Spaces:
Running
on
Zero
Running
on
Zero
ZeroGPUCompiledModel
Browse files- app.py +5 -33
- utils/zerogpu.py +60 -0
app.py
CHANGED
@@ -19,9 +19,8 @@ import spaces
|
|
19 |
import torch
|
20 |
import torch._inductor
|
21 |
from diffusers import FluxPipeline
|
22 |
-
|
23 |
-
from
|
24 |
-
from torch.export.pt2_archive._package_weights import Weights
|
25 |
|
26 |
|
27 |
pipeline = FluxPipeline.from_pretrained('black-forest-labs/FLUX.1-schnell', torch_dtype=torch.bfloat16).to('cuda')
|
@@ -61,43 +60,16 @@ def compile_transformer():
|
|
61 |
|
62 |
exported = torch.export.export(pipeline.transformer, args=(), kwargs=transformer_kwargs)
|
63 |
|
64 |
-
|
65 |
-
'aot_inductor.package_constants_in_so': False,
|
66 |
-
'aot_inductor.package_constants_on_disk': True,
|
67 |
-
'aot_inductor.package': True,
|
68 |
-
})
|
69 |
-
|
70 |
-
files = [file for file in artifacts if isinstance(file, str)]
|
71 |
-
package_aoti(package_path, files)
|
72 |
-
|
73 |
-
weights, = (artifact for artifact in artifacts if isinstance(artifact, Weights))
|
74 |
-
weights_: dict[str, torch.Tensor] = {}
|
75 |
-
|
76 |
-
for name in weights:
|
77 |
-
tensor, _properties = weights.get_weight(name)
|
78 |
-
tensor_ = torch.empty_like(tensor, device='cpu').pin_memory()
|
79 |
-
weights_[name] = tensor_.copy_(tensor).detach().share_memory_()
|
80 |
|
81 |
-
return weights_
|
82 |
-
|
83 |
-
|
84 |
-
weights = compile_transformer()
|
85 |
-
weights = {name: tensor.to('cuda') for name, tensor in weights.items()}
|
86 |
-
print('compile_transformer', -(t0 - (t0 := datetime.now())))
|
87 |
|
88 |
transformer_config = pipeline.transformer.config
|
89 |
-
pipeline.transformer =
|
90 |
-
|
91 |
|
92 |
@spaces.GPU
|
93 |
def _generate_image(prompt: str, t0: datetime):
|
94 |
print('@spaces.GPU', -(t0 - (t0 := datetime.now())))
|
95 |
-
compiled_transformer: AOTICompiledModel = torch._inductor.aoti_load_package(package_path)
|
96 |
-
print('aoti_load_package', -(t0 - (t0 := datetime.now())))
|
97 |
-
compiled_transformer.load_constants(weights, check_full_update=True, user_managed=True)
|
98 |
-
print('load_constants', -(t0 - (t0 := datetime.now())))
|
99 |
-
pipeline.transformer = compiled_transformer
|
100 |
-
pipeline.transformer.config = transformer_config
|
101 |
images = []
|
102 |
for _ in range(4):
|
103 |
images += pipeline(prompt, num_inference_steps=4).images
|
|
|
19 |
import torch
|
20 |
import torch._inductor
|
21 |
from diffusers import FluxPipeline
|
22 |
+
|
23 |
+
from .utils.zerogpu import aoti_compile
|
|
|
24 |
|
25 |
|
26 |
pipeline = FluxPipeline.from_pretrained('black-forest-labs/FLUX.1-schnell', torch_dtype=torch.bfloat16).to('cuda')
|
|
|
60 |
|
61 |
exported = torch.export.export(pipeline.transformer, args=(), kwargs=transformer_kwargs)
|
62 |
|
63 |
+
return aoti_compile(exported, inductor_configs)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
64 |
|
|
|
|
|
|
|
|
|
|
|
|
|
65 |
|
66 |
transformer_config = pipeline.transformer.config
|
67 |
+
pipeline.transformer = compile_transformer()
|
68 |
+
pipeline.transformer.config = transformer_config
|
69 |
|
70 |
@spaces.GPU
|
71 |
def _generate_image(prompt: str, t0: datetime):
