# Copyright (c) Microsoft Corporation.
# Licensed under the MIT license.
import contextlib
import os
import psutil
import ray
import torch
import torch.distributed as dist
from torch import Tensor, nn
from torch.backends import cudnn
from torch.cuda.amp import GradScaler
from torch.nn import SyncBatchNorm
from torch.nn.parallel import DistributedDataParallel
from torch.optim.optimizer import Optimizer
from archai.common import ml_utils, utils
from archai.common.config import Config
from archai.common.ordered_dict_logger import get_global_logger
from archai.supergraph.utils.multi_optim import MultiOptim
logger = get_global_logger()
[docs]class ApexUtils:
def __init__(self, apex_config:Config)->None:
# region conf vars
self._enabled = apex_config['enabled'] # global switch to disable anything apex
self._distributed_enabled = apex_config['distributed_enabled'] # enable/disable distributed mode
self._mixed_prec_enabled = apex_config['mixed_prec_enabled'] # enable/disable distributed mode
# torch.amp has default 'O1' optimization level and cannot be configured further
# torch.amp keeps BN in fp32
# There is no loss_scale option in torch.amp
self._sync_bn = apex_config['sync_bn'] # should be replace BNs with sync BNs for distributed model
self._scale_lr = apex_config['scale_lr'] # enable/disable distributed mode
self._min_world_size = apex_config['min_world_size'] # allows to confirm we are indeed in distributed setting
seed = apex_config['seed']
detect_anomaly = apex_config['detect_anomaly']
conf_gpu_ids = apex_config['gpus']
conf_ray = apex_config['ray']
self.ray_enabled = conf_ray['enabled']
self.ray_local_mode = conf_ray['local_mode']
# endregion
self._scaler = None
self._set_ranks(conf_gpu_ids)
#_log_info({'apex_config': apex_config.to_dict()})
self._log_info({'ray.enabled': self.is_ray(), 'apex.enabled': self._enabled})
self._log_info({'torch.distributed.is_available': dist.is_available(),
'apex.distributed_enabled': self._distributed_enabled,
'apex.mixed_prec_enabled': self._mixed_prec_enabled})
if dist.is_available():
# dist.* properties are otherwise not accessible
self._op_map = {'mean': dist.ReduceOp.SUM, 'sum': dist.ReduceOp.SUM,
'min': dist.ReduceOp.MIN, 'max': dist.ReduceOp.MAX}
self._log_info({'gloo_available': dist.is_gloo_available(),
'mpi_available': dist.is_mpi_available(),
'nccl_available': dist.is_nccl_available()})
if self.is_mixed():
# init enable mixed precision
assert cudnn.enabled, "Amp requires cudnn backend to be enabled."
self._scaler = GradScaler()
# enable distributed processing
if self.is_dist():
assert not self.is_ray(), "Ray is not yet enabled for Apex distributed mode"
assert dist.is_available() # distributed module is available
assert dist.is_nccl_available()
if not dist.is_initialized():
dist.init_process_group(backend='nccl', init_method='env://')
assert dist.is_initialized()
assert dist.get_world_size() == self.world_size
assert dist.get_rank() == self.global_rank
if self.is_ray():
assert not self.is_dist(), "Ray is not yet enabled for Apex distributed mode"
if not ray.is_initialized():
ray.init(local_mode=self.ray_local_mode, include_dashboard=False,
# for some reason Ray is detecting wrong number of GPUs
num_gpus=torch.cuda.device_count())
ray_cpus = ray.nodes()[0]['Resources']['CPU']
ray_gpus = ray.nodes()[0]['Resources']['GPU']
self._log_info({'ray_cpus': ray_cpus, 'ray_gpus':ray_gpus})
assert self.world_size >= 1
assert not self._min_world_size or self.world_size >= self._min_world_size
assert self.local_rank >= 0 and self.local_rank < self.world_size
assert self.global_rank >= 0 and self.global_rank < self.world_size
assert self._gpu < torch.cuda.device_count()
torch.cuda.set_device(self._gpu)
self.device = torch.device('cuda', self._gpu)
self._setup_gpus(seed, detect_anomaly)
self._log_info({'dist_initialized': dist.is_initialized() if dist.is_available() else False,
'world_size': self.world_size,
'gpu': self._gpu, 'gpu_ids':self.gpu_ids,
'local_rank': self.local_rank,
'global_rank': self.global_rank})
def _setup_gpus(self, seed:float, detect_anomaly:bool):
utils.setup_cuda(seed, local_rank=self.local_rank)
torch.autograd.set_detect_anomaly(detect_anomaly)
self._log_info({'set_detect_anomaly': detect_anomaly,
'is_anomaly_enabled': torch.is_anomaly_enabled()})
self._log_info({'gpu_names': utils.cuda_device_names(),
'gpu_count': torch.cuda.device_count(),
'CUDA_VISIBLE_DEVICES': os.environ['CUDA_VISIBLE_DEVICES']
if 'CUDA_VISIBLE_DEVICES' in os.environ else 'NotSet',
'cudnn.enabled': cudnn.enabled,
'cudnn.benchmark': cudnn.benchmark,
'cudnn.deterministic': cudnn.deterministic,
'cudnn.version': cudnn.version()
})
self._log_info({'memory': str(psutil.virtual_memory())})
