2023-03-04 12:08:11 +00:00
|
|
|
# this code is inspired by the DeepSpeed library and implemented with our own design from scratch
|
2022-12-09 10:09:17 +00:00
|
|
|
import math
|
2023-01-11 07:30:09 +00:00
|
|
|
import warnings
|
2022-11-02 08:11:34 +00:00
|
|
|
from enum import Enum
|
2022-11-16 06:44:28 +00:00
|
|
|
from typing import Any, Dict, Set, Tuple
|
2022-11-02 08:11:34 +00:00
|
|
|
|
2022-06-02 04:13:15 +00:00
|
|
|
import torch
|
|
|
|
import torch.distributed as dist
|
2022-10-09 01:18:51 +00:00
|
|
|
from torch.nn import Parameter
|
2022-11-02 08:11:34 +00:00
|
|
|
from torch.optim import Optimizer
|
|
|
|
|
2022-06-02 04:13:15 +00:00
|
|
|
from colossalai.amp.naive_amp.grad_scaler import DynamicGradScaler
|
|
|
|
from colossalai.logging import get_dist_logger
|
2022-12-12 09:58:57 +00:00
|
|
|
from colossalai.nn.optimizer import ColossalaiOptimizer, CPUAdam, FusedAdam, HybridAdam
|
2023-01-11 04:22:45 +00:00
|
|
|
from colossalai.utils import disposable, get_current_device, is_ddp_ignored
|
2022-06-02 04:13:15 +00:00
|
|
|
|
2023-04-04 05:48:16 +00:00
|
|
|
from .chunk import Chunk, ChunkManager
|
|
|
|
from .gemini_ddp import ZeroDDP
|
|
|
|
|
|
|
|
__all__ = ['ZeroOptimizer', 'GeminiAdamOptimizer']
|
|
|
|
|
2022-12-12 09:58:57 +00:00
|
|
|
_AVAIL_OPTIM_LIST = {FusedAdam, CPUAdam, HybridAdam}
|
|
|
|
|
2022-06-02 04:13:15 +00:00
|
|
|
|
|
|
|
class OptimState(Enum):
|
|
|
|
SCALED = 0
|
|
|
|
UNSCALED = 1
|
|
|
|
|
|
|
|
|
|
|
|
class ZeroOptimizer(ColossalaiOptimizer):
|
2022-07-21 07:54:53 +00:00
|
|
|
"""A wrapper for optimizer. ``ZeroDDP`` and ``ZeroOptimizer`` implement Zero Redundancy Optimizer (ZeRO state-3).
|
|
|
|
|
|
|
|
Note:
|
|
|
|
You must use ``ZeroDDP`` with ``ZeroOptimizer``.
|
|
|
|
|
|
|
|
Note:
|
|
|
|
Make sure you set ``placement_policy`` of ``GeminiManager`` to `"auto"`,
|
|
|
|
if you set ``gpu_margin_mem_ratio > 0``.
|
|
|
|
|
|
|
|
Args:
|
|
|
|
optim (Optimizer): An Optimizer instance.
|
|
|
|
module (ZeroDDP): A ``ZeroDDP`` instance.
|
2022-10-09 01:18:51 +00:00
|
|
|
gpu_margin_mem_ratio (float, optional): The ratio of GPU remaining memory (after the first forward-backward)
|
|
|
|
which will be used when using hybrid CPU optimizer.
|
2022-07-21 07:54:53 +00:00
|
|
|
This argument is meaningless when `placement_policy` of `GeminiManager` is not "auto".
|
|
|
|
Defaults to 0.0.
|
|
|
|
initial_scale (float, optional): Initial scale used by DynamicGradScaler. Defaults to 2**32.
|
|
|
|
min_scale (float, optional): Min scale used by DynamicGradScaler. Defaults to 1.
|
2023-04-10 03:11:28 +00:00
|
|
|
growth_factor (float, optional): Growth_factor used by DynamicGradScaler. Defaults to 2.
|
|
|
|
backoff_factor (float, optional): Backoff_factor used by DynamicGradScaler. Defaults to 0.5.
