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
34a11688
D
DeepSpeed
项目概览
Greenplum
/
DeepSpeed
上一次同步 大约 1 年
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 >>
未验证
提交
34a11688
编写于
1月 24, 2023
作者:
L
loadams
提交者:
GitHub
1月 24, 2023
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
Change zero_grad() argument to match pytorch (#2741)
上级
867da307
变更
4
隐藏空白更改
内联
并排
Showing
4 changed file
with
12 addition
and
12 deletion
+12
-12
deepspeed/runtime/fp16/fused_optimizer.py
deepspeed/runtime/fp16/fused_optimizer.py
+2
-2
deepspeed/runtime/fp16/unfused_optimizer.py
deepspeed/runtime/fp16/unfused_optimizer.py
+2
-2
deepspeed/runtime/zero/stage3.py
deepspeed/runtime/zero/stage3.py
+4
-4
deepspeed/runtime/zero/stage_1_and_2.py
deepspeed/runtime/zero/stage_1_and_2.py
+4
-4
未找到文件。
deepspeed/runtime/fp16/fused_optimizer.py
浏览文件 @
34a11688
...
...
@@ -130,14 +130,14 @@ class FP16_Optimizer(DeepSpeedOptimizer):
return
def
zero_grad
(
self
,
set_
grads_to_None
=
Tru
e
):
def
zero_grad
(
self
,
set_
to_none
=
Fals
e
):
"""
Zero FP16 parameter grads.
"""
# For speed, set model fp16 grad to None by default
for
group
in
self
.
fp16_groups
:
for
p
in
group
:
if
set_
grads_to_N
one
:
if
set_
to_n
one
:
p
.
grad
=
None
else
:
if
p
.
grad
is
not
None
:
...
...
deepspeed/runtime/fp16/unfused_optimizer.py
浏览文件 @
34a11688
...
...
@@ -110,7 +110,7 @@ class FP16_UnfusedOptimizer(DeepSpeedOptimizer):
self
.
initialize_optimizer_states
()
def
zero_grad
(
self
,
set_
grads_to_None
=
Tru
e
):
def
zero_grad
(
self
,
set_
to_none
=
Fals
e
):
"""
Zero FP16 parameter grads.
"""
...
...
@@ -118,7 +118,7 @@ class FP16_UnfusedOptimizer(DeepSpeedOptimizer):
# For speed, set model fp16 grad to None by default
for
group
in
self
.
fp16_groups
:
for
p
in
group
:
if
set_
grads_to_N
one
:
if
set_
to_n
one
:
p
.
grad
=
None
else
:
if
p
.
grad
is
not
None
:
...
...
deepspeed/runtime/zero/stage3.py
浏览文件 @
34a11688
...
...
@@ -1509,7 +1509,7 @@ class DeepSpeedZeroOptimizer_Stage3(ZeROOptimizer):
return
params_in_partition
,
params_not_in_partition
,
first_offset
@
instrument_w_nvtx
def
zero_grad
(
self
,
set_
grads_to_None
=
Tru
e
):
def
zero_grad
(
self
,
set_
to_none
=
Fals
e
):
"""
Zero FP16 parameter grads.
"""
...
...
@@ -1519,7 +1519,7 @@ class DeepSpeedZeroOptimizer_Stage3(ZeROOptimizer):
# For speed, set model fp16 grad to None by default
for
group
in
self
.
fp16_groups
:
for
p
in
group
:
if
set_
grads_to_N
one
:
if
set_
to_n
one
:
if
p
.
grad
is
not
None
and
p
.
grad
.
is_cuda
:
p
.
grad
.
record_stream
(
torch
.
cuda
.
current_stream
())
p
.
grad
=
None
...
...
@@ -1708,7 +1708,7 @@ class DeepSpeedZeroOptimizer_Stage3(ZeROOptimizer):
self
.
fp32_partitioned_groups_flat
[
sub_group_id
].
grad
=
single_grad_partition
# release all the gradient since we have already created a necessary copy in dp_grad_partition
self
.
zero_grad
()
self
.
zero_grad
(
set_to_none
=
True
)
for
grad
in
filter
(
lambda
g
:
g
.
is_cuda
,
self
.
averaged_gradients
[
sub_group_id
]):
grad
.
record_stream
(
torch
.
cuda
.
current_stream
())
...
...
@@ -1816,7 +1816,7 @@ class DeepSpeedZeroOptimizer_Stage3(ZeROOptimizer):
def
_overflow_clean_up
(
self
,
prev_scale
):
see_memory_usage
(
'After overflow before clearing gradients'
,
force
=
False
)
self
.
zero_grad
()
self
.
zero_grad
(
set_to_none
=
True
)
if
self
.
offload_optimizer
:
self
.
reset_cpu_buffers
()
...
...
deepspeed/runtime/zero/stage_1_and_2.py
浏览文件 @
34a11688
...
...
@@ -753,7 +753,7 @@ class DeepSpeedZeroOptimizer(ZeROOptimizer):
# No need to keep the gradients anymore.
# All gradients required by the step
# are in self.averaged_gradients
self
.
zero_grad
()
self
.
zero_grad
(
set_to_none
=
True
)
see_memory_usage
(
f
"End ipg_epilogue"
)
# resets all partition to no reduced
...
...
@@ -1526,7 +1526,7 @@ class DeepSpeedZeroOptimizer(ZeROOptimizer):
return
params_in_partition
,
params_not_in_partition
,
first_offset
def
zero_grad
(
self
,
set_
grads_to_None
=
Tru
e
):
def
zero_grad
(
self
,
set_
to_none
=
Fals
e
):
"""
Zero FP16 parameter grads.
"""
...
...
@@ -1534,7 +1534,7 @@ class DeepSpeedZeroOptimizer(ZeROOptimizer):
# For speed, set model fp16 grad to None by default
for
group
in
self
.
bit16_groups
:
for
p
in
group
:
if
set_
grads_to_N
one
:
if
set_
to_n
one
:
p
.
grad
=
None
# epilogue and in step
else
:
if
p
.
grad
is
not
None
:
...
...
@@ -1766,7 +1766,7 @@ class DeepSpeedZeroOptimizer(ZeROOptimizer):
self
.
loss_scale
))
see_memory_usage
(
'After overflow before clearing gradients'
)
self
.
zero_grad
()
self
.
zero_grad
(
set_to_none
=
True
)
if
self
.
cpu_offload
:
self
.
reset_cpu_buffers
()
else
:
...
...
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