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

[OPT] Fix default attention mask size (#22649)

Browse Source 
* Fix default attention mask size

* fixup

* add a test to make sure that even if attention mask are not provided, works

* style
This commit is contained in:
Arthur
2023-04-07 20:12:57 +02:00
committed by GitHub
parent b1b3dc3e52
commit f33419261a
2 changed files with 23 additions and 8 deletions
Download Patch File Download Diff File Expand all files Collapse all files

13
tests/models/opt/test_modeling_opt.py
View File

@@ -182,6 +182,19 @@ class OPTModelTester:
# test that outputs are equal for slice
self.parent.assertTrue(torch.allclose(output_from_past_slice, output_from_no_past_slice, atol=1e-3))
# test no attention_mask works
outputs = model(input_ids, attention_mask=attention_mask, head_mask=head_mask, use_cache=True)
_, past_key_values = outputs.to_tuple()
output_from_no_past = model(next_input_ids)["last_hidden_state"]
output_from_past = model(next_tokens, past_key_values=past_key_values)["last_hidden_state"]
random_slice_idx = ids_tensor((1,), output_from_past.shape[-1]).item()
output_from_no_past_slice = output_from_no_past[:, -3:, random_slice_idx].detach()
output_from_past_slice = output_from_past[:, :, random_slice_idx].detach()
# test that outputs are equal for slice
self.parent.assertTrue(torch.allclose(output_from_past_slice, output_from_no_past_slice, atol=1e-3))
@require_torch
class OPTModelTest(ModelTesterMixin, GenerationTesterMixin, PipelineTesterMixin, unittest.TestCase):