Spaces:
Running
on
Zero
Running
on
Zero
File size: 3,924 Bytes
dc155d4 59e1cbc dc155d4 59e1cbc dc155d4 |
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 |
"""
"""
import contextlib
from contextvars import ContextVar
from io import BytesIO
from typing import Any
from typing import cast
from unittest.mock import patch
import torch
from torch._inductor.package.package import package_aoti
from torch.export.pt2_archive._package import AOTICompiledModel
from torch.export.pt2_archive._package_weights import Weights
INDUCTOR_CONFIGS_OVERRIDES = {
'aot_inductor.package_constants_in_so': False,
'aot_inductor.package_constants_on_disk': True,
'aot_inductor.package': True,
}
class ZeroGPUWeights:
def __init__(self, constants_map: dict[str, torch.Tensor], to_cuda: bool = False):
if to_cuda:
self.constants_map = {name: tensor.to('cuda') for name, tensor in constants_map.items()}
else:
self.constants_map = constants_map
def __reduce__(self):
print('ZeroGPUWeights.__reduce__', f'{id(self)=}', sum([tensor.nbytes for tensor in self.constants_map.values()]))
constants_map: dict[str, torch.Tensor] = {}
for name, tensor in self.constants_map.items():
tensor_ = torch.empty_like(tensor, device='cpu').pin_memory()
constants_map[name] = tensor_.copy_(tensor).detach().share_memory_()
return ZeroGPUWeights, (constants_map, True)
class ZeroGPUCompiledModel:
def __init__(self, archive_file: torch.types.FileLike, weights: ZeroGPUWeights):
self.archive_file = archive_file
self.weights = weights
self.compiled_model: ContextVar[AOTICompiledModel | None] = ContextVar('compiled_model', default=None)
def __call__(self, *args, **kwargs):
if (compiled_model := self.compiled_model.get()) is None:
compiled_model = cast(AOTICompiledModel, torch._inductor.aoti_load_package(self.archive_file))
compiled_model.load_constants(self.weights.constants_map, check_full_update=True, user_managed=True)
self.compiled_model.set(compiled_model)
return compiled_model(*args, **kwargs)
def __reduce__(self):
print('ZeroGPUCompiledModel.__reduce__', f'{id(self.weights)=}')
return ZeroGPUCompiledModel, (self.archive_file, self.weights)
def aoti_compile(
exported_program: torch.export.ExportedProgram,
inductor_configs: dict[str, Any] | None = None,
):
inductor_configs = (inductor_configs or {}) | INDUCTOR_CONFIGS_OVERRIDES
gm = cast(torch.fx.GraphModule, exported_program.module())
assert exported_program.example_inputs is not None
args, kwargs = exported_program.example_inputs
artifacts = torch._inductor.aot_compile(gm, args, kwargs, options=inductor_configs)
archive_file = BytesIO()
files: list[str | Weights] = [file for file in artifacts if isinstance(file, str)]
package_aoti(archive_file, files)
weights, = (artifact for artifact in artifacts if isinstance(artifact, Weights))
zerogpu_weights = ZeroGPUWeights({name: weights.get_weight(name)[0] for name in weights})
return ZeroGPUCompiledModel(archive_file, zerogpu_weights)
@contextlib.contextmanager
def capture_component_call(
pipeline: Any,
component_name: str,
component_method='forward',
):
class CapturedCallException(Exception):
def __init__(self, *args, **kwargs):
super().__init__()
self.args = args
self.kwargs = kwargs
class CapturedCall:
def __init__(self):
self.args: tuple[Any, ...] = ()
self.kwargs: dict[str, Any] = {}
component = getattr(pipeline, component_name)
captured_call = CapturedCall()
def capture_call(*args, **kwargs):
raise CapturedCallException(*args, **kwargs)
with patch.object(component, component_method, new=capture_call):
try:
yield captured_call
except CapturedCallException as e:
captured_call.args = e.args
captured_call.kwargs = e.kwargs
|