[Bugfix] Fix OpenAI parallel sampling when using xgrammar (#11637)

Signed-off-by: mgoin <michael@neuralmagic.com>

vllm/model_executor/guided_decoding/xgrammar_decoding.py

@@ -1,6 +1,7 @@
 # noqa: UP007
 from __future__ import annotations
 
+import copy
 import json
 from dataclasses import dataclass, field
 from typing import TYPE_CHECKING, Any
@@ -309,3 +310,7 @@ class XGrammarLogitsProcessor:
             scores = scores.to(device_type).squeeze()
 
         return scores
+
+    def clone(self) -> XGrammarLogitsProcessor:
+        """Deepcopy due to per-sequence state in the matchers"""
+        return copy.deepcopy(self)

vllm/sampling_params.py

@@ -450,15 +450,16 @@ class SamplingParams(
         return self._all_stop_token_ids
 
     def clone(self) -> "SamplingParams":
-        """Deep copy excluding LogitsProcessor objects.
+        """Deep copy, but maybe not the LogitsProcessor objects.
 
-        LogitsProcessor objects are excluded because they may contain an
-        arbitrary, nontrivial amount of data.
+        LogitsProcessor objects may contain an arbitrary, nontrivial amount of
+        data that is expensive to copy. However, if not copied, the processor
+        needs to support parallel decoding for multiple sequences
         See https://github.com/vllm-project/vllm/issues/3087
         """
 
         logit_processor_refs = None if self.logits_processors is None else {
-            id(lp): lp
+            id(lp): lp.clone() if hasattr(lp, 'clone') else lp
             for lp in self.logits_processors
         }
         return copy.deepcopy(self, memo=logit_processor_refs)

vllm/sequence.py

@@ -1372,7 +1372,7 @@ class ParallelSampleSequenceGroup(SequenceGroupBase):
     @staticmethod
     def add_request(request_id: str, engine, params, **kwargs):
         original_params = params
-        params = copy.deepcopy(original_params)
+        params = original_params.clone()
         params.n = 1
         group = ParallelSampleSequenceGroup(request_id)
         seqs = []