Spaces:
Running
on
Zero
Running
on
Zero
File size: 16,506 Bytes
a49cc2f |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 |
import gc
import unittest
import numpy as np
import torch
import torch.nn as nn
from diffusers import (
AuraFlowPipeline,
AuraFlowTransformer2DModel,
FluxPipeline,
FluxTransformer2DModel,
GGUFQuantizationConfig,
SD3Transformer2DModel,
StableDiffusion3Pipeline,
)
from diffusers.utils.testing_utils import (
is_gguf_available,
nightly,
numpy_cosine_similarity_distance,
require_accelerate,
require_big_gpu_with_torch_cuda,
require_gguf_version_greater_or_equal,
torch_device,
)
if is_gguf_available():
from diffusers.quantizers.gguf.utils import GGUFLinear, GGUFParameter
@nightly
@require_big_gpu_with_torch_cuda
@require_accelerate
@require_gguf_version_greater_or_equal("0.10.0")
class GGUFSingleFileTesterMixin:
ckpt_path = None
model_cls = None
torch_dtype = torch.bfloat16
expected_memory_use_in_gb = 5
def test_gguf_parameters(self):
quant_storage_type = torch.uint8
quantization_config = GGUFQuantizationConfig(compute_dtype=self.torch_dtype)
model = self.model_cls.from_single_file(self.ckpt_path, quantization_config=quantization_config)
for param_name, param in model.named_parameters():
if isinstance(param, GGUFParameter):
assert hasattr(param, "quant_type")
assert param.dtype == quant_storage_type
def test_gguf_linear_layers(self):
quantization_config = GGUFQuantizationConfig(compute_dtype=self.torch_dtype)
model = self.model_cls.from_single_file(self.ckpt_path, quantization_config=quantization_config)
for name, module in model.named_modules():
if isinstance(module, torch.nn.Linear) and hasattr(module.weight, "quant_type"):
assert module.weight.dtype == torch.uint8
if module.bias is not None:
assert module.bias.dtype == torch.float32
def test_gguf_memory_usage(self):
quantization_config = GGUFQuantizationConfig(compute_dtype=self.torch_dtype)
model = self.model_cls.from_single_file(
self.ckpt_path, quantization_config=quantization_config, torch_dtype=self.torch_dtype
)
model.to("cuda")
assert (model.get_memory_footprint() / 1024**3) < self.expected_memory_use_in_gb
inputs = self.get_dummy_inputs()
torch.cuda.reset_peak_memory_stats()
torch.cuda.empty_cache()
with torch.no_grad():
model(**inputs)
max_memory = torch.cuda.max_memory_allocated()
assert (max_memory / 1024**3) < self.expected_memory_use_in_gb
def test_keep_modules_in_fp32(self):
r"""
A simple tests to check if the modules under `_keep_in_fp32_modules` are kept in fp32.
Also ensures if inference works.
"""
_keep_in_fp32_modules = self.model_cls._keep_in_fp32_modules
self.model_cls._keep_in_fp32_modules = ["proj_out"]
quantization_config = GGUFQuantizationConfig(compute_dtype=self.torch_dtype)
model = self.model_cls.from_single_file(self.ckpt_path, quantization_config=quantization_config)
for name, module in model.named_modules():
if isinstance(module, torch.nn.Linear):
if name in model._keep_in_fp32_modules:
assert module.weight.dtype == torch.float32
self.model_cls._keep_in_fp32_modules = _keep_in_fp32_modules
def test_dtype_assignment(self):
quantization_config = GGUFQuantizationConfig(compute_dtype=self.torch_dtype)
model = self.model_cls.from_single_file(self.ckpt_path, quantization_config=quantization_config)
with self.assertRaises(ValueError):
# Tries with a `dtype`
model.to(torch.float16)
with self.assertRaises(ValueError):
# Tries with a `device` and `dtype`
model.to(device="cuda:0", dtype=torch.float16)
with self.assertRaises(ValueError):
# Tries with a cast
model.float()
with self.assertRaises(ValueError):
# Tries with a cast
model.half()
# This should work
model.to("cuda")
def test_dequantize_model(self):
quantization_config = GGUFQuantizationConfig(compute_dtype=self.torch_dtype)
model = self.model_cls.from_single_file(self.ckpt_path, quantization_config=quantization_config)
model.dequantize()
def _check_for_gguf_linear(model):
has_children = list(model.children())
if not has_children:
return
for name, module in model.named_children():
if isinstance(module, nn.Linear):
assert not isinstance(module, GGUFLinear), f"{name} is still GGUFLinear"
