2022-03-17 07:05:41 +00:00
|
|
|
import math
|
2022-11-17 05:42:33 +00:00
|
|
|
from typing import Optional
|
|
|
|
|
2022-03-22 06:56:59 +00:00
|
|
|
import torch
|
|
|
|
|
2023-01-06 12:50:26 +00:00
|
|
|
from colossalai.kernel.op_builder import CPUAdamBuilder
|
2022-11-17 05:42:33 +00:00
|
|
|
|
2022-07-26 09:25:24 +00:00
|
|
|
from .nvme_optimizer import NVMeOptimizer
|
2022-03-30 10:14:59 +00:00
|
|
|
|
2022-03-04 08:05:15 +00:00
|
|
|
|
2022-07-26 09:25:24 +00:00
|
|
|
class CPUAdam(NVMeOptimizer):
|
2022-03-30 10:14:59 +00:00
|
|
|
"""Implements Adam algorithm.
|
|
|
|
|
2023-06-06 06:07:36 +00:00
|
|
|
Supports parameters updating on both GPU and CPU, depending on the device of parameters.
|
2022-11-17 05:42:33 +00:00
|
|
|
But the parameters and gradients should on the same device:
|
2022-03-30 10:14:59 +00:00
|
|
|
* Parameters on CPU and gradients on CPU is allowed.
|
|
|
|
* Parameters on GPU and gradients on GPU is allowed.
|
|
|
|
* Parameters on GPU and gradients on CPU is **not** allowed.
|
|
|
|
|
2023-01-09 09:13:53 +00:00
|
|
|
`CPUAdam` requires CUDA extensions which can be built during installation or runtime.
|
2022-03-30 10:14:59 +00:00
|
|
|
|
2023-06-05 08:04:27 +00:00
|
|
|
This version of CPU Adam accelerates parameters updating on CPU with SIMD.
|
2022-03-30 10:14:59 +00:00
|
|
|
Support of AVX2 or AVX512 is required.
|
|
|
|
|
|
|
|
The GPU part is implemented in an naive way.
|
|
|
|
|
|
|
|
CPU Adam also supports the hybrid precision calculation, eg. fp32 parameters and fp16 gradients.
|
|
|
|
|
|
|
|
:class:`colossalai.nn.optimizer.CPUAdam` may be used as a drop-in replacement for ``torch.optim.AdamW``,
|
|
|
|
or ``torch.optim.Adam`` with ``adamw_mode=False``
|
|
|
|
|
|
|
|
Adam was been proposed in `Adam: A Method for Stochastic Optimization`_.
|
|
|
|
|
|
|
|
Arguments:
|
|
|
|
model_params (iterable): iterable of parameters of dicts defining
|
|
|
|
parameter groups.
|
|
|
|
lr (float, optional): learning rate. (default: 1e-3)
|
|
|
|
betas (Tuple[float, float], optional): coefficients used for computing
|
|
|
|
running averages of gradient and its square. (default: (0.9, 0.999))
|
|
|
|
eps (float, optional): term added to the denominator to improve
|
|
|
|
numerical stability. (default: 1e-8)
|
|
|
|
weight_decay (float, optional): weight decay (L2 penalty) (default: 0)
|
|
|
|
amsgrad (boolean, optional): whether to use the AMSGrad variant of this
|
|
|
|
algorithm from the paper `On the Convergence of Adam and Beyond`_
|
|
|
|
(default: False) NOT SUPPORTED yet in CPUAdam!
|
|
|
|
adamw_mode (boolean, optional): Apply L2 regularization or weight decay
|
|
|
|
True for decoupled weight decay(also known as AdamW) (default: True)
|
2022-11-17 05:42:33 +00:00
|
|
|
simd_log (boolean, optional): whether to show if you are using SIMD to
|
2022-03-30 10:14:59 +00:00
|
|
|
accelerate. (default: False)
|
2022-08-12 10:01:02 +00:00
|
|
|
nvme_offload_fraction (float, optional): Fraction of optimizer states to be offloaded to NVMe. Defaults to 0.0.
|
|
|
|
nvme_offload_dir (Optional[str], optional): Directory to save NVMe offload files.
|
2022-07-26 09:25:24 +00:00
|
|
|
If it's ``None``, a random temporary directory will be used. Defaults to None.
