Add checkpoint to quantize #28612
Add checkpoint to quantize #28612
Conversation
|
Thanks for your contribution! |
| def init_data(self): | ||
| self.bs = 8 | ||
| self.d_type = np.float32 | ||
| self.shape_x = [12, 1, 1] |
There was a problem hiding this comment.
Some time ago, baidu require test dimension to be over 100? Maybe self.shape_x = [12, 1, 10]
| self.enable_mkldnn = True | ||
|
|
||
| def test_check_output(self): | ||
| self.enable_mkldnn = True |
There was a problem hiding this comment.
There are two self.enable_mkldnn = True, the one in the init_data(self) is enough
The test shows that if self.check_output_with_option(use_gpu, flatten=True, bfloat16=False) then this test will do automatically float32? I thought settings should be done in init, before checking output. like self.enable_mkldnn_bfloat = True, But both are fine.
There was a problem hiding this comment.
Yes, I agree that just adding self.enable_mkldnn_bfloat = True will be better solution.
|
LGTM |
PR types
Bug fixes
PR changes
OPs
Describe
This PR :
quantize opwhere new attributebfloat16was added.cpu_bfloat16_passwhich is related to the change inquantize op.InferencePassTest, which allowed applying cpu_bfloat16_pass to the graph