Skip to content
体验新版
项目
组织
正在加载...
登录
切换导航
打开侧边栏
Greenplum
DeepSpeed
提交
ade3bbc0
D
DeepSpeed
项目概览
Greenplum
/
DeepSpeed
上一次同步 1 年多
通知
10
Star
0
Fork
0
代码
文件
提交
分支
Tags
贡献者
分支图
Diff
Issue
0
列表
看板
标记
里程碑
合并请求
0
DevOps
流水线
流水线任务
计划
Wiki
0
Wiki
分析
仓库
DevOps
项目成员
Pages
D
DeepSpeed
项目概览
项目概览
详情
发布
仓库
仓库
文件
提交
分支
标签
贡献者
分支图
比较
Issue
0
Issue
0
列表
看板
标记
里程碑
合并请求
0
合并请求
0
Pages
DevOps
DevOps
流水线
流水线任务
计划
分析
分析
仓库分析
DevOps
Wiki
0
Wiki
成员
成员
收起侧边栏
关闭侧边栏
动态
分支图
创建新Issue
流水线任务
提交
Issue看板
体验新版 GitCode,发现更多精彩内容 >>
提交
ade3bbc0
编写于
2月 12, 2020
作者:
O
Olatunji Ruwase
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
Support legacy optimizer fusion as config option
上级
4f7d016d
变更
3
显示空白变更内容
内联
并排
Showing
3 changed file
with
29 addition
and
13 deletion
+29
-13
deepspeed/pt/deepspeed_config.py
deepspeed/pt/deepspeed_config.py
+9
-0
deepspeed/pt/deepspeed_constants.py
deepspeed/pt/deepspeed_constants.py
+2
-0
deepspeed/pt/deepspeed_light.py
deepspeed/pt/deepspeed_light.py
+18
-13
未找到文件。
deepspeed/pt/deepspeed_config.py
浏览文件 @
ade3bbc0
...
...
@@ -157,6 +157,14 @@ def get_optimizer_gradient_clipping(param_dict):
return
None
def
get_optimizer_legacy_fusion
(
param_dict
):
if
OPTIMIZER
in
param_dict
.
keys
()
and
\
LEGACY_FUSION
in
param_dict
[
OPTIMIZER
].
keys
():
return
param_dict
[
OPTIMIZER
][
LEGACY_FUSION
]
else
:
return
LEGACY_FUSION_DEFAULT
def
get_scheduler_name
(
param_dict
):
if
SCHEDULER
in
param_dict
.
keys
()
and
\
TYPE
in
param_dict
[
SCHEDULER
].
keys
():
...
...
@@ -261,6 +269,7 @@ class DeepSpeedConfig(object):
self
.
optimizer_name
=
self
.
optimizer_name
.
lower
()
self
.
optimizer_params
=
get_optimizer_params
(
param_dict
)
self
.
optimizer_legacy_fusion
=
get_optimizer_legacy_fusion
(
param_dict
)
self
.
scheduler_name
=
get_scheduler_name
(
param_dict
)
self
.
scheduler_params
=
get_scheduler_params
(
param_dict
)
...
...
deepspeed/pt/deepspeed_constants.py
浏览文件 @
ade3bbc0
...
...
@@ -24,6 +24,8 @@ OPTIMIZER = "optimizer"
OPTIMIZER_TYPE_DEFAULT
=
None
OPTIMIZER_PARAMS
=
"params"
TYPE
=
"type"
LEGACY_FUSION
=
"legacy_fusion"
LEGACY_FUSION_DEFAULT
=
True
SCHEDULER
=
"scheduler"
SCHEDULER_TYPE_DEFAULT
=
None
SCHEDULER_PARAMS
=
"params"
...
...
deepspeed/pt/deepspeed_light.py
浏览文件 @
ade3bbc0
...
...
@@ -211,6 +211,9 @@ class DeepSpeedLight(Module):
def
optimizer_params
(
self
):
return
self
.
_config
.
optimizer_params
def
optimizer_legacy_fusion
(
self
):
return
self
.
_config
.
optimizer_legacy_fusion
def
scheduler_name
(
self
):
return
self
.
_config
.
scheduler_name
...
...
@@ -411,21 +414,23 @@ class DeepSpeedLight(Module):
if
self
.
optimizer_name
()
==
ADAM_OPTIMIZER
:
if
self
.
dynamic_loss_scale
():
logging
.
info
(
'Creating fp16 optimizer with dynamic loss scale'
)
optimizer
=
FP16_Optimizer
(
optimizer
,
optimizer
=
FP16_Optimizer
(
optimizer
,
dynamic_loss_scale
=
True
,
initial_dynamic_scale
=
initial_dynamic_scale
,
dynamic_loss_args
=
dynamic_loss_args
,
mpu
=
self
.
mpu
,
clip_grad
=
clip_grad
,
fused_adam_legacy
=
True
)
fused_adam_legacy
=
self
.
optimizer_legacy_fusion
()
)
else
:
logging
.
info
(
'Creating fp16 optimizer with static loss scale: {}'
.
format
(
self
.
loss_scale
()))
optimizer
=
FP16_Optimizer
(
optimizer
,
optimizer
=
FP16_Optimizer
(
optimizer
,
static_loss_scale
=
self
.
loss_scale
(),
mpu
=
self
.
mpu
,
clip_grad
=
clip_grad
,
fused_adam_legacy
=
True
)
fused_adam_legacy
=
self
.
optimizer_legacy_fusion
()
)
else
:
logging
.
info
(
'Creating fp16 unfused optimizer with dynamic loss scale'
)
optimizer
=
FP16_UnfusedOptimizer
(
...
...
@@ -434,7 +439,7 @@ class DeepSpeedLight(Module):
dynamic_loss_args
=
dynamic_loss_args
,
mpu
=
self
.
mpu
,
clip_grad
=
clip_grad
,
fused_lamb_legacy
=
True
fused_lamb_legacy
=
self
.
optimizer_legacy_fusion
()
if
self
.
optimizer_name
()
==
LAMB_OPTIMIZER
else
False
)
return
optimizer
...
...
编辑
预览
Markdown
is supported
0%
请重试
或
添加新附件
.
添加附件
取消
You are about to add
0
people
to the discussion. Proceed with caution.
先完成此消息的编辑!
取消
想要评论请
注册
或
登录