|
|
|
|
growth_interval (float, optional): Growth_interval used by DynamicGradScaler. Defaults to 1000.
|
|
|
|
hysteresis (float, optional): Hysteresis used by DynamicGradScaler. Defaults to 2.
|
|
|
|
max_scale (int, optional): Max_scale used by DynamicGradScaler. Defaults to 2**32.
|
|
|
|
clipping_norm (float, optional): The norm value used to clip gradient. Defaults to 0.0.
|
|
|
|
norm_type (float, optional): The type of norm used for gradient clipping. Currently, only L2-norm (norm_type=2.0)
|
|
|
|
is supported in ZeroOptimizer. Defaults to 2.0.
|
2023-04-17 03:25:35 +00:00
|
|
|
verbose (bool, optional): Whether to print verbose information, including grad overflow info. Defaults to False.
|
2023-04-10 03:11:28 +00:00
|
|
|
"""
|
2022-06-02 04:13:15 +00:00
|
|
|
|
|
|
|
def __init__(self,
|
|
|
|
optim: Optimizer,
|
2022-06-21 08:35:23 +00:00
|
|
|
module: ZeroDDP,
|
2022-06-10 06:48:28 +00:00
|
|
|
gpu_margin_mem_ratio: float = 0.0,
|
2022-06-02 04:13:15 +00:00
|
|
|
initial_scale: float = 2**32,
|
|
|
|
min_scale: float = 1,
|
|
|
|
growth_factor: float = 2,
|
|
|
|
backoff_factor: float = 0.5,
|
|
|
|
growth_interval: int = 1000,
|
|
|
|
hysteresis: int = 2,
|
2022-11-16 06:44:28 +00:00
|
|
|
max_scale: float = 2**32,
|
2022-12-09 10:09:17 +00:00
|
|
|
clipping_norm: float = 0.0,
|
|
|
|
norm_type: float = 2.0,
|
2023-04-17 03:25:35 +00:00
|
|
|
verbose: bool = False,
|
2022-11-16 06:44:28 +00:00
|
|
|
**defaults: Any):
|
2022-06-02 04:13:15 +00:00
|
|
|
super().__init__(optim)
|
2022-06-21 08:35:23 +00:00
|
|
|
assert isinstance(module, ZeroDDP)
|
2023-01-29 09:52:58 +00:00
|
|
|
assert type(optim) in _AVAIL_OPTIM_LIST, "You should use an optimizer in the available list:\n" \
|
|
|
|
f"{_AVAIL_OPTIM_LIST}"
|
2022-06-02 04:13:15 +00:00
|
|
|
self.module = module
|
2022-06-10 06:48:28 +00:00
|
|
|
self.gemini_manager = module.gemini_manager
|
2022-10-09 01:18:51 +00:00
|
|
|
self.chunk_manager: ChunkManager = self.gemini_manager.chunk_manager
|
2022-06-02 04:13:15 +00:00
|
|
|
self.optim_state = OptimState.UNSCALED
|
2022-10-09 01:18:51 +00:00
|
|
|
self.param_to_range: Dict[Parameter, Tuple[int, int]] = dict()
|
|
|
|
self.param_to_chunk32: Dict[Parameter, Chunk] = dict()
|
|
|
|
self.chunk16_set: Set[Chunk] = set()
|
2022-12-09 10:09:17 +00:00
|
|
|
self.clipping_flag = clipping_norm > 0.0
|
|
|
|
self.max_norm = clipping_norm
|
2023-04-17 03:25:35 +00:00
|
|
|
self.verbose = verbose
|
2022-12-09 10:09:17 +00:00
|
|
|
|
|
|
|
if self.clipping_flag:
|
|
|
|
assert norm_type == 2.0, "ZeroOptimizer only supports L2 norm now"
|
2022-10-09 01:18:51 +00:00
|
|
|
|
2023-01-11 07:30:09 +00:00
|
|
|
ddp_param_list = []
|
|
|
|
for name, param in module.named_parameters():
|
|
|
|
if is_ddp_ignored(param):
|
|
|
|
if param.requires_grad:
|
|
|
|
warnings.warn(f"Parameter `{name}` is ignored by DDP but requires gradient! "
|
|
|
|
"You should handle its optimizer update by yourself!")
