[Core] Optimize SPMD architecture with delta + serialization optimization (#7109)

tests/spec_decode/utils.py

@@ -1,3 +1,4 @@
+from array import array
 from itertools import count
 from typing import Callable, Dict, List, Optional
 from typing import Sequence as GenericSequence
@@ -9,7 +10,8 @@ import torch
 from vllm.engine.arg_utils import EngineArgs
 from vllm.model_executor.utils import set_random_seed
 from vllm.sampling_params import SamplingParams
-from vllm.sequence import (CompletionSequenceGroupOutput, Logprob,
+from vllm.sequence import (VLLM_TOKEN_ID_ARRAY_TYPE,
+                           CompletionSequenceGroupOutput, Logprob,
                            SamplerOutput, SequenceData, SequenceGroupMetadata,
                            SequenceOutput)
 from vllm.utils import get_distributed_init_method, get_ip, get_open_port
@@ -138,8 +140,9 @@ def create_seq_group_metadata_from_prompts(
             seq_data={
                 i:
                 SequenceData(
-                    prompt_token_ids=prompt_token_ids[:],
-                    output_token_ids=cont_token_ids[:],
+                    array(VLLM_TOKEN_ID_ARRAY_TYPE, prompt_token_ids[:]),
+                    _output_token_ids=array(VLLM_TOKEN_ID_ARRAY_TYPE,
+                                            cont_token_ids[:]),
                 ),
             },
             sampling_params=SamplingParams(temperature=0.0, ),