[Core] Optimize SPMD architecture with delta + serialization optimization (#7109)
This commit is contained in:
@@ -1,5 +1,6 @@
|
||||
"""Minimal implementation of BlipVisionModel intended to be only used
|
||||
within a vision language model."""
|
||||
from array import array
|
||||
from typing import Optional, Union
|
||||
|
||||
import torch
|
||||
@@ -16,7 +17,7 @@ from vllm.model_executor.layers.linear import (ColumnParallelLinear,
|
||||
from vllm.model_executor.layers.quantization import QuantizationConfig
|
||||
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_blip_patch_grid_length(*, image_size: int, patch_size: int) -> int:
|
||||
@@ -53,8 +54,10 @@ def dummy_seq_data_for_blip(
|
||||
else:
|
||||
image_feature_size = image_feature_size_override
|
||||
|
||||
token_ids = [image_token_id] * image_feature_size
|
||||
token_ids += [0] * (seq_len - image_feature_size)
|
||||
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)
|
||||
|
||||
|
||||
|
||||
Reference in New Issue
Block a user