Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

[XPU]correct adamw bf16 unit test and the way to get data type #60565

Merged
merged 2 commits into from
Jan 10, 2024
Merged

[XPU]correct adamw bf16 unit test and the way to get data type #60565

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
72db2a5
correct adamw bf16 unit test and the way to get data type
runzhech Jan 4, 2024
ebbcbd0
Merge branch 'PaddlePaddle:develop' into develop
runzhech Jan 4, 2024
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
11 changes: 10 additions & 1 deletion test/xpu/op_test_xpu.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -106,8 +106,13 @@ def check_output_with_place(
if not core.is_float16_supported(place):
return

if self.dtype == np.float16:
if self.dtype == np.uint16:
if not core.is_bfloat16_supported(place):
return

if self.dtype == np.float16 or self.dtype == np.uint16:
atol = 0.1
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

原始文件里float16统一定为0.1是什么样的考虑啊，实际上bf16很多算子精度都高于这个，fp16按理说应该是更高的呀

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

原来那个pr也没细讲，可能是kl2不支持fp16所以经常截断导致的误差吧，这个之后确实可以改一下，0.1有点大了

Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

应该是统一设小一点，遇到过不了的特例在对应的单测里改大点，后边有时间我们可以统一改下，感觉可能会扫出来一些过不了的


return super().check_output_with_place(
place,
atol,
Expand Down Expand Up @@ -183,6 +188,10 @@ def check_grad_with_place(
if not core.is_float16_supported(place):
return

if self.dtype == np.uint16:
if not core.is_bfloat16_supported(place):
return

if self.dtype == np.float16 or self.dtype == np.uint16:
max_relative_error = 0.1
return super().check_grad_with_place(
Expand Down
22 changes: 19 additions & 3 deletions test/xpu/test_adamw_op_xpu.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -21,6 +21,7 @@
create_test_class,
get_xpu_op_support_types,
)
from op_test import convert_float_to_uint16
from op_test_xpu import XPUOpTest

import paddle
Expand Down Expand Up @@ -85,8 +86,8 @@ def setUp(self):
self.op_type = "adamw"
self.init_shape()
self.dtype = self.in_type
param = np.random.uniform(-1, 1, self.shape).astype(self.dtype)
grad = np.random.uniform(-1, 1, self.shape).astype(self.dtype)
param = np.random.uniform(-1, 1, self.shape)
grad = np.random.uniform(-1, 1, self.shape)
moment1 = np.random.uniform(-1, 1, self.shape).astype("float32")
# The second moment is positive
moment2 = np.random.random(self.shape).astype("float32")
Expand All @@ -97,7 +98,9 @@ def setUp(self):
epsilon = 1e-4
beta1_pow = beta1**10
beta2_pow = beta2**10

if self.dtype != np.uint16:
param = param.astype(self.dtype)
grad = grad.astype(self.dtype)
self.inputs = {
'Param': param,
'Grad': grad,
Expand Down Expand Up @@ -128,13 +131,26 @@ def setUp(self):
'Beta2PowOut': np.array([beta2_pow]).astype("float32") * beta2,
}

if self.dtype == np.uint16:
self.inputs['Param'] = convert_float_to_uint16(
self.inputs['Param']
)
self.inputs['Grad'] = convert_float_to_uint16(
self.inputs['Grad']
)
self.outputs['ParamOut'] = convert_float_to_uint16(param_out)

def init_shape(self):
self.shape = [102, 105]

def test_check_output(self):
paddle.enable_static()
self.check_output_with_place(place=paddle.XPUPlace(0))

def infer_dtype_from_inputs_outputs(self, inputs, outputs):
self.__class__.dtype = self.dtype
self.output_dtype = self.dtype

class TestAdamW2(TestAdamW):
def init_shape(self):
self.shape = [
Expand Down
7 changes: 4 additions & 3 deletions test/xpu/test_elementwise_sub_op_xpu.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -66,16 +66,17 @@ def init_input_output(self):
if self.dtype == np.uint16:
tmp_x = self.reshape_data(self.x, self.y)
tmp_y = self.reshape_data(self.y, self.x)
self.outputs = {'Out': tmp_x - tmp_y}
tmp_out = tmp_x - tmp_y
self.outputs = {'Out': convert_float_to_uint16(tmp_out)}
self.x = convert_float_to_uint16(self.x)
self.y = convert_float_to_uint16(self.y)
else:
tmp_x = self.reshape_data(self.x, self.y).astype(self.dtype)
tmp_y = self.reshape_data(self.y, self.x).astype(self.dtype)
self.outputs = {'Out': tmp_x - tmp_y}
self.inputs = {
'X': self.x,
'Y': self.y,
'X': self.x.astype(self.dtype),
'Y': self.y.astype(self.dtype),
}

def init_shape(self):
Expand Down