|
2022-04-01 08:27:03 +00:00
|
|
|
|
|
|
|
.. _Adam\: A Method for Stochastic Optimization:
|
2022-03-30 10:14:59 +00:00
|
|
|
https://arxiv.org/abs/1412.6980
|
|
|
|
.. _On the Convergence of Adam and Beyond:
|
|
|
|
https://openreview.net/forum?id=ryQu7f-RZ
|
|
|
|
"""
|
|
|
|
|
2022-03-22 06:56:59 +00:00
|
|
|
# Number of fp32 shards for per parameter
|
|
|
|
# Param weight, grad, momentum and variance
|
|
|
|
num_fp32_shards_per_param = 4
|
2022-03-04 08:05:15 +00:00
|
|
|
|
|
|
|
def __init__(self,
|
|
|
|
model_params,
|
|
|
|
lr=1e-3,
|
|
|
|
bias_correction=True,
|
2022-03-17 07:05:41 +00:00
|
|
|
betas=(0.9, 0.999),
|
2022-03-04 08:05:15 +00:00
|
|
|
eps=1e-8,
|
|
|
|
weight_decay=0,
|
|
|
|
adamw_mode=True,
|
2022-07-26 09:25:24 +00:00
|
|
|
nvme_offload_fraction: float = 0.0,
|
|
|
|
nvme_offload_dir: Optional[str] = None):
|
2022-03-17 09:24:25 +00:00
|
|
|
|
2022-03-17 07:05:41 +00:00
|
|
|
default_args = dict(lr=lr, betas=betas, eps=eps, weight_decay=weight_decay, bias_correction=bias_correction)
|
2022-07-26 09:25:24 +00:00
|
|
|
super(CPUAdam, self).__init__(model_params, default_args, nvme_offload_fraction, nvme_offload_dir)
|
2022-03-25 10:03:54 +00:00
|
|
|
self.adamw_mode = adamw_mode
|
2023-01-06 12:50:26 +00:00
|
|
|
cpu_adam = CPUAdamBuilder().load()
|
|
|
|
self.cpu_adam_op = cpu_adam.CPUAdamOptimizer(lr, betas[0], betas[1], eps, weight_decay, adamw_mode)
|
2022-03-04 08:05:15 +00:00
|
|
|
|
2022-03-17 07:05:41 +00:00
|
|
|
def torch_adam_update(self,
|
|
|
|
data,
|
|
|
|
grad,
|
|
|
|
exp_avg,
|
|
|
|
exp_avg_sq,
|
|
|
|
lr,
|
|
|
|
beta1,
|
|
|
|
beta2,
|
|
|
|
eps,
|
|
|
|
weight_decay,
|
|
|
|
bias_correction1,
|
|
|
|
bias_correction2,
|
|
|
|
use_adamw=False):
|
2023-06-05 07:58:31 +00:00
|
|
|
grad = grad.to(data.dtype)
|
2022-03-17 07:05:41 +00:00
|
|
|
|
|
|
|
if weight_decay != 0:
|
|
|
|
if use_adamw:
|
|
|
|
data.mul_(1 - lr * weight_decay)
|
|
|
|
else:
|
|
|
|
grad = grad.add(data, alpha=weight_decay)
|
|
|
|
|
|
|
|
# Decay the first and second moment running average coefficient
|
|
|
|
exp_avg.mul_(beta1).add_(grad, alpha=1 - beta1)
|
|
|
|
exp_avg_sq.mul_(beta2).addcmul_(grad, grad, value=1 - beta2)
|
|
|
|
|
|
|
|
# TODO(jiaruifang) dose not support amsgrad
|
|
|
|
denom = (exp_avg_sq.sqrt() / math.sqrt(bias_correction2)).add_(eps)
|
|
|
|
|
|
|
|
step_size = lr / bias_correction1
|
|
|
|
|
|
|
|
data.addcdiv_(exp_avg, denom, value=-step_size)
|
|
|
|
|
2022-03-04 08:05:15 +00:00
|
|
|
@torch.no_grad()
|
2022-12-12 09:58:57 +00:00
|
|
|
def step(self, closure=None, div_scale: float = -1):
|
2022-03-04 08:05:15 +00:00
|
|
|
loss = None
|
|
|
|
if closure is not None:
|
|
|
|
with torch.enable_grad():
|
|
|
|
loss = closure()
|
|
|
|
|
2022-07-26 09:25:24 +00:00
|
|
|
self._pre_step('exp_avg', 'exp_avg_sq')
|
2022-03-17 07:05:41 +00:00
|
|
|
for _, group in enumerate(self.param_groups):
|
|
|
|
for _, p in enumerate(group['params']):
|
2022-03-04 08:05:15 +00:00
|
|
|
|
|
|
|
if p.grad is None:
|
|
|
|
continue
|
|
|
|
|
|
|
|
state = self.state[p]
