fix tests s.t. they mock forgetting the logits processor

tests/serve/test_vllm.py

@@ -1,5 +1,6 @@
 import re
 
+import pytest
 import torch
 
 from outlines.serve.vllm import RegexLogitsProcessor, _patched_apply_logits_processors
@@ -41,14 +42,21 @@ def sample_from_logits(logits):
     return torch.multinomial(probs, 1).item()
 
 
-def test_time_regexp():
+@pytest.mark.parametrize("forget_logits_processor", [True, False])
+def test_time_regexp(forget_logits_processor):
     pattern = r"(0?[1-9]|1[0-2]):[0-5]\d\s?(am|pm)?"
     llm = MockModel()
     logits_processor = RegexLogitsProcessor(pattern, llm)
 
     token_ids = []
     while True:
         random_scores = -10 + 20 * torch.rand(len(llm.tokenizer.vocabulary))
+
+        # mock "forgetting" the logits processor behavior in
+        # vLLM tensor-parallel world size > 1
+        if forget_logits_processor:
+            logits_processor = RegexLogitsProcessor(pattern, llm)
+
         logits = logits_processor(
             input_ids=token_ids,
             scores=random_scores,