|
|
|
|
else:
|
|
|
|
ddp_param_list.append(param)
|
|
|
|
|
|
|
|
for p, fp32_p in zip(ddp_param_list, module.fp32_params):
|
2022-10-09 01:18:51 +00:00
|
|
|
chunk_16 = self.chunk_manager.get_chunk(p)
|
|
|
|
if chunk_16 not in self.chunk16_set:
|
2022-12-09 10:09:17 +00:00
|
|
|
chunk_16.l2_norm_flag = self.clipping_flag
|
2022-10-09 01:18:51 +00:00
|
|
|
self.chunk16_set.add(chunk_16)
|
|
|
|
|
|
|
|
self.__init__optimizer()
|
2022-06-02 04:13:15 +00:00
|
|
|
|
|
|
|
# Grad scaler
|
|
|
|
self.grad_scaler = DynamicGradScaler(initial_scale=initial_scale,
|
|
|
|
min_scale=min_scale,
|
|
|
|
growth_factor=growth_factor,
|
|
|
|
backoff_factor=backoff_factor,
|
|
|
|
growth_interval=growth_interval,
|
|
|
|
hysteresis=hysteresis,
|
|
|
|
max_scale=max_scale)
|
2022-10-09 01:18:51 +00:00
|
|
|
self._found_overflow: torch.Tensor = torch.zeros(1, dtype=torch.int64, device=get_current_device())
|
2022-06-02 04:13:15 +00:00
|
|
|
self._logger = get_dist_logger()
|
|
|
|
|
2022-06-10 06:48:28 +00:00
|
|
|
self.gpu_margin_mem_ratio: float = float(gpu_margin_mem_ratio)
|
|
|
|
assert 0.0 <= self.gpu_margin_mem_ratio <= 1.0, f'gpu_margin_mem_ratio must >=0.0 and <=1.0'
|
|
|
|
# Only move fp32 shards from CPU to GPU when user allows and inner optimizer is valid
|
|
|
|
# Inner optimizer must support optimizing hybrid (CPU and CUDA) tensors,
|
|
|
|
# and it must set `num_fp32_shards_per_param` correctly
|
|
|
|
self._should_move_fp32_params_h2d: bool = self.gemini_manager.is_cuda_margin_mem_avail and self.gpu_margin_mem_ratio > 0.0 and getattr(
|
|
|
|
optim, 'num_fp32_shards_per_param', 0) >= 2
|
|
|
|
if self.gpu_margin_mem_ratio > 0.0 and not self.gemini_manager.is_cuda_margin_mem_avail:
|
|
|
|
self._logger.warning(f'gpu_margin_mem_ratio is meaningless when placement_policy is not "auto"', ranks=[0])
|
|
|
|
|
|
|
|
self._register_states = disposable(self._register_states_)
|
|
|
|
|
2022-10-09 01:18:51 +00:00
|
|
|
def _set_grad_ptr(self):
|
|
|
|
for group in self.param_groups:
|
|
|
|
for fake_param in group['params']:
|
|
|
|
chunk32 = self.param_to_chunk32[fake_param]
|
|
|
|
begin, end = self.param_to_range[fake_param]
|
|
|
|
chunk16 = chunk32.paired_chunk
|
|
|
|
|
|
|
|
fake_param.data = chunk16.payload[begin:end]
|
|
|
|
fake_param.grad = fake_param.data
|
|
|
|
fake_param.data = chunk32.payload[begin:end]
|
2022-06-02 04:13:15 +00:00
|
|
|
|
|
|
|
def _update_fp16_params(self):
|
2022-10-09 01:18:51 +00:00
|
|
|
none_tensor = torch.empty([0])
|
|
|
|
for group in self.param_groups:
|
|
|
|
for fake_param in group['params']:
|
|
|
|
assert fake_param.grad is None
|
2023-02-17 07:29:07 +00:00
|
|
|
fake_param.data = none_tensor.to(fake_param.device)
|
2022-10-09 01:18:51 +00:00
|
|
|
|
|
|
|
for chunk16 in self.chunk16_set:
|
|
|
|
