-
Notifications
You must be signed in to change notification settings - Fork 4.3k
/
mixed_precision_optimizer.py
214 lines (181 loc) · 7.77 KB
/
mixed_precision_optimizer.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
from typing import Dict, List, Tuple
import torch
from torch import Tensor, inf
from torch.nn import Module, Parameter
from torch.optim import Optimizer
from colossalai.interface import OptimizerWrapper
from .mixed_precision_mixin import BF16MixedPrecisionMixin, FP16MixedPrecisionMixin
class NaiveFP16MixedPrecisionMixin(FP16MixedPrecisionMixin):
def __init__(
self,
working_params: List[Parameter],
initial_scale: float = 2**16,
min_scale: float = 1,
growth_factor: float = 2,
backoff_factor: float = 0.5,
growth_interval: int = 1000,
hysteresis: int = 2,
max_scale: float = 2**32,
) -> None:
super().__init__(
initial_scale, min_scale, growth_factor, backoff_factor, growth_interval, hysteresis, max_scale
)
self.params = working_params
def check_local_overflow(self) -> bool:
for p in self.params:
if p.grad is not None and not torch.isfinite(p.grad).all():
return True
return False
class MixedPrecisionOptimizer(OptimizerWrapper):
def __init__(
self,
optim: Optimizer,
precision: str = "fp16",
initial_scale: float = 2**16,
min_scale: float = 1,
growth_factor: float = 2,
backoff_factor: float = 0.5,
growth_interval: int = 1000,
hysteresis: int = 2,
max_scale: float = 2**32,
max_norm: float = 0.0,
):
super().__init__(optim)
if precision == "fp16":
working_params = []
for group in self.optim.param_groups:
for p in group["params"]:
working_params.append(p)
self.mixed_precision = NaiveFP16MixedPrecisionMixin(
working_params,
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,
)
elif precision == "bf16":
self.mixed_precision = BF16MixedPrecisionMixin()
else:
raise ValueError(f"Unsupported precision: {precision}")
self.max_norm = max_norm
self.working_to_master_map: Dict[Parameter, Tensor] = {}
self.master_to_working_map: Dict[Tensor, Parameter] = {}
# create master weights
for group in self.optim.param_groups:
master_params = []
for p in group["params"]:
if p.requires_grad:
master_p = p
if p.dtype != torch.float:
master_p = p.detach().float()
self.working_to_master_map[p] = master_p
self.master_to_working_map[master_p] = p
master_params.append(master_p)
group["params"] = master_params
def backward(self, loss: Tensor, *args, **kwargs):
loss = self.mixed_precision.pre_backward(loss)
loss.backward(*args, **kwargs)
def backward_by_grad(self, tensor: Tensor, grad: Tensor):
grad = self.mixed_precision.pre_backward_by_grad(tensor, grad)
tensor.backward(grad)
def zero_grad(self, *args, **kwargs):
for p in self.working_to_master_map.keys():
p.grad = None
self.mixed_precision.pre_zero_grad()
return super().zero_grad(*args, **kwargs)
def _unscale_and_clip_grads(self, total_norm: float) -> None:
"""
Unscale and clip gradients before performing the optimization step.
Args:
total_norm (float): The computed total gradient norm.
Returns:
None
"""
div_scale = 1.0
# If mixed-precision training is used, get the gradient division scale from the mixed-precision handler.
if self.mixed_precision is not None:
div_scale = self.mixed_precision.get_grad_div_scale()
if self.max_norm > 0.0:
# Calculate the scaling factor for gradient clipping
# The gradient norm is scaled by 'div_scale' and then clipped to 'max_norm'
clip = ((total_norm / div_scale) + 1e-6) / self.max_norm
# If the clip factor exceeds 1, adjust 'div_scale' accordingly to ensure clipping
if clip > 1:
div_scale = clip * div_scale
# Apply the scaling factor to gradients
for group in self.param_groups:
for p in group["params"]:
if p.grad is None:
continue
p.grad.data.mul_(1.0 / div_scale)
def _compute_grad_norm(self, param_gradient_pairs: List[Tuple[Tensor]], norm_type: int = 2) -> int:
r"""
Compute and return the gradient norm for gradient clipping.
Args:
param_gradient_pairs (List[Tuple[Tensor]]): List of (parameter, gradient) pairs; gradients are used for norm calculation.
norm_type (int, optional): Type of the norm used (e.g., 2 for L2 norm). Defaults to 2.
Returns:
float: The total norm of the given gradients.
"""
if len(param_gradient_pairs) == 0:
return 0.0
# gradients used for norm calculation.
gradients = [grad for param, grad in param_gradient_pairs]
if norm_type == inf:
total_norm = max(grad.data.abs().max() for grad in gradients)
else:
total_norm_exponentiated = 0.0
for grad in gradients:
total_norm_exponentiated += grad.data.double().norm(norm_type) ** norm_type
total_norm = total_norm_exponentiated ** (1.0 / norm_type)
return total_norm
def step(self, *args, **kwargs):
if self.mixed_precision.should_skip_step():
self.zero_grad()
return
# prepare grads
for group in self.optim.param_groups:
for p in group["params"]:
working_param = self.master_to_working_map[p]
if p is working_param:
continue
if working_param.grad is not None:
p.grad = working_param.grad.data.float()
working_param.grad = None
# gradient unscale and clip.
if self.max_norm <= 0:
# no need to compute gradient norm.
total_norm = 0.0
else:
# compute the total norm.
param_gradient_pairs = [
(self.master_to_working_map[p], p.grad)
for group in self.param_groups
for p in group["params"]
if p.grad is not None
]
total_norm = self._compute_grad_norm(param_gradient_pairs)
self._unscale_and_clip_grads(total_norm)
self.optim.step(*args, **kwargs)
# update working params
for group in self.optim.param_groups:
for p in group["params"]:
working_param = self.master_to_working_map[p]
if p is working_param:
continue
working_param.data.copy_(p.data)
def update_master_params(self, model: Module):
# Update master params from working params
with torch.no_grad():
for p in model.parameters():
if (p is None) or (p not in self.working_to_master_map):
continue
master_param = self.working_to_master_map[p]
master_param.data.copy_(p.data)
def get_working_to_master_map(self) -> Dict[int, torch.Tensor]:
return {id(working_p): master_p for working_p, master_p in self.working_to_master_map.items()}
def get_master_to_working_map(self) -> Dict[int, torch.Tensor]:
return {id(master_p): working_p for master_p, working_p in self.master_to_working_map.items()}