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

This commit is contained in:
SangBin Cho
2024-08-18 17:57:20 -07:00
committed by GitHub
parent 200a2ffa6b
commit ff7ec82c4d
36 changed files with 722 additions and 346 deletions
+7 -3
View File
@@ -1,3 +1,4 @@
from array import array
from functools import cached_property
from typing import (Any, Dict, Iterable, List, Literal, Mapping, Optional,
Tuple, TypedDict)
@@ -31,7 +32,8 @@ from vllm.model_executor.utils import set_weight_attrs
from vllm.multimodal import MULTIMODAL_REGISTRY
from vllm.multimodal.image import (cached_get_tokenizer,
repeat_and_pad_image_tokens)
from vllm.sequence import IntermediateTensors, SamplerOutput, SequenceData
from vllm.sequence import (VLLM_TOKEN_ID_ARRAY_TYPE, IntermediateTensors,
SamplerOutput, SequenceData)
from vllm.utils import print_warning_once
from .interfaces import SupportsMultiModal
@@ -70,8 +72,10 @@ def dummy_seq_data_for_chameleon(
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 * num_images
token_ids += array(VLLM_TOKEN_ID_ARRAY_TYPE,
[0]) * (seq_len - image_feature_size * num_images)
return SequenceData(token_ids)