245 lines
7.7 KiB
Python
245 lines
7.7 KiB
Python
# Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved.
|
||
#
|
||
# 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 copy
|
||
|
||
import paddle
|
||
from paddle import _C_ops
|
||
from paddle.base import core
|
||
from paddle.base.data_feeder import check_variable_and_dtype
|
||
from paddle.base.framework import EagerParamBase
|
||
from paddle.base.layer_helper import LayerHelper
|
||
from paddle.framework import in_dynamic_or_pir_mode
|
||
|
||
__all__ = []
|
||
|
||
|
||
# TODO(qili93): remove this op after custom op and custom device
|
||
# integrated and then move this op along with its code to plugin.
|
||
def _npu_identity(x, format=-1):
|
||
"""
|
||
|
||
This OP takes in the Tensor :attr:`x` and change it to output with
|
||
aclFormat with int value. This API is only used for Ascend NPU.
|
||
|
||
Args:
|
||
x(Tensor): An input N-D Tensor with data type bool, float16,
|
||
float32, float64, int32, int64, int16, int8, uint8.
|
||
format(int): Storage data format of the output in aclFormat,
|
||
default value is -1.
|
||
|
||
Returns:
|
||
Tensor: A Tensor with acl storage format on Ascend NPU.
|
||
|
||
Examples:
|
||
.. code-block:: pycon
|
||
|
||
>>> # doctest: +REQUIRES(env:NPU)
|
||
>>> import paddle
|
||
>>> paddle.device.set_device('npu')
|
||
|
||
>>> x = paddle.ones(shape=[6])
|
||
>>> y = paddle.incubate._npu_identity(x, 3) # ACL_FORMAT_NC1HWC0 = 3
|
||
>>> print(y.shape)
|
||
paddle.Size([1, 1, 1, 1, 16])
|
||
"""
|
||
if in_dynamic_or_pir_mode():
|
||
return _C_ops.npu_identity(x, format)
|
||
else:
|
||
check_variable_and_dtype(
|
||
x,
|
||
'x',
|
||
[
|
||
'bool',
|
||
'int8',
|
||
'uint8',
|
||
'int16',
|
||
'int32',
|
||
'int64',
|
||
'float16',
|
||
'float32',
|
||
'float64',
|
||
],
|
||
'npu_identity',
|
||
)
|
||
|
||
helper = LayerHelper('npu_identity', **locals())
|
||
out = helper.create_variable_for_type_inference(
|
||
dtype=x.dtype, stop_gradient=x.stop_gradient
|
||
)
|
||
helper.append_op(
|
||
type='npu_identity',
|
||
inputs={'x': [x]},
|
||
outputs={'out': [out]},
|
||
attrs={'format': format},
|
||
)
|
||
return out
|
||
|
||
|
||
def _load_reload_impl(src_tensor, func):
|
||
"""
|
||
Helper to create a new destination tensor and call 'func(dst, src)'
|
||
which is either offload or reload.
|
||
"""
|
||
if isinstance(src_tensor, EagerParamBase):
|
||
state = copy.deepcopy(src_tensor.__dict__)
|
||
new_param = EagerParamBase(src_tensor.shape, src_tensor.dtype, **state)
|
||
task = func(new_param, src_tensor)
|
||
return new_param, task
|
||
elif isinstance(src_tensor, paddle.Tensor):
|
||
new_varbase = core.eager.Tensor()
|
||
task = func(new_varbase, src_tensor)
|
||
return new_varbase, task
|
||
|
||
|
||
def create_async_load():
|
||
"""
|
||
Constructs a new AsyncLoad object.
|
||
It is used to load/reload data asynchronously on GPU.
|
||
"""
|
||
custom_devices = paddle.device.get_all_custom_device_type()
|
||
if paddle.is_compiled_with_xpu():
|
||
return core.XpuAsyncLoad()
|
||
elif any(
|
||
paddle.is_compiled_with_custom_device(dev) for dev in custom_devices
|
||
):
|
||
return None
|
||
else: # default is GPU or CUDA
|
||
return core.AsyncLoad()
|
||
|
||
|
||
def create_xpu_async_load():
|
||
"""
|
||
Constructs a new AsyncLoad object.
|
||
It is used to load/reload data asynchronously on XPU.
|
||
"""
|
||
return core.XpuAsyncLoad()
|
||
|
||
|
||
class _NoopAsyncTask:
|
||
"""A dummy Task for sync‐fallback on XPU."""