chunk16.optim_update()
|
2022-06-02 04:13:15 +00:00
|
|
|
|
|
|
|
def _check_overflow(self):
|
|
|
|
# clear previous overflow record
|
|
|
|
self._found_overflow.fill_(self.module.overflow_counter)
|
|
|
|
|
|
|
|
# all-reduce across global group
|
|
|
|
dist.all_reduce(self._found_overflow)
|
|
|
|
|
|
|
|
return self._found_overflow.item() > 0
|
|
|
|
|
2023-01-10 07:43:06 +00:00
|
|
|
def _clear_global_norm(self) -> None:
|
|
|
|
for c16 in self.chunk16_set:
|
|
|
|
c16.l2_norm = None
|
|
|
|
|
2022-12-09 10:09:17 +00:00
|
|
|
def _calc_global_norm(self) -> float:
|
|
|
|
norm_sqr: float = 0.0
|
|
|
|
group_to_norm = dict()
|
|
|
|
for c16 in self.chunk16_set:
|
|
|
|
assert c16.l2_norm is not None
|
|
|
|
|
|
|
|
if c16.is_gathered:
|
|
|
|
norm_sqr += c16.l2_norm
|
|
|
|
else:
|
|
|
|
# this chunk is sharded, use communication to collect total norm
|
|
|
|
if c16.torch_pg not in group_to_norm:
|
|
|
|
group_to_norm[c16.torch_pg] = 0.0
|
|
|
|
group_to_norm[c16.torch_pg] += c16.l2_norm
|
|
|
|
|
|
|
|
c16.l2_norm = None # clear l2 norm
|
|
|
|
|
|
|
|
comm_buffer = torch.zeros(1, dtype=torch.float, device=get_current_device())
|
|
|
|
for group, part_norm in group_to_norm.items():
|
|
|
|
comm_buffer.fill_(part_norm)
|
|
|
|
dist.all_reduce(comm_buffer, group=group)
|
|
|
|
norm_sqr += comm_buffer.item()
|
|
|
|
|
|
|
|
global_norm = math.sqrt(norm_sqr)
|
|
|
|
return global_norm
|
|
|
|
|
2022-12-12 09:58:57 +00:00
|
|
|
def _get_combined_scale(self):
|
|
|
|
loss_scale = 1
|
|
|
|
|
|
|
|
if self.optim_state == OptimState.SCALED:
|
|
|
|
loss_scale = self.loss_scale
|
|
|
|
self.optim_state = OptimState.UNSCALED
|
2022-12-09 10:09:17 +00:00
|
|
|
|
2022-12-12 09:58:57 +00:00
|
|
|
combined_scale = loss_scale
|
2022-12-09 10:09:17 +00:00
|
|
|
if self.clipping_flag:
|
|
|
|
total_norm = self._calc_global_norm()
|
2022-12-12 09:58:57 +00:00
|
|
|
clip = ((total_norm / loss_scale) + 1e-6) / self.max_norm
|
2022-12-09 10:09:17 +00:00
|
|
|
if clip > 1:
|
2022-12-12 09:58:57 +00:00
|
|
|
combined_scale = clip * loss_scale
|
2022-12-09 10:09:17 +00:00
|
|
|
|
2022-12-12 09:58:57 +00:00
|
|
|
if combined_scale == 1:
|
|
|
|
return -1
|
|
|
|
else:
|
|
|
|
return combined_scale
|
2022-06-02 04:13:15 +00:00
|
|
|
|
|
|
|
@property
|
|
|
|
def loss_scale(self):
|
|
|
|
return self.grad_scaler.scale.item()
|
|
|
|
|
|
|
|
def zero_grad(self, *args, **kwargs):
|
|
|
|
self.module.overflow_counter = 0
|
|
|
|
return self.optim.zero_grad(set_to_none=True)
|
|
|
|
|
|
|
|
def step(self, *args, **kwargs):
|
2022-06-10 06:48:28 +00:00
|
|
|
self._maybe_move_fp32_params()
|
2022-10-09 01:18:51 +00:00
|
|
|
self._set_grad_ptr()
|
2022-12-09 10:09:17 +00:00
|
|
|
|
2022-06-02 04:13:15 +00:00
|
|
|
found_inf = self._check_overflow()
|
|
|
|
if found_inf:
|
