2022-08-01 10:31:48 +00:00
|
|
|
# Copyright (c) Facebook, Inc. and its affiliates.
|
|
|
|
#
|
|
|
|
# This source code is licensed under the MIT license found in the
|
2021-10-06 02:16:20 +00:00
|
|
|
# LICENSE file in the root directory of this source tree.
|
|
|
|
import torch
|
|
|
|
from torch.optim import Optimizer
|
2022-08-01 10:31:48 +00:00
|
|
|
|
2021-10-06 02:16:20 +00:00
|
|
|
from bitsandbytes.optim.optimizer import Optimizer1State
|
|
|
|
|
2022-08-01 10:31:48 +00:00
|
|
|
|
2021-10-06 02:16:20 +00:00
|
|
|
class LARS(Optimizer1State):
|
2022-08-01 10:31:48 +00:00
|
|
|
def __init__(
|
|
|
|
self,
|
|
|
|
params,
|
|
|
|
lr,
|
|
|
|
momentum=0,
|
|
|
|
dampening=0,
|
|
|
|
weight_decay=0,
|
|
|
|
nesterov=False,
|
|
|
|
optim_bits=32,
|
|
|
|
args=None,
|
|
|
|
min_8bit_size=4096,
|
|
|
|
percentile_clipping=100,
|
|
|
|
max_unorm=0.02,
|
|
|
|
):
|
2021-10-06 02:16:20 +00:00
|
|
|
if momentum == 0:
|
2022-08-01 16:32:47 +00:00
|
|
|
raise NotImplementedError(
|
2022-10-27 11:36:39 +00:00
|
|
|
"LARS without momentum is not supported!"
|
2022-08-01 16:32:47 +00:00
|
|
|
)
|
2022-10-27 11:14:13 +00:00
|
|
|
super().__init__(
|
2022-08-01 10:31:48 +00:00
|
|
|
"lars",
|
|
|
|
params,
|
|
|
|
lr,
|
|
|
|
(momentum, dampening),
|
|
|
|
0.0,
|
|
|
|
weight_decay,
|
|
|
|
optim_bits,
|
|
|
|
args,
|
|
|
|
min_8bit_size,
|
|
|
|
percentile_clipping,
|
|
|
|
max_unorm=max_unorm,
|
|
|
|
block_wise=False,
|
|
|
|
)
|
|
|
|
|
2021-10-06 02:16:20 +00:00
|
|
|
|
|
|
|
class LARS8bit(Optimizer1State):
|
2022-08-01 10:31:48 +00:00
|
|
|
def __init__(
|
|
|
|
self,
|
|
|
|
params,
|
|
|
|
lr,
|
|
|
|
momentum=0,
|
|
|
|
dampening=0,
|
|
|
|
weight_decay=0,
|
|
|
|
nesterov=False,
|
|
|
|
args=None,
|
|
|
|
min_8bit_size=4096,
|
|
|
|
percentile_clipping=100,
|
|
|
|
max_unorm=0.02,
|
|
|
|
):
|
2021-10-06 02:16:20 +00:00
|
|
|
if momentum == 0:
|
2022-08-01 16:32:47 +00:00
|
|
|
raise NotImplementedError(
|
2022-10-27 11:36:39 +00:00
|
|
|
"LARS without momentum is not supported!"
|
2022-08-01 16:32:47 +00:00
|
|
|
)
|
2022-10-27 11:14:13 +00:00
|
|
|
super().__init__(
|
2022-08-01 10:31:48 +00:00
|
|
|
"lars",
|
|
|
|
params,
|
|
|
|
lr,
|
|
|
|
(momentum, dampening),
|
|
|
|
0.0,
|
|
|
|
weight_decay,
|
|
|
|
8,
|
|
|
|
args,
|
|
|
|
min_8bit_size,
|
|
|
|
percentile_clipping,
|
|
|
|
max_unorm=max_unorm,
|
|
|
|
block_wise=False,
|
|
|
|
)
|
|
|
|
|
2021-10-06 02:16:20 +00:00
|
|
|
|
|
|
|
class LARS32bit(Optimizer1State):
|
2022-08-01 10:31:48 +00:00
|
|
|
def __init__(
|
|
|
|
self,
|
|
|
|
params,
|
|
|
|
lr,
|
|
|
|
momentum=0,
|
|
|
|
dampening=0,
|
|
|
|
weight_decay=0,
|
|
|
|
nesterov=False,
|
|
|
|
args=None,
|
|
|
|
min_8bit_size=4096,
|
|
|
|
percentile_clipping=100,
|
|
|
|
max_unorm=0.02,
|
|
|
|
):
|
2021-10-06 02:16:20 +00:00
|
|
|
if momentum == 0:
|
2022-08-01 16:32:47 +00:00
|
|
|
raise NotImplementedError(
|
2022-10-27 11:36:39 +00:00
|
|
|
"LARS without momentum is not supported!"