assert not isinstance(module.weight, GGUFParameter), f"{name} weight is still GGUFParameter"
for name, module in model.named_children():
_check_for_gguf_linear(module)
class FluxGGUFSingleFileTests(GGUFSingleFileTesterMixin, unittest.TestCase):
ckpt_path = "https://huggingface.co/city96/FLUX.1-dev-gguf/blob/main/flux1-dev-Q2_K.gguf"
torch_dtype = torch.bfloat16
model_cls = FluxTransformer2DModel
expected_memory_use_in_gb = 5
def setUp(self):
gc.collect()
torch.cuda.empty_cache()
def tearDown(self):
gc.collect()
torch.cuda.empty_cache()
def get_dummy_inputs(self):
return {
"hidden_states": torch.randn((1, 4096, 64), generator=torch.Generator("cpu").manual_seed(0)).to(
torch_device, self.torch_dtype
),
"encoder_hidden_states": torch.randn(
(1, 512, 4096),
generator=torch.Generator("cpu").manual_seed(0),
).to(torch_device, self.torch_dtype),
"pooled_projections": torch.randn(
(1, 768),
generator=torch.Generator("cpu").manual_seed(0),
).to(torch_device, self.torch_dtype),
"timestep": torch.tensor([1]).to(torch_device, self.torch_dtype),
"img_ids": torch.randn((4096, 3), generator=torch.Generator("cpu").manual_seed(0)).to(
torch_device, self.torch_dtype
),
"txt_ids": torch.randn((512, 3), generator=torch.Generator("cpu").manual_seed(0)).to(
torch_device, self.torch_dtype
),
"guidance": torch.tensor([3.5]).to(torch_device, self.torch_dtype),
}
def test_pipeline_inference(self):
quantization_config = GGUFQuantizationConfig(compute_dtype=self.torch_dtype)
transformer = self.model_cls.from_single_file(
self.ckpt_path, quantization_config=quantization_config, torch_dtype=self.torch_dtype
)
pipe = FluxPipeline.from_pretrained(
"black-forest-labs/FLUX.1-dev", transformer=transformer, torch_dtype=self.torch_dtype
)
pipe.enable_model_cpu_offload()
prompt = "a cat holding a sign that says hello"
output = pipe(
prompt=prompt, num_inference_steps=2, generator=torch.Generator("cpu").manual_seed(0), output_type="np"
).images[0]
output_slice = output[:3, :3, :].flatten()
expected_slice = np.array(
[
0.47265625,
0.43359375,
0.359375,
0.47070312,
0.421875,
0.34375,
0.46875,
0.421875,
0.34765625,
0.46484375,
0.421875,
0.34179688,
0.47070312,
0.42578125,
0.34570312,
0.46875,
0.42578125,
0.3515625,
0.45507812,
0.4140625,
0.33984375,
0.4609375,
0.41796875,
0.34375,
0.45898438,
0.41796875,
0.34375,
]
)
max_diff = numpy_cosine_similarity_distance(expected_slice, output_slice)
assert max_diff < 1e-4
class SD35LargeGGUFSingleFileTests(GGUFSingleFileTesterMixin, unittest.TestCase):
ckpt_path = "https://huggingface.co/city96/stable-diffusion-3.5-large-gguf/blob/main/sd3.5_large-Q4_0.gguf"
torch_dtype = torch.bfloat16
model_cls = SD3Transformer2DModel
expected_memory_use_in_gb = 5
def setUp(self):
gc.collect()
torch.cuda.empty_cache()
def tearDown(self):
gc.collect()
torch.cuda.empty_cache()
def get_dummy_inputs(self):
return {
"hidden_states": torch.randn((1, 16, 64, 64), generator=torch.Generator("cpu").manual_seed(0)).to(
torch_device, self.torch_dtype
),
"encoder_hidden_states": torch.randn(
(1, 512, 4096),
generator=torch.Generator("cpu").manual_seed(0),
).to(torch_device, self.torch_dtype),
"pooled_projections": torch.randn(
(1, 2048),
generator=torch.Generator("cpu").manual_seed(0),
).to(torch_device, self.torch_dtype),
"timestep": torch.tensor([1]).to(torch_device, self.torch_dtype),
}
def test_pipeline_inference(self):
quantization_config = GGUFQuantizationConfig(compute_dtype=self.torch_dtype)
transformer = self.model_cls.from_single_file(
self.ckpt_path, quantization_config=quantization_config, torch_dtype=self.torch_dtype
)
pipe = StableDiffusion3Pipeline.from_pretrained(
"stabilityai/stable-diffusion-3.5-large", transformer=transformer, torch_dtype=self.torch_dtype
)
pipe.enable_model_cpu_offload()
prompt = "a cat holding a sign that says hello"
output = pipe(
prompt=prompt, num_inference_steps=2, generator=torch.Generator("cpu").manual_seed(0), output_type="np"
).images[0]
output_slice = output[:3, :3, :].flatten()
expected_slice = np.array(
[
0.17578125,
0.27539062,
0.27734375,
0.11914062,
0.26953125,
0.25390625,
0.109375,
0.25390625,
0.25,
0.15039062,
