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

vllm/model_executor/models/siglip.py

@@ -2,6 +2,7 @@
 within a vision language model."""
 
 import math
+from array import array
 from typing import Iterable, Optional, Tuple
 
 import torch
@@ -25,7 +26,7 @@ from vllm.model_executor.layers.vocab_parallel_embedding import (
 from vllm.model_executor.model_loader.weight_utils import default_weight_loader
 from vllm.multimodal.image import (cached_get_tokenizer,
                                    repeat_and_pad_image_tokens)
-from vllm.sequence import SequenceData
+from vllm.sequence import VLLM_TOKEN_ID_ARRAY_TYPE, SequenceData
 
 
 def get_siglip_patch_grid_length(*, image_size: int, patch_size: int) -> int:
@@ -62,8 +63,10 @@ def dummy_seq_data_for_siglip(
     else:
         image_feature_size = image_feature_size_override
 
-    token_ids = [image_token_id] * image_feature_size * num_images
-    token_ids += [0] * (seq_len - image_feature_size * num_images)
+    token_ids = array(VLLM_TOKEN_ID_ARRAY_TYPE,
+                      [image_token_id]) * image_feature_size
+    token_ids += array(VLLM_TOKEN_ID_ARRAY_TYPE,
+                       [0]) * (seq_len - image_feature_size)
     return SequenceData(token_ids)