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
PaddlePaddle
Paddle
提交
69dd43d1
P
Paddle
项目概览
PaddlePaddle
/
Paddle
大约 1 年 前同步成功
通知
2292
Star
20931
Fork
5422
代码
文件
提交
分支
Tags
贡献者
分支图
Diff
Issue
1423
列表
看板
标记
里程碑
合并请求
543
Wiki
0
Wiki
分析
仓库
DevOps
项目成员
Pages
P
Paddle
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
1,423
Issue
1,423
列表
看板
标记
里程碑
合并请求
543
合并请求
543
Pages
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
提交
Issue看板
体验新版 GitCode,发现更多精彩内容 >>
未验证
提交
69dd43d1
编写于
3月 15, 2022
作者:
F
furnace
提交者:
GitHub
3月 15, 2022
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
[NPU] add AMP O1 support (#40362)
* [NPU] add AMP O1 support * [NPU] fix NOTE and warnings
上级
2c5edb4f
变更
4
隐藏空白更改
内联
并排
Showing
4 changed file
with
37 addition
and
14 deletion
+37
-14
paddle/fluid/imperative/amp_auto_cast.cc
paddle/fluid/imperative/amp_auto_cast.cc
+3
-1
paddle/fluid/pybind/op_function_generator.h
paddle/fluid/pybind/op_function_generator.h
+1
-0
python/paddle/fluid/dygraph/amp/auto_cast.py
python/paddle/fluid/dygraph/amp/auto_cast.py
+8
-3
python/paddle/fluid/dygraph/amp/loss_scaler.py
python/paddle/fluid/dygraph/amp/loss_scaler.py
+25
-10
未找到文件。
paddle/fluid/imperative/amp_auto_cast.cc
浏览文件 @
69dd43d1
...
...
@@ -209,7 +209,9 @@ inline bool NeedCast(const std::shared_ptr<VarType>& var) {
auto
data_type
=
GetDataType
<
VarType
>
(
var
);
if
(
paddle
::
platform
::
is_gpu_place
(
place
)
||
paddle
::
platform
::
is_cuda_pinned_place
(
place
)
||
paddle
::
platform
::
is_xpu_place
(
place
))
{
paddle
::
platform
::
is_xpu_place
(
place
)
||
paddle
::
platform
::
is_npu_place
(
place
)
||
paddle
::
platform
::
is_npu_pinned_place
(
place
))
{
// CudaPinndePlace is added for varbase created by dataloader
if
(
data_type
==
paddle
::
framework
::
proto
::
VarType
::
FP32
||
data_type
==
paddle
::
framework
::
proto
::
VarType
::
FP16
||
...
...
paddle/fluid/pybind/op_function_generator.h
浏览文件 @
69dd43d1
...
...
@@ -88,6 +88,7 @@ std::map<std::string, std::set<std::string>> op_ins_map = {
{
"nce"
,
{
"Input"
,
"Label"
,
"Weight"
,
"Bias"
,
"SampleWeight"
,
"CustomDistProbs"
,
"CustomDistAlias"
,
"CustomDistAliasProbs"
}},
{
"check_finite_and_unscale"
,
{
"X"
,
"Scale"
,
"FloatStatus"
}},
};
// NOTE(zhiqiu): Like op_ins_map.
...
...
python/paddle/fluid/dygraph/amp/auto_cast.py
浏览文件 @
69dd43d1
...
...
@@ -271,14 +271,19 @@ def amp_guard(enable=True,
"current_tracer is None, maybe it is not in imperative mode."
)
# check device_type:
# NOTE: Now, amp only support gpu for float16 and bfloat16, xpu for float16.
# NOTE: Now, amp only support gpu for float16 and bfloat16, xpu for float16
, npu for float16
.
# Maybe we will support cpu for bfloat16.
if
enable
and
not
(
tracer
.
_expected_place
.
is_gpu_place
()
or
tracer
.
