forked from mrq/bitsandbytes-rocm
33 lines
1.8 KiB
Python
33 lines
1.8 KiB
Python
# Copyright (c) Facebook, Inc. and its affiliates.
|
|
#
|
|
# This source code is licensed under the MIT license found in the
|
|
# LICENSE file in the root directory of this source tree.
|
|
from bitsandbytes.optim.optimizer import Optimizer1State
|
|
|
|
class SGD(Optimizer1State):
|
|
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, block_wise=True):
|
|
if momentum == 0:
|
|
raise NotImplementedError(f'SGD without momentum is not supported!')
|
|
super(SGD, self).__init__('momentum', params, lr, (momentum, dampening), 0.0,
|
|
weight_decay, optim_bits, args, min_8bit_size, percentile_clipping, block_wise)
|
|
|
|
class SGD8bit(Optimizer1State):
|
|
def __init__(self, params, lr, momentum=0, dampening=0,
|
|
weight_decay=0, nesterov=False, args=None,
|
|
min_8bit_size=4096, percentile_clipping=100, block_wise=True):
|
|
if momentum == 0:
|
|
raise NotImplementedError(f'SGD without momentum is not supported!')
|
|
super(SGD8bit, self).__init__('momentum', params, lr, (momentum, dampening), 0.0,
|
|
weight_decay, 8, args, min_8bit_size, percentile_clipping, block_wise)
|
|
|
|
class SGD32bit(Optimizer1State):
|
|
def __init__(self, params, lr, momentum=0, dampening=0,
|
|
weight_decay=0, nesterov=False, args=None,
|
|
min_8bit_size=4096, percentile_clipping=100, block_wise=True):
|
|
if momentum == 0:
|
|
raise NotImplementedError(f'SGD without momentum is not supported!')
|
|
super(SGD32bit, self).__init__('momentum', params, lr, (momentum, dampening), 0.0,
|
|
weight_decay, 32, args, min_8bit_size, percentile_clipping, block_wise)
|