|
||
|
||
def is_completed(self):
|
||
return True
|
||
|
||
def cpu_wait(self):
|
||
pass
|
||
|
||
def xpu_wait(self):
|
||
pass
|
||
|
||
|
||
def async_offload(src_tensor, async_load):
|
||
"""
|
||
Loads the source tensor into the destination tensor asynchronously.
|
||
|
||
Args:
|
||
src_tensor (EagerParamBase|paddle.Tensor): The source tensor.
|
||
async_load (core.AsyncLoad): The AsyncLoad object.
|
||
|
||
Returns:
|
||
tuple: A tuple containing two elements:
|
||
- dest_tensor (EagerParamBase|paddle.Tensor): The destination tensor.
|
||
- task (Task): The task that loads the source tensor into the destination tensor.
|
||
"""
|
||
is_xpu_tensor = (
|
||
paddle.is_compiled_with_xpu()
|
||
and hasattr(src_tensor, "place")
|
||
and src_tensor.place.is_xpu_place()
|
||
)
|
||
|
||
# async_offload does not support custom device now
|
||
custom_devices = paddle.device.get_all_custom_device_type()
|
||
is_custom_tensor = (
|
||
any(
|
||
paddle.is_compiled_with_custom_device(dev) for dev in custom_devices
|
||
)
|
||
and hasattr(src_tensor, "place")
|
||
and src_tensor.place.custom_device_type() in custom_devices
|
||
)
|
||
|
||
if is_xpu_tensor or is_custom_tensor:
|
||
# sync fallback
|
||
host_tensor = src_tensor.cpu()
|
||
out = paddle.to_tensor(host_tensor.numpy(), place=paddle.CPUPlace())
|
||
return out, _NoopAsyncTask()
|
||
|
||
return _load_reload_impl(src_tensor, async_load.offload)
|
||
|
||
|
||
def async_reload(src_tensor, async_load):
|
||
"""
|
||
Reloads the source tensor into the destination tensor asynchronously.
|
||
|
||
Args:
|
||
src_tensor (EagerParamBase|paddle.Tensor): The source tensor.
|
||
async_load (core.AsyncLoad): The AsyncLoad object.
|
||
|
||
Returns:
|
||
tuple: A tuple containing two elements:
|
||
- dest_tensor (EagerParamBase|paddle.Tensor): The destination tensor.
|
||
- task (Task): The task that reloads the source tensor into the destination tensor.
|
||
"""
|
||
|
||
if (
|
||
paddle.is_compiled_with_xpu()
|
||
and hasattr(src_tensor, "place")
|
||
and src_tensor.place.is_cpu_place()
|
||
):
|
||
arr = src_tensor.numpy()
|
||
xpu = paddle.to_tensor(arr, place=paddle.XPUPlace(0))
|
||
return xpu, _NoopAsyncTask()
|
||
|
||
return _load_reload_impl(src_tensor, async_load.reload)
|
||
|
||
|
||
def async_offload_with_offset(
|
||
src_tensor, dst_tensor, src_offset, dst_offset, offload_size, async_loader
|
||
):
|
||
"""
|
||
Offloading the source tensor into the destination tensor asynchronously with offset and size customized.
|
||
|
||
Args:
|
||
src_tensor (EagerParamBase|paddle.Tensor): The source tensor.
|
||
dst_tensor (EagerParamBase|paddle.Tensor): The destination tensor.
|
||
src_offset (int): The element offset of the source tensor.
|
||
dst_offset (int): The element offset of the destination tensor.
|
||
offload_size (int): The size of the data to be loaded.
|
||
async_loader (core.AsyncLoad): The AsyncLoad object.
|
||
|
||
Returns:
|
||
task (Task): The task that operates partial offloading.
|
||
"""
|
||
assert len(src_tensor.shape) <= 1, "Only support 1-D tensor"
|
||
assert len(dst_tensor.shape) <= 1, "Only support 1-D tensor"
|
||
assert src_tensor.dtype == dst_tensor.dtype, "Only support same dtype"
|
||
|
||
return async_loader.offload_with_offset(
|
||
dst_tensor, src_tensor, dst_offset, src_offset, offload_size
|
||
)
|
||
|
||
|
||
def enable_activation_offload(model, enable=True, retry_times=1):
|
||
"""
|
||
Enable activation offload
|
||
"""
|
||
if enable:
|
||
paddle.set_flags({"FLAGS_offload_retry_times": retry_times})
|
||
paddle.core.register_offload_callback()
|
||
paddle.core.set_skip_offload_callback_tensors(model.parameters())
|
||
else:
|
||
paddle.set_flags({"FLAGS_offload_retry_times": -1})
|
||
paddle.core.clear_offload_callback()
|
||
paddle.core.set_skip_offload_callback_tensors([])
|