self._log_info({'CPUs': str(psutil.cpu_count())})
# gpu_usage = os.popen(
# 'nvidia-smi --query-gpu=memory.total,memory.used --format=csv,nounits,noheader'
# ).read().split('\n')
# for i, line in enumerate(gpu_usage):
# vals = line.split(',')
# if len(vals) == 2:
# _log_info('GPU {} mem: {}, used: {}'.format(i, vals[0], vals[1]))
def _set_ranks(self, conf_gpu_ids:str)->None:
# this function needs to work even when torch.distributed is not available
if 'WORLD_SIZE' in os.environ:
self.world_size = int(os.environ['WORLD_SIZE'])
else:
self.world_size = 1
if 'LOCAL_RANK' in os.environ:
self.local_rank = int(os.environ['LOCAL_RANK'])
else:
self.local_rank = 0
if 'RANK' in os.environ:
self.global_rank = int(os.environ['RANK'])
else:
self.global_rank = 0
assert self.local_rank < torch.cuda.device_count(), \
f'local_rank={self.local_rank} but device_count={torch.cuda.device_count()}' \
' Possible cause may be Pytorch is not GPU enabled or you have too few GPUs'
self.gpu_ids = [int(i) for i in conf_gpu_ids.split(',') if i]
# which GPU to use, we will use only 1 GPU per process to avoid complications with apex
# remap if GPU IDs are specified
if len(self.gpu_ids):
assert len(self.gpu_ids) > self.local_rank
self._gpu = self.gpu_ids[self.local_rank]
else:
self._gpu = self.local_rank % torch.cuda.device_count()
[docs] def is_mixed(self)->bool:
return self._enabled and self._mixed_prec_enabled
[docs] def is_dist(self)->bool:
return self._enabled and self._distributed_enabled and self.world_size > 1
[docs] def is_master(self)->bool:
return self.global_rank == 0
[docs] def is_ray(self)->bool:
return self.ray_enabled
def _log_info(self, d:dict)->None:
if logger is not None:
logger.info(d, override_key=True)
[docs] def sync_devices(self)->None:
if self.is_dist():
torch.cuda.synchronize(self.device)
[docs] def barrier(self)->None:
if self.is_dist():
dist.barrier() # wait for all processes to come to this point
[docs] def reduce(self, val, op='mean'):
if self.is_dist():
if not isinstance(val, Tensor):
rt = torch.tensor(val).to(self.device)
converted = True
else:
rt = val.clone().to(self.device)
converted = False
r_op = self._op_map[op]
dist.all_reduce(rt, op=r_op)
if op=='mean':
rt /= self.world_size
if converted and len(rt.shape)==0:
return rt.item()
return rt
else:
return val
def _get_one_optim(self, multi_optim:MultiOptim)->Optimizer:
assert len(multi_optim)==1, \
'Mixed precision is only supported for one optimizer' \
f' but {len(multi_optim)} optimizers were supplied'
return multi_optim[0].optim
[docs] def backward(self, loss:torch.Tensor)->None:
if self.is_mixed():
self._scaler.scale(loss).backward() # pyright: ignore[reportGeneralTypeIssues, reportOptionalMemberAccess]
else:
loss.backward()
[docs] def autocast(self):
if self.is_mixed():
return torch.cuda.amp.autocast()
else:
return contextlib.nullcontext()
[docs] def step(self, multi_optim:MultiOptim)->None:
if self.is_mixed():
# self._scaler.unscale_ will be called automatically if it isn't called yet from grad clipping
# https://pytorch.org/docs/stable/amp.html#torch.cuda.amp.GradScaler.step
for optim_shed in multi_optim:
self._scaler.step(optim_shed.optim) # pyright: ignore[reportOptionalMemberAccess]
self._scaler.update() # pyright: ignore[reportOptionalMemberAccess]
else:
multi_optim.step()
[docs] def to_amp(self, model:nn.Module, multi_optim:MultiOptim, batch_size:int)\
->nn.Module:
# conver BNs to sync BNs in distributed mode
if self.is_dist() and self._sync_bn:
model = SyncBatchNorm.convert_sync_batchnorm(model)
self._log_info({'BNs_converted': True})
model = model.to(self.device)
# scale LR
if self.is_dist() and self._scale_lr:
for optim_shed in multi_optim:
optim = optim_shed.optim
lr = ml_utils.get_optim_lr(optim)
scaled_lr = lr * self.world_size / float(batch_size)
ml_utils.set_optim_lr(optim, scaled_lr)
self._log_info({'lr_scaled': True, 'old_lr': lr, 'new_lr': scaled_lr})
if self.is_dist():
model = DistributedDataParallel(model, device_ids=[self._gpu], output_device=self._gpu)
return model
[docs] def clip_grad(self, clip:float, model:nn.Module, multi_optim:MultiOptim)->None:
if clip > 0.0:
if self.is_mixed():
# https://pytorch.org/docs/stable/notes/amp_examples.html#working-with-multiple-models-losses-and-optimizers
self._scaler.unscale_(multi_optim[0].optim) # pyright: ignore[reportOptionalMemberAccess]
nn.utils.clip_grad_norm_(model.parameters(), clip)
else:
nn.utils.clip_grad_norm_(model.parameters(), clip)
[docs] def state_dict(self):
if self.is_mixed():
return self._scaler.state_dict() # pyright: ignore[reportOptionalMemberAccess]
else:
return None
[docs] def load_state_dict(self, state_dict):
if self.is_mixed():
self._scaler.load_state_dict(state_dict) # pyright: ignore[reportOptionalMemberAccess]