|
2022-08-01 16:32:47 +00:00
|
|
|
)
|
2022-10-27 11:14:13 +00:00
|
|
|
super().__init__(
|
2022-08-01 10:31:48 +00:00
|
|
|
"lars",
|
|
|
|
params,
|
|
|
|
lr,
|
|
|
|
(momentum, dampening),
|
|
|
|
0.0,
|
|
|
|
weight_decay,
|
|
|
|
32,
|
|
|
|
args,
|
|
|
|
min_8bit_size,
|
|
|
|
percentile_clipping,
|
|
|
|
max_unorm=max_unorm,
|
|
|
|
block_wise=False,
|
|
|
|
)
|
2021-10-06 02:16:20 +00:00
|
|
|
|
|
|
|
|
|
|
|
class PytorchLARS(Optimizer):
|
2022-08-01 10:31:48 +00:00
|
|
|
def __init__(
|
|
|
|
self,
|
|
|
|
params,
|
|
|
|
lr=0.01,
|
|
|
|
momentum=0,
|
|
|
|
dampening=0,
|
|
|
|
weight_decay=0,
|
|
|
|
nesterov=False,
|
|
|
|
max_unorm=0.02,
|
|
|
|
):
|
2021-10-06 02:16:20 +00:00
|
|
|
if lr < 0.0:
|
2022-10-27 11:14:13 +00:00
|
|
|
raise ValueError(f"Invalid learning rate: {lr}")
|
2021-10-06 02:16:20 +00:00
|
|
|
if momentum < 0.0:
|
2022-10-27 11:14:13 +00:00
|
|
|
raise ValueError(f"Invalid momentum value: {momentum}")
|
2021-10-06 02:16:20 +00:00
|
|
|
if weight_decay < 0.0:
|
2022-08-01 16:32:47 +00:00
|
|
|
raise ValueError(
|
2022-10-27 11:14:13 +00:00
|
|
|
f"Invalid weight_decay value: {weight_decay}"
|
2022-08-01 16:32:47 +00:00
|
|
|
)
|
2021-10-06 02:16:20 +00:00
|
|
|
|
2022-08-01 10:31:48 +00:00
|
|
|
defaults = dict(
|
|
|
|
lr=lr,
|
|
|
|
momentum=momentum,
|
|
|
|
dampening=dampening,
|
|
|
|
weight_decay=weight_decay,
|
|
|
|
nesterov=nesterov,
|
|
|
|
max_unorm=max_unorm,
|
|
|
|
)
|
2021-10-06 02:16:20 +00:00
|
|
|
if nesterov and (momentum <= 0 or dampening != 0):
|
2022-08-01 16:32:47 +00:00
|
|
|
raise ValueError(
|
|
|
|
"Nesterov momentum requires a momentum and zero dampening"
|
|
|
|
)
|
2022-10-27 11:14:13 +00:00
|
|
|
super().__init__(params, defaults)
|
2021-10-06 02:16:20 +00:00
|
|
|
|
|
|
|
def __setstate__(self, state):
|
2022-10-27 11:14:13 +00:00
|
|
|
super().__setstate__(state)
|
2021-10-06 02:16:20 +00:00
|
|
|
for group in self.param_groups:
|
2022-08-01 10:31:48 +00:00
|
|
|
group.setdefault("nesterov", False)
|
2021-10-06 02:16:20 +00:00
|
|
|
|
|
|
|
@torch.no_grad()
|
|
|
|
def step(self, closure=None):
|
|
|
|
"""Performs a single optimization step.
|
|
|
|
|
|
|
|
Args:
|
|
|
|
closure (callable, optional): A closure that reevaluates the model
|
|
|
|
and returns the loss.
|
|
|
|
"""
|
|
|
|
loss = None
|
|
|
|
if closure is not None:
|
|
|
|
with torch.enable_grad():
|
|
|
|
loss = closure()
|
|
|
|
|
|
|
|
for group in self.param_groups:
|
|
|
|
params_with_grad = []
|
|
|
|
d_p_list = []
|
|
|
|
momentum_buffer_list = []
|
2022-08-01 10:31:48 +00:00
|
|
|
weight_decay = group["weight_decay"]
|
|
|
|
momentum = group["momentum"]
|
|
|
|
dampening = group["dampening"]
|
|
|
|
nesterov = group["nesterov"]
|
|
|
|
max_unorm = group["max_unorm"]
|
|
|
|
lr = group["lr"]
|
2021-10-06 02:16:20 +00:00
|
|
|
|
2022-08-01 10:31:48 +00:00
|
|
|
for p in group["params"]:
|
|
|
|
if p.grad is None:
|
|
|
|
continue
|
2021-10-06 02:16:20 +00:00
|
|
|
|
|
|
|
state = self.state[p]
|
|
|
|
d_p = p.grad
|
|
|
|
if weight_decay != 0:
|
2022-10-27 11:19:09 +00:00
|
|
|
d_p = d_p.add(p, alpha=weight_decay)
|
2021-10-06 02:16:20 +00:00
|
|
|
|
|
|
|
if momentum != 0:
|
2022-08-01 10:31:48 +00:00
|
|
|
buf = state.get("momentum_buffer", None)
|
2021-10-06 02:16:20 +00:00
|
|
|
|
|
|
|
if buf is None:
|
|
|
|
buf = torch.clone(d_p).detach()
|
2022-08-01 10:31:48 +00:00
|
|
|
state["momentum_buffer"] = buf
|
2021-10-06 02:16:20 +00:00
|
|
|
else:
|
|
|
|
buf.mul_(momentum).add_(d_p, alpha=1 - dampening)
|
|
|
|
|
|
|
|
if nesterov:
|
2022-08-01 10:31:48 +00:00
|
|
|
update = d_p + buf * momentum
|
2021-10-06 02:16:20 +00:00
|
|
|
else:
|
|
|
|
update = buf
|
|
|
|
|
|
|
|
update_scale = 1.0
|
|
|
|
if max_unorm > 0.0:
|
|
|
|
assert p.dtype == torch.float32
|
|
|
|
pnorm = torch.norm(p.detach())
|
|
|
|
unorm = torch.norm(update)
|
2022-08-01 10:31:48 +00:00
|
|
|
if unorm > max_unorm * pnorm:
|
|
|
|
update_scale = max_unorm * pnorm / unorm
|
2021-10-06 02:16:20 +00:00
|
|
|
|
2022-08-01 10:31:48 +00:00
|
|
|
p.add_(update, alpha=-lr * update_scale)
|
2021-10-06 02:16:20 +00:00
|
|
|
|
|
|
|
return loss
|