Generate: fix flaky tests (#27543)

src/transformers/generation/logits_process.py

@@ -1301,8 +1301,9 @@ class InfNanRemoveLogitsProcessor(LogitsProcessor):
         # set all nan values to 0.0
         scores[scores != scores] = 0.0
 
-        # set all inf values to max possible value
+        # set all +/-inf values to max/min possible value
         scores[scores == float("inf")] = torch.finfo(scores.dtype).max
+        scores[scores == float("-inf")] = torch.finfo(scores.dtype).min
 
         return scores
 

tests/generation/test_logits_process.py

@@ -692,7 +692,7 @@ class LogitsProcessorTest(unittest.TestCase):
             torch.allclose(
                 scores,
                 torch.tensor(
-                    [[0.0, 0.7, 0.8, 0.0], [0.1, torch.finfo(scores.dtype).max, 0.3, float("-inf")]],
+                    [[0.0, 0.7, 0.8, 0.0], [0.1, torch.finfo(scores.dtype).max, 0.3, torch.finfo(scores.dtype).min]],
                     device=torch_device,
                 ),
                 atol=1e-6,

tests/generation/test_utils.py

@@ -124,9 +124,14 @@ class GenerationTesterMixin:
         process_kwargs = {
             "min_length": input_length + 1 if max_length is None else max_length - 1,
             "bad_words_ids": [[1, 0]],
-            "no_repeat_ngram_size": 2,
             "repetition_penalty": 1.2,
+            "remove_invalid_values": True,
         }
+        # NoRepeatNGramLogitsProcessor + forced tokens may result in no valid continuations
+        if forced_bos_token_id is None and forced_eos_token_id is None:
+            process_kwargs["no_repeat_ngram_size"] = 2
+
+        # NOTE: the order of operations here should match `generate` for accurate testing
         logits_processor = LogitsProcessorList(
             (
                 [
@@ -154,12 +159,16 @@ class GenerationTesterMixin:
                 if forced_eos_token_id is not None
                 else []
             )
-            + [
-                NoBadWordsLogitsProcessor(process_kwargs["bad_words_ids"], eos_token_id),
-                NoRepeatNGramLogitsProcessor(process_kwargs["no_repeat_ngram_size"]),
-                RepetitionPenaltyLogitsProcessor(process_kwargs["repetition_penalty"]),
-            ]
+            + [NoBadWordsLogitsProcessor(process_kwargs["bad_words_ids"], eos_token_id)]
+            + (
+                [NoRepeatNGramLogitsProcessor(process_kwargs["no_repeat_ngram_size"])]
+                if forced_bos_token_id is None and forced_eos_token_id is None
+                else []
+            )
+            + [RepetitionPenaltyLogitsProcessor(process_kwargs["repetition_penalty"])]
+            + [InfNanRemoveLogitsProcessor()]  # prevent flaky generation test failures
         )
+
         return process_kwargs, logits_processor
 
     @staticmethod
@@ -282,7 +291,6 @@ class GenerationTesterMixin:
             output_hidden_states=output_hidden_states,
             output_scores=output_scores,
             return_dict_in_generate=return_dict_in_generate,
-            remove_invalid_values=True,
             **logits_process_kwargs,
             **model_kwargs,
         )
@@ -340,7 +348,6 @@ class GenerationTesterMixin:
             output_attentions=output_attentions,
             output_hidden_states=output_hidden_states,
             return_dict_in_generate=return_dict_in_generate,
-            remove_invalid_values=True,
             **logits_warper_kwargs,
             **process_kwargs,
             **model_kwargs,
@@ -361,9 +368,6 @@ class GenerationTesterMixin:
         elif attention_mask is not None:
             attention_mask = attention_mask.repeat_interleave(num_return_sequences, dim=0)
 
-        # prevent flaky generation test failures
-        logits_processor.append(InfNanRemoveLogitsProcessor())
-
         with torch.no_grad():
             model_kwargs = {"attention_mask": attention_mask} if attention_mask is not None else {}
             output_sample = model.sample(
@@ -405,7 +409,6 @@ class GenerationTesterMixin:
             output_attentions=output_attentions,
             output_hidden_states=output_hidden_states,
             return_dict_in_generate=return_dict_in_generate,
-            remove_invalid_values=True,
             **beam_kwargs,
             **logits_process_kwargs,
             **model_kwargs,
@@ -467,7 +470,6 @@ class GenerationTesterMixin:
             output_attentions=output_attentions,
             output_hidden_states=output_hidden_states,
             return_dict_in_generate=return_dict_in_generate,
-            remove_invalid_values=True,
             **beam_kwargs,
             **logits_warper_kwargs,
             **model_kwargs,
@@ -534,7 +536,6 @@ class GenerationTesterMixin:
             output_attentions=output_attentions,
             output_hidden_states=output_hidden_states,
             return_dict_in_generate=return_dict_in_generate,
-            remove_invalid_values=True,
             **beam_kwargs,
             **logits_process_kwargs,
             **model_kwargs,
@@ -596,7 +597,6 @@ class GenerationTesterMixin:
             output_attentions=output_attentions,
             output_hidden_states=output_hidden_states,
             return_dict_in_generate=return_dict_in_generate,
-            remove_invalid_values=True,
             constraints=constraints,
             **beam_kwargs,
             **logits_process_kwargs,
@@ -671,7 +671,6 @@ class GenerationTesterMixin:
             output_hidden_states=output_hidden_states,
             output_scores=output_scores,
             return_dict_in_generate=return_dict_in_generate,
-            remove_invalid_values=True,
             **logits_process_kwargs,
             **model_kwargs,
             **contrastive_search_kwargs,
@@ -1284,13 +1283,8 @@ class GenerationTesterMixin:
 
             # check `generate()` and `constrained_beam_search()` are equal
             # Sample constraints
-            if not input_ids.dtype == torch.float32:
-                min_id = torch.min(input_ids) + 3
-                max_id = torch.max(input_ids)
-            else:
-                # otherwise this throws an error for Speech2TextModel since its inputs are floating points
-                min_id = 3
-                max_id = 100
+            min_id = 3
+            max_id = config.vocab_size
 
             force_tokens = torch.randint(min_id, max_id, (1, 2)).tolist()[0]
             constraints = [