|
72 |
print('@spaces.GPU', -(t0 - (t0 := datetime.now())))
|
|
|
|
|
|
|
|
|
|
|
|
|
73 |
images = []
|
74 |
for _ in range(4):
|
75 |
images += pipeline(prompt, num_inference_steps=4).images
|
utils/zerogpu.py
ADDED
@@ -0,0 +1,60 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
"""
|
2 |
+
"""
|
3 |
+
from io import BytesIO
|
4 |
+
from typing import Any
|
5 |
+
|
6 |
+
import torch
|
7 |
+
from torch._inductor.package.package import package_aoti
|
8 |
+
from torch.export.pt2_archive._package import AOTICompiledModel
|
9 |
+
from torch.export.pt2_archive._package_weights import TensorProperties
|
10 |
+
from torch.export.pt2_archive._package_weights import Weights
|
11 |
+
|
12 |
+
|
13 |
+
INDUCTOR_CONFIGS_OVERRIDES = {
|
14 |
+
'aot_inductor.package_constants_in_so': False,
|
15 |
+
'aot_inductor.package_constants_on_disk': True,
|
16 |
+
'aot_inductor.package': True,
|
17 |
+
}
|
18 |
+
|
19 |
+
|
20 |
+
class ZeroGPUCompiledModel:
|
21 |
+
def __init__(self, archive_file: BytesIO, weights: Weights, cuda: bool = False):
|
22 |
+
self.archive_file = archive_file
|
23 |
+
self.weights = weights
|
24 |
+
if cuda:
|
25 |
+
self.weights_to_cuda_()
|
26 |
+
self.compiled_model: AOTICompiledModel | None = None
|
27 |
+
def weights_to_cuda_(self):
|
28 |
+
for name in self.weights:
|
29 |
+
tensor, properties = self.weights.get_weight(name)
|
30 |
+
self.weights[name] = (tensor.to('cuda'), properties)
|
31 |
+
def __call__(self, *args, **kwargs):
|
32 |
+
if self.compiled_model is None:
|
33 |
+
constants_map = {name: value[1] for name, value in self.weights.items()}
|
34 |
+
compiled_model: AOTICompiledModel = torch._inductor.aoti_load_package(self.archive_file)
|
35 |
+
compiled_model.load_constants(constants_map, check_full_update=True, user_managed=True)
|
36 |
+
self.compiled_model = compiled_model
|
37 |
+
return self.compiled_model(*args, **kwargs)
|
38 |
+
def __reduce__(self):
|
39 |
+
weight_dict: dict[str, tuple[torch.Tensor, TensorProperties]] = {}
|
40 |
+
for name in self.weights:
|
41 |
+
tensor, properties = self.weights.get_weight(name)
|
42 |
+
tensor_ = torch.empty_like(tensor, device='cpu').pin_memory()
|
43 |
+
weight_dict[name] = (tensor_.copy_(tensor).detach().share_memory_(), properties)
|
44 |
+
return ZeroGPUCompiledModel, (self.archive_file, Weights(weight_dict), True)
|
45 |
+
|
46 |
+
|
47 |
+
def aoti_compile(
|
48 |
+
exported_program: torch.export.ExportedProgram,
|
49 |
+
inductor_configs: dict[str, Any] | None = None,
|
50 |
+
):
|
51 |
+
inductor_configs = inductor_configs or {} | INDUCTOR_CONFIGS_OVERRIDES
|
52 |
+
gm = exported_program.module()
|
53 |
+
assert exported_program.example_inputs is not None
|
54 |
+
args, kwargs = exported_program.example_inputs
|
55 |
+
artifacts = torch._inductor.aot_compile(gm, args, kwargs, options=inductor_configs)
|
56 |
+
archive_file = BytesIO()
|
57 |
+
files = [file for file in artifacts if isinstance(file, str)]
|
58 |
+
package_aoti(archive_file, files)
|
59 |
+
weights, = (artifact for artifact in artifacts if isinstance(artifact, Weights))
|
60 |
+
return ZeroGPUCompiledModel(archive_file, weights)
|