((batch_size, num_channels, num_frames, height, width)).to(torch_device) |
| 50 | + encoder_hidden_states = torch.randn((batch_size, sequence_length, text_encoder_embedding_dim)).to(torch_device) |
| 51 | + pooled_projections = torch.randn((batch_size, pooled_projection_dim)).to(torch_device) |
| 52 | + encoder_attention_mask = torch.ones((batch_size, sequence_length)).to(torch_device) |
| 53 | + image_embeds = torch.randn((batch_size, sequence_length, image_encoder_embedding_dim)).to(torch_device) |
| 54 | + indices_latents = torch.ones((3,)).to(torch_device) |
| 55 | + latents_clean = torch.randn((batch_size, num_channels, num_frames - 1, height, width)).to(torch_device) |
| 56 | + indices_latents_clean = torch.ones((num_frames - 1,)).to(torch_device) |
| 57 | + latents_history_2x = torch.randn((batch_size, num_channels, num_frames - 1, height, width)).to(torch_device) |
| 58 | + indices_latents_history_2x = torch.ones((num_frames - 1,)).to(torch_device) |
| 59 | + latents_history_4x = torch.randn((batch_size, num_channels, (num_frames - 1) * 4, height, width)).to( |
| 60 | + torch_device |
| 61 | + ) |
| 62 | + indices_latents_history_4x = torch.ones(((num_frames - 1) * 4,)).to(torch_device) |
| 63 | + timestep = torch.randint(0, 1000, size=(batch_size,)).to(torch_device) |
| 64 | + guidance = torch.randint(0, 1000, size=(batch_size,)).to(torch_device) |
| 65 | + |
| 66 | + return { |
| 67 | + "hidden_states": hidden_states, |
| 68 | + "timestep": timestep, |
| 69 | + "encoder_hidden_states": encoder_hidden_states, |
| 70 | + "pooled_projections": pooled_projections, |
| 71 | + "encoder_attention_mask": encoder_attention_mask, |
| 72 | + "guidance": guidance, |
| 73 | + "image_embeds": image_embeds, |
| 74 | + "indices_latents": indices_latents, |
| 75 | + "latents_clean": latents_clean, |
| 76 | + "indices_latents_clean": indices_latents_clean, |
| 77 | + "latents_history_2x": latents_history_2x, |
| 78 | + "indices_latents_history_2x": indices_latents_history_2x, |
| 79 | + "latents_history_4x": latents_history_4x, |
| 80 | + "indices_latents_history_4x": indices_latents_history_4x, |
| 81 | + } |
| 82 | + |
| 83 | + @property |
| 84 | + def input_shape(self): |
| 85 | + return (4, 3, 4, 4) |
| 86 | + |
| 87 | + @property |
| 88 | + def output_shape(self): |
| 89 | + return (4, 3, 4, 4) |
| 90 | + |
| 91 | + def prepare_init_args_and_inputs_for_common(self): |
| 92 | + init_dict = { |
| 93 | + "in_channels": 4, |
| 94 | + "out_channels": 4, |
| 95 | + "num_attention_heads": 2, |
| 96 | + "attention_head_dim": 10, |
| 97 | + "num_layers": 1, |
| 98 | + "num_single_layers": 1, |
| 99 | + "num_refiner_layers": 1, |
| 100 | + "patch_size": 2, |
| 101 | + "patch_size_t": 1, |
| 102 | + "guidance_embeds": True, |
| 103 | + "text_embed_dim": 16, |
| 104 | + "pooled_projection_dim": 8, |
| 105 | + "rope_axes_dim": (2, 4, 4), |
| 106 | + "image_condition_type": None, |
| 107 | + "has_image_proj": True, |
| 108 | + "image_proj_dim": 16, |
| 109 | + "has_clean_x_embedder": True, |
| 110 | + } |
| 111 | + inputs_dict = self.dummy_input |
| 112 | + return init_dict, inputs_dict |
| 113 | + |
| 114 | + def test_gradient_checkpointing_is_applied(self): |
| 115 | + expected_set = {"HunyuanVideoFramepackTransformer3DModel"} |
| 116 | + super().test_gradient_checkpointing_is_applied(expected_set=expected_set) |
0 commit comments