_expected_place
.
is_xpu_place
()):
tracer
.
_expected_place
.
is_xpu_place
()
or
tracer
.
_expected_place
.
is_npu_place
()):
warnings
.
warn
(
'amp_guard can only be enabled on CUDAPlace
and X
PUPlace, current place is %s, so it makes no effect.'
'amp_guard can only be enabled on CUDAPlace
, XPUPlace, and N
PUPlace, current place is %s, so it makes no effect.'
%
tracer
.
_expected_place
)
enable
=
False
# For npu:
if
tracer
.
_expected_place
.
is_npu_place
()
and
(
dtype
==
'bfloat16'
):
warnings
.
warn
(
'NPUPlace only support float16 amp.'
)
enable
=
False
# For xpu:
if
tracer
.
_expected_place
.
is_xpu_place
()
and
(
dtype
==
'bfloat16'
):
warnings
.
warn
(
'XPUPlace only support float16 amp.'
)
...
...
python/paddle/fluid/dygraph/amp/loss_scaler.py
浏览文件 @
69dd43d1
...
...
@@ -105,9 +105,10 @@ class AmpScaler(object):
"current_tracer is None, maybe it is not in imperative mode."
)
if
enable
and
not
(
tracer
.
_expected_place
.
is_gpu_place
()
or
tracer
.
_expected_place
.
is_xpu_place
()):
tracer
.
_expected_place
.
is_xpu_place
()
or
tracer
.
_expected_place
.
is_npu_place
()):
warnings
.
warn
(
'AmpScaler can only be enabled on CUDAPlace
and X
PUPlace, current place is %s, so it makes no effect.'
'AmpScaler can only be enabled on CUDAPlace
, XPUPlace and N
PUPlace, current place is %s, so it makes no effect.'
%
tracer
.
_expected_place
)
enable
=
False
...
...
@@ -286,14 +287,28 @@ class AmpScaler(object):
)
and
(
param
.
_grad_ivar
().
dtype
==
core
.
VarDesc
.
VarType
.
FP32
)
]
if
len
(
param_grads_fp16
):
_C_ops
.
check_finite_and_unscale
(
param_grads_fp16
,
self
.
_scale
,
param_grads_fp16
,
self
.
_temp_found_inf_fp16
)
if
len
(
param_grads_fp32
):
_C_ops
.
check_finite_and_unscale
(
param_grads_fp32
,
self
.
_scale
,
param_grads_fp32
,
self
.
_temp_found_inf_fp32
)
if
core
.
is_compiled_with_npu
():
float_status
=
_C_ops
.
alloc_float_status
()
_C_ops
.
clear_float_status
(
float_status
,
float_status
)
if
len
(
param_grads_fp16
):
_C_ops
.
check_finite_and_unscale
(
param_grads_fp16
,
self
.
_scale
,
float_status
,
param_grads_fp16
,
self
.
_temp_found_inf_fp16
)
if
len
(
param_grads_fp32
):
_C_ops
.
check_finite_and_unscale
(
param_grads_fp32
,
self
.
_scale
,
float_status
,
param_grads_fp32
,
self
.
_temp_found_inf_fp32
)
else
:
if
len
(
param_grads_fp16
):
_C_ops
.
check_finite_and_unscale
(
param_grads_fp16
,
self
.
_scale
,
param_grads_fp16
,
self
.
_temp_found_inf_fp16
)
if
len
(
param_grads_fp32
):
_C_ops
.
check_finite_and_unscale
(
param_grads_fp32
,
self
.
_scale
,
param_grads_fp32
,
self
.
_temp_found_inf_fp32
)
if
len
(
param_grads_fp16
)
and
len
(
param_grads_fp32
):
self
.
_found_inf
=
self
.
_temp_found_inf_fp16
or
self
.
_temp_found_inf_fp32
elif
len
(
param_grads_fp16
):
...
...
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