SUMIN/HuggingFace_transformer
1
0
Fork 0
Code Issues Pull Requests Actions 7 Packages Projects Releases Wiki Activity

Fix expected value for OPT test test_inference_no_head (#17395)

Browse Source 
* Fix expected value

* 5e-5

Co-authored-by: ydshieh <ydshieh@users.noreply.github.com>
This commit is contained in:
Yih-Dar
2022-05-25 11:19:06 +02:00
committed by GitHub
parent 1ef9a1ed4a
commit 4d727bd2df
1 changed files with 7 additions and 10 deletions
Download Patch File Download Diff File Expand all files Collapse all files

17
tests/models/opt/test_modeling_opt.py
View File

@@ -277,19 +277,16 @@ class OPTModelIntegrationTests(unittest.TestCase):
expected_shape = torch.Size((1, 11, 512))
self.assertEqual(output.shape, expected_shape)
# expected value works for CPU, as well as GPU (with TF32 disabled)
expected_slice = torch.tensor(
[[-0.2873, -1.9242, -0.3059], [-1.2738, -0.1333, -0.1877], [0.4116, 0.1192, -1.3107]],
[
[-0.28726277, -1.9241608, -0.3058734],
[-1.2737825, -0.13332152, -0.18766522],
[0.41159445, 0.1191957, -1.3107123],
],
device=torch_device,
)
# Getting different logits results on GPU depending on PyTorch version (1.10+cu11.0 vs. 1.11+cu11.4)
# and results also differ between CPU and GPU. Only on CPU it seems to be deterministic.
# It's not because the weights are saved & loaded in FP16
# checked that the same happens when weights are stored in fp32 and loaded in fp32.
# The differences start to creep in in the first linear projection matrix project_in_dim
# It however also happens for BART (maybe related to training model in fp16?)
atol = 1e-2 if torch_device != "cpu" else 1e-3
assert_tensors_close(output[0, :3, :3], expected_slice, atol=atol)
assert_tensors_close(output[0, :3, :3], expected_slice, atol=5e-5)
@require_torch