[BugFix] Fix use of per-request seed with pipeline parallel (#6698)

vllm/sequence.py

@@ -411,14 +411,6 @@ class Sequence:
                 f"num_blocks={self.n_blocks}, ")
 
 
-@dataclass
-class SequenceGroupState:
-    """Mutable state tied to a specific sequence group"""
-
-    # torch.Generator used in seeded sampling
-    generator: Optional = None  # type: ignore
-
-
 class SequenceGroup:
     """A group of sequences that are generated from the same prompt.
 
@@ -461,7 +453,6 @@ class SequenceGroup:
                                       time_in_queue=None)
         self.lora_request = lora_request
         self.prompt_logprobs: Optional[PromptLogprobs] = None
-        self.state = SequenceGroupState()
         self.embeddings = embeddings
         self.pooling_params = pooling_params
         self.prompt_adapter_request = prompt_adapter_request
@@ -648,7 +639,6 @@ class SequenceGroupMetadata:
         lora_request: LoRA request.
         computed_block_nums: The block numbers that are already computed,
             used in prefix caching.
-        state: Internal state tied to this sequence group.
         multi_modal_data: Multi modal data.
         encoder_seq_data: Optional sequence data for encoder prompt
                           (SequenceGroup.encoder_seq). Should be None 
@@ -674,7 +664,6 @@ class SequenceGroupMetadata:
         token_chunk_size: Optional[int] = None,
         lora_request: Optional[LoRARequest] = None,
         computed_block_nums: Optional[List[int]] = None,
-        state: Optional[SequenceGroupState] = None,
         multi_modal_data: Optional["MultiModalDataDict"] = None,
         encoder_seq_data: Optional[SequenceData] = None,
         cross_block_table: Optional[List[int]] = None,
@@ -690,7 +679,6 @@ class SequenceGroupMetadata:
         self.prompt_adapter_request = prompt_adapter_request
         self.computed_block_nums = computed_block_nums
         self.multi_modal_data = multi_modal_data
-        self.state = SequenceGroupState() if state is None else state
         self.encoder_seq_data = encoder_seq_data
         self.cross_block_table = cross_block_table
         self._token_chunk_size = token_chunk_size