2023-02-22 02:27:59 +00:00
|
|
|
from typing import Callable, List, Tuple
|
2022-11-16 15:12:31 +00:00
|
|
|
|
|
|
|
import torch
|
|
|
|
|
2023-04-04 09:40:45 +00:00
|
|
|
from colossalai._analyzer._subclasses.flop_tensor import ewise_flop_counter as elementwise_flop_counter
|
|
|
|
from colossalai._analyzer.fx.node_util import compute_size_in_bytes as activation_size
|
2022-11-16 15:12:31 +00:00
|
|
|
from colossalai.auto_parallel.tensor_shard.sharding_strategy import MemoryCost, OperationDataType, TrainCycleItem
|
|
|
|
|
|
|
|
from ..registry import meta_register
|
|
|
|
|
2023-02-22 02:27:59 +00:00
|
|
|
__all__ = ["elementwise_meta_info"]
|
2022-11-16 15:12:31 +00:00
|
|
|
|
|
|
|
|
2023-02-22 02:27:59 +00:00
|
|
|
def elementwise_meta_info(temp_mem_scale: float = 0, buffer_mem_scale: float = 0) -> Callable:
|
|
|
|
"""This is a function to create the meta information generator for elementwise operations
|
2022-11-16 15:12:31 +00:00
|
|
|
|
2023-02-22 02:27:59 +00:00
|
|
|
Args:
|
|
|
|
temp_mem_scale (float, optional): temp memory scaling factor for backward. Defaults to 0.
|
|
|
|
buffer_mem_scale (float, optional): buffer memory scaling factor for forward. Defaults to 0.
|
2022-11-16 15:12:31 +00:00
|
|
|
|
2023-02-13 08:09:22 +00:00
|
|
|
Returns:
|
2023-02-22 02:27:59 +00:00
|
|
|
Callable: meta information generator
|
2023-02-13 08:09:22 +00:00
|
|
|
"""
|
2023-02-22 02:27:59 +00:00
|
|
|
|
|
|
|
def meta_func(*args, **kwargs) -> Tuple[TrainCycleItem, TrainCycleItem, List[torch.Tensor]]:
|
|
|
|
input_tensor = next(
|
|
|
|
filter(
|
2023-09-19 06:20:26 +00:00
|
|
|
lambda x: (x.type == OperationDataType.ARG or x.type == OperationDataType.PARAM)
|
|
|
|
and x.name != "softmax_dim",
|
|
|
|
args,
|
|
|
|
)
|
|
|
|
).data
|
2023-02-22 02:27:59 +00:00
|
|
|
output_tensor = next(filter(lambda x: x.type == OperationDataType.OUTPUT, args)).data
|
2023-09-19 06:20:26 +00:00
|
|
|
is_inplace = 1 if kwargs.get("inplace", False) else 0
|
2023-02-22 02:27:59 +00:00
|
|
|
|
|
|
|
flop_counter = elementwise_flop_counter(1, 0)
|
|
|
|
# calculate compute cost
|
|
|
|
fwd_compute_cost = flop_counter([input_tensor], [output_tensor])
|
|
|
|
bwd_compute_cost = flop_counter([output_tensor], [input_tensor])
|
|
|
|
|
2023-09-19 06:20:26 +00:00
|
|
|
compute_cost = TrainCycleItem(
|
|
|
|
fwd=fwd_compute_cost, bwd=bwd_compute_cost, total=fwd_compute_cost + bwd_compute_cost
|
|
|
|
)
|
2023-02-22 02:27:59 +00:00
|
|
|
|
|
|
|
# calculate memory cost
|
|
|
|
# NOTE: currently in SPMD solver we always believe that there will be a new tensor created in forward
|
|
|
|
# NOTE: if in_place is True, we will not create a new tensor in forward
|
2023-09-19 06:20:26 +00:00
|
|
|
fwd_memory_cost = MemoryCost(
|
|
|
|
activation=activation_size(input_tensor) * (2 - is_inplace),
|
|
|
|
parameter=0,
|
|
|
|
temp=0,
|
|
|
|
buffer=activation_size(input_tensor) * buffer_mem_scale,
|
|
|
|
)
|
2023-02-22 02:27:59 +00:00
|
|
|
|
|
|
|
# temp_mem_scale is for situation like softmax backward
|
|
|
|
# the buffer will be removed during backward phase
|
|
|
|
bwd_memory_cost = MemoryCost(
|
|
|
|
activation=activation_size(input_tensor) - activation_size(input_tensor) * buffer_mem_scale,
|
|
|
|
parameter=0,
|
|
|
|
temp=activation_size(input_tensor) * temp_mem_scale + activation_size(input_tensor) * buffer_mem_scale,
|
2023-09-19 06:20:26 +00:00
|
|
|
buffer=0,
|
|
|
|
)
|
2023-02-22 02:27:59 +00:00
|
|
|
|
|
|
|
# total cost is the sum of forward and backward cost
|
2023-09-19 06:20:26 +00:00
|
|
|
total_cost = MemoryCost(
|
|
|
|
activation=fwd_memory_cost.activation + bwd_memory_cost.activation,
|
|
|
|
parameter=fwd_memory_cost.parameter + bwd_memory_cost.parameter,
|
|
|
|
temp=fwd_memory_cost.temp + bwd_memory_cost.temp,
|
|
|
|
buffer=fwd_memory_cost.buffer + bwd_memory_cost.buffer,
|
|
|
|
)
|
2023-02-22 02:27:59 +00:00
|
|
|
|
|
|
|
memory_cost = TrainCycleItem(fwd=fwd_memory_cost, bwd=bwd_memory_cost, total=total_cost)
|
|
|
|
|
|
|
|
# store fwd_in, fwd_buffer, fwd_out
|
|
|
|
fwd_in = []
|
2023-09-19 06:20:26 +00:00
|
|
|
fwd_buffer = [torch.zeros_like(output_tensor, device="meta")]
|
|
|
|
fwd_out = [torch.zeros_like(output_tensor, device="meta")]
|
2023-02-22 02:27:59 +00:00
|
|
|
|
|
|
|
return compute_cost, memory_cost, fwd_in, fwd_buffer, fwd_out
|
|
|
|
|
|
|
|
return meta_func
|
|
|
|
|
|
|
|
|
|
|
|
# register meta information
|
|
|
|
# (0, 0)
|
|
|
|
meta_register.register([torch.nn.ReLU, torch.nn.functional.relu, torch.tanh])(elementwise_meta_info(0, 0))
|
|
|
|
|
|
|
|
# (1, 0)
|
|
|
|
meta_register.register([torch.nn.Softmax, torch.nn.functional.softmax])(elementwise_meta_info(1, 0))
|
|
|
|
|
|
|
|
# (0, 0.25) for dropout, the buffer is in bool type so that the buffer memory cost is 0.25 times of input tensor
|
|
|
|
meta_register.register([torch.nn.Dropout, torch.nn.functional.dropout])(elementwise_meta_info(0, 0.25))
|