2022-12-09 10:09:17 +00:00
|
|
|
self.optim_state = OptimState.UNSCALED # no need to unscale grad
|
|
|
|
self.grad_scaler.update(found_inf) # update gradient scaler
|
2023-04-17 03:25:35 +00:00
|
|
|
if self.verbose:
|
|
|
|
self._logger.info(f'Found overflow. Skip step')
|
2023-01-10 07:43:06 +00:00
|
|
|
self._clear_global_norm() # clear recorded norm
|
2022-12-09 10:09:17 +00:00
|
|
|
self.zero_grad() # reset all gradients
|
2022-06-02 04:13:15 +00:00
|
|
|
self._update_fp16_params()
|
|
|
|
return
|
2022-12-09 10:09:17 +00:00
|
|
|
|
2022-12-12 09:58:57 +00:00
|
|
|
# get combined scale. combined scale = loss scale * clipping norm
|
|
|
|
# so that gradient = gradient / combined scale
|
|
|
|
combined_scale = self._get_combined_scale()
|
2022-12-09 10:09:17 +00:00
|
|
|
self.grad_scaler.update(found_inf)
|
|
|
|
|
2022-12-12 09:58:57 +00:00
|
|
|
ret = self.optim.step(div_scale=combined_scale, *args, **kwargs)
|
2022-06-10 06:48:28 +00:00
|
|
|
self._register_states()
|
2022-08-09 08:08:12 +00:00
|
|
|
self.zero_grad()
|
2022-06-02 04:13:15 +00:00
|
|
|
self._update_fp16_params()
|
|
|
|
return ret
|
|
|
|
|
2022-08-11 14:58:58 +00:00
|
|
|
def clip_grad_norm(self, model: torch.nn.Module, max_norm: float, norm_type: float = 2.0):
|
2022-10-09 01:18:51 +00:00
|
|
|
raise NotImplementedError
|
2022-06-02 04:13:15 +00:00
|
|
|
|
|
|
|
def backward(self, loss: torch.Tensor):
|
|
|
|
loss = self.loss_scale * loss
|
|
|
|
self.optim_state = OptimState.SCALED
|
|
|
|
self.module.backward(loss)
|
|
|
|
|
|
|
|
def backward_by_grad(self, tensor: torch.Tensor, grad: torch.Tensor):
|
2022-07-21 08:43:58 +00:00
|
|
|
# This function is called except the last stage of pipeline parallel
|
|
|
|
# It receives the scaled grad from the previous rank
|
|
|
|
# No need to scale the grad again
|
|
|
|
# Need to unscale when optimizing
|
|
|
|
self.optim_state = OptimState.SCALED
|
2022-06-02 04:13:15 +00:00
|
|
|
self.module.backward_by_grad(tensor, grad)
|
2022-06-10 06:48:28 +00:00
|
|
|
|
|
|
|
def _maybe_move_fp32_params(self):
|
|
|
|
if self._should_move_fp32_params_h2d:
|
|
|
|
self._should_move_fp32_params_h2d = False
|
|
|
|
available_cuda_margin_mem = self.gemini_manager.cuda_margin_mem * self.gpu_margin_mem_ratio
|
|
|
|
fp32_params_available_cuda_margin_mem = available_cuda_margin_mem / self.optim.num_fp32_shards_per_param
|
|
|
|
fp32_params_used_cuda_margin_mem = 0
|
2022-10-09 01:18:51 +00:00
|
|
|
|
|
|
|
for group in self.param_groups:
|
|
|
|
for fake_param in group['params']:
|
|
|
|
chunk32 = self.param_to_chunk32[fake_param]
|
|
|
|
chunk16 = chunk32.paired_chunk
|
|
|
|
|
|
|
|
if chunk32.device_type == 'cuda':
|
|
|
|
continue
|
|
|
|
|
|
|
|
if fp32_params_used_cuda_margin_mem + chunk32.payload_mem < fp32_params_available_cuda_margin_mem:
