File size: 13,704 Bytes
dde5d93 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 |
# 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 AutoencoderKLCogVideoX, CogVideoXPipeline, CogVideoXTransformer3DModel, DDIMScheduler
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,
check_qkv_fusion_matches_attn_procs_length,
check_qkv_fusion_processors_exist,
to_np,
)
enable_full_determinism()
class CogVideoXPipelineFastTests(PipelineTesterMixin, unittest.TestCase):
pipeline_class = CogVideoXPipeline
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",
]
)
def get_dummy_components(self):
torch.manual_seed(0)
transformer = CogVideoXTransformer3DModel(
# Product of num_attention_heads * attention_head_dim must be divisible by 16 for 3D positional embeddings
# But, since we are using tiny-random-t5 here, we need the internal dim of CogVideoXTransformer3DModel
# to be 32. The internal dim is product of num_attention_heads and attention_head_dim
num_attention_heads=4,
attention_head_dim=8,
in_channels=4,
out_channels=4,
time_embed_dim=2,
text_embed_dim=32, # Must match with tiny-random-t5
num_layers=1,
sample_width=16, # latent width: 2 -> final width: 16
sample_height=16, # latent height: 2 -> final height: 16
sample_frames=9, # latent frames: (9 - 1) / 4 + 1 = 3 -> final frames: 9
patch_size=2,
temporal_compression_ratio=4,
max_text_seq_length=16,
)
torch.manual_seed(0)
vae = AutoencoderKLCogVideoX(
in_channels=3,
out_channels=3,
down_block_types=(
"CogVideoXDownBlock3D",
"CogVideoXDownBlock3D",
"CogVideoXDownBlock3D",
"CogVideoXDownBlock3D",
),
up_block_types=(
"CogVideoXUpBlock3D",
"CogVideoXUpBlock3D",
"CogVideoXUpBlock3D",
"CogVideoXUpBlock3D",
),
block_out_channels=(8, 8, 8, 8),
latent_channels=4,
layers_per_block=1,
norm_num_groups=2,
temporal_compression_ratio=4,
)
torch.manual_seed(0)
scheduler = DDIMScheduler()
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": 6.0,
# Cannot reduce because convolution kernel becomes bigger than sample
"height": 16,
"width": 16,
"num_frames": 8,
"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, (8, 3, 16, 16))
expected_video = torch.randn(8, 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_overlap_factor_height=1 / 12,
tile_overlap_factor_width=1 / 12,
)
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",
)
@unittest.skip("xformers attention processor does not exist for CogVideoX")
def test_xformers_attention_forwardGenerator_pass(self):
pass
def test_fused_qkv_projections(self):
device = "cpu" # ensure determinism for the device-dependent torch.Generator
components = self.get_dummy_components()
pipe = self.pipeline_class(**components)
pipe = pipe.to(device)
pipe.set_progress_bar_config(disable=None)
inputs = self.get_dummy_inputs(device)
frames = pipe(**inputs).frames # [B, F, C, H, W]
original_image_slice = frames[0, -2:, -1, -3:, -3:]
pipe.fuse_qkv_projections()
assert check_qkv_fusion_processors_exist(
pipe.transformer
), "Something wrong with the fused attention processors. Expected all the attention processors to be fused."
assert check_qkv_fusion_matches_attn_procs_length(
pipe.transformer, pipe.transformer.original_attn_processors
), "Something wrong with the attention processors concerning the fused QKV projections."
inputs = self.get_dummy_inputs(device)
frames = pipe(**inputs).frames
image_slice_fused = frames[0, -2:, -1, -3:, -3:]
pipe.transformer.unfuse_qkv_projections()
inputs = self.get_dummy_inputs(device)
frames = pipe(**inputs).frames
image_slice_disabled = frames[0, -2:, -1, -3:, -3:]
assert np.allclose(
original_image_slice, image_slice_fused, atol=1e-3, rtol=1e-3
), "Fusion of QKV projections shouldn't affect the outputs."
assert np.allclose(
image_slice_fused, image_slice_disabled, atol=1e-3, rtol=1e-3
), "Outputs, with QKV projection fusion enabled, shouldn't change when fused QKV projections are disabled."
assert np.allclose(
original_image_slice, image_slice_disabled, atol=1e-2, rtol=1e-2
), "Original outputs should match when fused QKV projections are disabled."
@slow
@require_torch_gpu
class CogVideoXPipelineIntegrationTests(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 = CogVideoXPipeline.from_pretrained("THUDM/CogVideoX-2b", torch_dtype=torch.float16)
pipe.enable_model_cpu_offload()
prompt = self.prompt
videos = pipe(
prompt=prompt,
height=480,
width=720,
num_frames=16,
generator=generator,
num_inference_steps=2,
output_type="pt",
).frames
video = videos[0]
expected_video = torch.randn(1, 16, 480, 720, 3).numpy()
max_diff = numpy_cosine_similarity_distance(video, expected_video)
assert max_diff < 1e-3, f"Max diff is too high. got {video}"
|