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

Generate: remove flakyness in test_generate_from_inputs_embeds_decoder_only (#33602)

Browse Source 
almost zero is not zero
This commit is contained in:
Joao Gante
2024-09-20 13:50:42 +01:00
committed by GitHub
parent ec1424c6a3
commit 266d0a6375
1 changed files with 28 additions and 12 deletions
Download Patch File Download Diff File Expand all files Collapse all files

40
tests/generation/test_utils.py
View File

@@ -1647,26 +1647,42 @@ class GenerationTesterMixin:
continue continue
# Traditional way of generating text # Traditional way of generating text
outputs_from_ids = model.generate(input_ids, max_new_tokens=5) outputs_from_ids = model.generate(
self.assertEqual(outputs_from_ids.shape, (input_ids.shape[0], input_ids.shape[1] + 5)) input_ids, max_new_tokens=5, return_dict_in_generate=True, output_scores=True
)
self.assertEqual(outputs_from_ids.sequences.shape, (input_ids.shape[0], input_ids.shape[1] + 5))
# Same thing, but from input embeddings (`input_ids` is passed so the prompt is present in the output) # Same thing, but from input embeddings (`input_ids` is passed so the prompt is present in the output)
inputs_embeds = model.get_input_embeddings()(input_ids) inputs_embeds = model.get_input_embeddings()(input_ids)
outputs_from_embeds = model.generate(input_ids, inputs_embeds=inputs_embeds, max_new_tokens=5) outputs_from_embeds = model.generate(
self.assertListEqual(outputs_from_ids.tolist(), outputs_from_embeds.tolist()) input_ids,
inputs_embeds=inputs_embeds,
max_new_tokens=5,
return_dict_in_generate=True,
output_scores=True,
)
self.assertListEqual(outputs_from_ids.sequences.tolist(), outputs_from_embeds.sequences.tolist())
# But if we pass different inputs_embeds, we should get different outputs # But if we pass different inputs_embeds, we should get different outputs (the output text may be the
torch.manual_seed(0) # same, but the logits will almost surely be different)
random_embeds = torch.rand_like(inputs_embeds) random_embeds = torch.rand_like(inputs_embeds)
outputs_from_rand_embeds = model.generate(input_ids, inputs_embeds=random_embeds, max_new_tokens=5) outputs_from_rand_embeds = model.generate(
with self.assertRaises(AssertionError): input_ids,
self.assertListEqual(outputs_from_rand_embeds.tolist(), outputs_from_embeds.tolist()) inputs_embeds=random_embeds,
max_new_tokens=5,
return_dict_in_generate=True,
output_scores=True,
)
for i in range(len(outputs_from_rand_embeds.scores)):
self.assertFalse(torch.allclose(outputs_from_embeds.scores[i], outputs_from_rand_embeds.scores[i]))
# input_ids is not a required input -- if we don't pass it, the newly generated tokens will be the same # input_ids is not a required input -- if we don't pass it, the newly generated tokens will be the same
outputs_from_embeds_wo_ids = model.generate(inputs_embeds=inputs_embeds, max_new_tokens=5) outputs_from_embeds_wo_ids = model.generate(
inputs_embeds=inputs_embeds, max_new_tokens=5, return_dict_in_generate=True, output_scores=True
)
self.assertListEqual( self.assertListEqual(
outputs_from_embeds[:, inputs_embeds.shape[1] :].tolist(), outputs_from_embeds.sequences[:, inputs_embeds.shape[1] :].tolist(),
outputs_from_embeds_wo_ids.tolist(), outputs_from_embeds_wo_ids.sequences.tolist(),
) )
@pytest.mark.generate @pytest.mark.generate