|
|
|
|
self.chunk_manager.move_chunk(chunk32, get_current_device())
|
|
|
|
# stores grad now
|
|
|
|
self.chunk_manager.move_chunk(chunk16, get_current_device())
|
|
|
|
self.module.set_chunk_grad_device(chunk16, get_current_device())
|
|
|
|
fp32_params_used_cuda_margin_mem += chunk32.payload_mem
|
|
|
|
|
|
|
|
for group in self.param_groups:
|
|
|
|
for fake_param in group['params']:
|
|
|
|
chunk32 = self.param_to_chunk32[fake_param]
|
|
|
|
if chunk32.device_type == 'cuda':
|
|
|
|
state = self.optim.state[fake_param]
|
2022-06-24 09:25:57 +00:00
|
|
|
for k, v in state.items():
|
|
|
|
if isinstance(v, torch.Tensor):
|
|
|
|
state[k] = v.to(get_current_device())
|
|
|
|
|
2022-06-10 06:48:28 +00:00
|
|
|
def _register_states_(self):
|
|
|
|
for group in self.optim.param_groups:
|
|
|
|
for p in group['params']:
|
|
|
|
state = self.optim.state[p]
|
|
|
|
for val in state.values():
|
|
|
|
if isinstance(val, torch.Tensor):
|
|
|
|
self.chunk_manager.add_extern_static_tensor(val)
|
2022-06-24 09:25:57 +00:00
|
|
|
|
2022-10-09 01:18:51 +00:00
|
|
|
def __init__optimizer(self):
|
2022-09-24 11:58:18 +00:00
|
|
|
|
2022-10-09 01:18:51 +00:00
|
|
|
def get_range_pair(local_chunk: Chunk, local_param: Parameter):
|
|
|
|
param_info = local_chunk.tensors_info[local_param]
|
2022-11-02 08:11:34 +00:00
|
|
|
if local_chunk.keep_gathered:
|
|
|
|
return param_info.offset, param_info.end
|
2022-10-09 01:18:51 +00:00
|
|
|
begin = max(0, param_info.offset - local_chunk.shard_begin)
|
|
|
|
end = min(local_chunk.shard_size, param_info.end - local_chunk.shard_begin)
|
|
|
|
return begin, end
|
2022-09-24 11:58:18 +00:00
|
|
|
|
2022-10-09 01:18:51 +00:00
|
|
|
for group in self.optim.param_groups:
|
|
|
|
fake_params_list = list()
|
|
|
|
|
|
|
|
for param in group['params']:
|
2023-01-11 07:30:09 +00:00
|
|
|
if is_ddp_ignored(param):
|
|
|
|
continue
|
2022-10-09 01:18:51 +00:00
|
|
|
chunk16 = self.chunk_manager.get_chunk(param)
|
|
|
|
range_pair = get_range_pair(chunk16, param)
|
|
|
|
if range_pair[0] >= range_pair[1]:
|
|
|
|
continue
|
|
|
|
|
2023-02-17 07:29:07 +00:00
|
|
|
grad_device = self.module.grads_device[param]
|
|
|
|
fake_param = torch.nn.Parameter(torch.empty([0], device=grad_device))
|
2022-10-09 01:18:51 +00:00
|
|
|
self.param_to_chunk32[fake_param] = chunk16.paired_chunk
|
|
|
|
self.param_to_range[fake_param] = range_pair
|
|
|
|
|
|
|
|
fake_params_list.append(fake_param)
|
|
|
|
|
|
|
|
group['params'] = fake_params_list
|
2023-04-04 05:48:16 +00:00
|
|
|
|
|
|
|
|
|
|
|
class GeminiAdamOptimizer(ZeroOptimizer):
|
|
|
|
|
|
|
|
def __init__(self, model: torch.nn.Module, **defaults: Any) -> None:
|
|
|
|
optimizer = HybridAdam(model.parameters(), **defaults)
|
|
|
|
super().__init__(optimizer, model, **defaults)
|