【Hackathon No.47】为 Paddle cross 算子实现 float16 数据类型支持 #50924
Conversation
|
你的PR提交成功，感谢你对开源项目的贡献! |
|
Zhang Ting seems not to be a GitHub user. You need a GitHub account to be able to sign the CLA. If you have already a GitHub account, please add the email address used for this commit to your account. You have signed the CLA already but the status is still pending? Let us recheck it. |
|
|
可以参考#51168 |
| @@ -65,6 +65,21 @@ def init_output(self): | |||
| self.outputs = {'Out': np.array(z_list).reshape(self.shape)} | |||
|
|
|||
|
|
|||
| class TestCrossOpCase2(TestCrossOp): | |||
| def initTestCase(self): | |||
There was a problem hiding this comment.
需要参考低精度单测规范修改单测命名，TestCrossOpCase2 -> TestCrossFP16Op
https://www.paddlepaddle.org.cn/documentation/docs/zh/develop/dev_guides/amp_precision/amp_test_dev_guide_cn.html#optest
| self.outputs = {'Out': np.array(z_list).reshape(self.shape)} | ||
|
|
||
| def test_check_output(self): | ||
| self.check_output(atol=1e-3, check_eager=True) |
There was a problem hiding this comment.
test_check_output可以删掉，因为OpTest内部对fp16类型的阈值设置默认是1e-3。这里不需要为FP16实现这个方法
There was a problem hiding this comment.
老师，这里如果不设置atol=1e-3那么测试就会失败，设置后才能测试通过。不确定是不是内部对cross这个op单独设置了阈值。
There was a problem hiding this comment.
这个阈值大概也是近期刚刚默认加入的，是不是你开发环境的代码没有合入paddle的dev的内容，所以导致无法通过，我先在PR中删除，重新跑CI看看
|
@zhangting2020 这是中文文档pr:PaddlePaddle/docs#5708 |
|
请根据 PR-CI-Codestyle-Check 流水线提示修改下代码风格 |
|
hi, @denglianbin
|
PR types
Others
PR changes
Others
Describe
性能数据（op benchmark）