0.26171875,
0.28515625,
0.13671875,
0.27734375,
0.28515625,
0.12109375,
0.26757812,
0.265625,
0.16210938,
0.29882812,
0.28515625,
0.15625,
0.30664062,
0.27734375,
0.14648438,
0.29296875,
0.26953125,
]
)
max_diff = numpy_cosine_similarity_distance(expected_slice, output_slice)
assert max_diff < 1e-4
class SD35MediumGGUFSingleFileTests(GGUFSingleFileTesterMixin, unittest.TestCase):
ckpt_path = "https://huggingface.co/city96/stable-diffusion-3.5-medium-gguf/blob/main/sd3.5_medium-Q3_K_M.gguf"
torch_dtype = torch.bfloat16
model_cls = SD3Transformer2DModel
expected_memory_use_in_gb = 2
def setUp(self):
gc.collect()
torch.cuda.empty_cache()
def tearDown(self):
gc.collect()
torch.cuda.empty_cache()
def get_dummy_inputs(self):
return {
"hidden_states": torch.randn((1, 16, 64, 64), generator=torch.Generator("cpu").manual_seed(0)).to(
torch_device, self.torch_dtype
),
"encoder_hidden_states": torch.randn(
(1, 512, 4096),
generator=torch.Generator("cpu").manual_seed(0),
).to(torch_device, self.torch_dtype),
"pooled_projections": torch.randn(
(1, 2048),
generator=torch.Generator("cpu").manual_seed(0),
).to(torch_device, self.torch_dtype),
"timestep": torch.tensor([1]).to(torch_device, self.torch_dtype),
}
def test_pipeline_inference(self):
quantization_config = GGUFQuantizationConfig(compute_dtype=self.torch_dtype)
transformer = self.model_cls.from_single_file(
self.ckpt_path, quantization_config=quantization_config, torch_dtype=self.torch_dtype
)
pipe = StableDiffusion3Pipeline.from_pretrained(
"stabilityai/stable-diffusion-3.5-medium", transformer=transformer, torch_dtype=self.torch_dtype
)
pipe.enable_model_cpu_offload()
prompt = "a cat holding a sign that says hello"
output = pipe(
prompt=prompt, num_inference_steps=2, generator=torch.Generator("cpu").manual_seed(0), output_type="np"
).images[0]
output_slice = output[:3, :3, :].flatten()
expected_slice = np.array(
[
0.625,
0.6171875,
0.609375,
0.65625,
0.65234375,
0.640625,
0.6484375,
0.640625,
0.625,
0.6484375,
0.63671875,
0.6484375,
0.66796875,
0.65625,
0.65234375,
0.6640625,
0.6484375,
0.6328125,
0.6640625,
0.6484375,
0.640625,
0.67578125,
0.66015625,
0.62109375,
0.671875,
0.65625,
0.62109375,
]
)
max_diff = numpy_cosine_similarity_distance(expected_slice, output_slice)
assert max_diff < 1e-4
class AuraFlowGGUFSingleFileTests(GGUFSingleFileTesterMixin, unittest.TestCase):
ckpt_path = "https://huggingface.co/city96/AuraFlow-v0.3-gguf/blob/main/aura_flow_0.3-Q2_K.gguf"
torch_dtype = torch.bfloat16
model_cls = AuraFlowTransformer2DModel
expected_memory_use_in_gb = 4
def setUp(self):
gc.collect()
torch.cuda.empty_cache()
def tearDown(self):
gc.collect()
torch.cuda.empty_cache()
def get_dummy_inputs(self):
return {
"hidden_states": torch.randn((1, 4, 64, 64), generator=torch.Generator("cpu").manual_seed(0)).to(
torch_device, self.torch_dtype
),
"encoder_hidden_states": torch.randn(
(1, 512, 2048),
generator=torch.Generator("cpu").manual_seed(0),
).to(torch_device, self.torch_dtype),
"timestep": torch.tensor([1]).to(torch_device, self.torch_dtype),
}
def test_pipeline_inference(self):
quantization_config = GGUFQuantizationConfig(compute_dtype=self.torch_dtype)
transformer = self.model_cls.from_single_file(
self.ckpt_path, quantization_config=quantization_config, torch_dtype=self.torch_dtype
)
pipe = AuraFlowPipeline.from_pretrained(
"fal/AuraFlow-v0.3", transformer=transformer, torch_dtype=self.torch_dtype
)
pipe.enable_model_cpu_offload()
prompt = "a pony holding a sign that says hello"
output = pipe(
prompt=prompt, num_inference_steps=2, generator=torch.Generator("cpu").manual_seed(0), output_type="np"
).images[0]
output_slice = output[:3, :3, :].flatten()
expected_slice = np.array(
[
0.46484375,
0.546875,
0.64453125,
0.48242188,
0.53515625,
0.59765625,
0.47070312,
0.5078125,
0.5703125,
0.42773438,
0.50390625,
0.5703125,
0.47070312,
0.515625,
0.57421875,
0.45898438,
0.48632812,
0.53515625,
0.4453125,
0.5078125,
0.56640625,
0.47851562,
0.5234375,
0.57421875,
0.48632812,
0.5234375,
0.56640625,
]
)
max_diff = numpy_cosine_similarity_distance(expected_slice, output_slice)
assert max_diff < 1e-4
|