|
2022-03-17 07:05:41 +00:00
|
|
|
|
|
|
|
target_device = p.device
|
2022-03-04 08:05:15 +00:00
|
|
|
if len(state) == 0:
|
|
|
|
state['step'] = 0
|
|
|
|
|
2023-06-05 07:58:31 +00:00
|
|
|
# FIXME(ver217): CPU adam kernel only supports fp32 states now
|
|
|
|
assert p.dtype is torch.float, "CPUAdam only support fp32 parameters"
|
2022-03-04 08:05:15 +00:00
|
|
|
# gradient momentums
|
2023-06-05 07:58:31 +00:00
|
|
|
state['exp_avg'] = torch.zeros_like(p, device=target_device)
|
2022-03-04 08:05:15 +00:00
|
|
|
# gradient variances
|
2023-06-05 07:58:31 +00:00
|
|
|
state['exp_avg_sq'] = torch.zeros_like(p, device=target_device)
|
2022-07-26 09:25:24 +00:00
|
|
|
self._post_state_init(p)
|
2022-03-04 08:05:15 +00:00
|
|
|
|
|
|
|
state['step'] += 1
|
|
|
|
beta1, beta2 = group['betas']
|
|
|
|
|
2022-03-17 07:05:41 +00:00
|
|
|
if target_device.type == 'cpu':
|
2022-04-11 05:38:51 +00:00
|
|
|
assert p.data.numel() == p.grad.data.numel(), "parameter and gradient should have the same size"
|
2022-03-17 07:05:41 +00:00
|
|
|
assert state['exp_avg'].device.type == 'cpu', "exp_avg should stay on cpu"
|
|
|
|
assert state['exp_avg_sq'].device.type == 'cpu', "exp_avg should stay on cpu"
|
2022-07-26 09:25:24 +00:00
|
|
|
self._pre_update(p, 'exp_avg', 'exp_avg_sq')
|
2023-06-05 07:58:31 +00:00
|
|
|
if p.grad.dtype is torch.bfloat16:
|
|
|
|
# cpu adam kernel does not support bf16 now
|
|
|
|
bias_correction1 = 1 - beta1**state['step']
|
|
|
|
bias_correction2 = 1 - beta2**state['step']
|
|
|
|
self.torch_adam_update(p.data, p.grad.data, state['exp_avg'], state['exp_avg_sq'], group['lr'],
|
|
|
|
beta1, beta2, group['eps'], group['weight_decay'], bias_correction1,
|
|
|
|
bias_correction2, self.adamw_mode)
|
|
|
|
else:
|
|
|
|
self.cpu_adam_op.step(state['step'], group['lr'], beta1, beta2, group['eps'],
|
|
|
|
group['weight_decay'], group['bias_correction'], p.data, p.grad.data,
|
|
|
|
state['exp_avg'], state['exp_avg_sq'], div_scale)
|
2022-07-26 09:25:24 +00:00
|
|
|
self._post_update(p, 'exp_avg', 'exp_avg_sq')
|
2022-03-17 07:05:41 +00:00
|
|
|
elif target_device.type == 'cuda':
|
2022-12-12 09:58:57 +00:00
|
|
|
assert div_scale == -1, "div_scale should remain default"
|
2022-03-17 07:05:41 +00:00
|
|
|
assert state['exp_avg'].device.type == 'cuda', "exp_avg should stay on cuda"
|
|
|
|
assert state['exp_avg_sq'].device.type == 'cuda', "exp_avg should stay on cuda"
|
|
|
|
|
2022-04-11 05:38:51 +00:00
|
|
|
bias_correction1 = 1 - beta1**state['step']
|
|
|
|
bias_correction2 = 1 - beta2**state['step']
|
2022-03-17 07:05:41 +00:00
|
|
|
|
|
|
|
# adam on cuda
|
|
|
|
self.torch_adam_update(p.data, p.grad.data, state['exp_avg'], state['exp_avg_sq'], group['lr'],
|
|
|
|
beta1, beta2, group['eps'], group['weight_decay'], bias_correction1,
|
2022-03-25 10:03:54 +00:00
|
|
|
bias_correction2, self.adamw_mode)
|
2022-03-17 07:05:41 +00:00
|
|
|
else:
|
|
|
|
raise RuntimeError
|
2022-07-26 09:25:24 +00:00
|
|
|
self._post_step()
|
2022-03-04 08:05:15 +00:00
|
|
|
return loss
|