biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

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

Browse Source
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
Download Patch File Download Diff File Expand all files Collapse all files

9
vllm/model_executor/models/clip.py
View File

@@ -1,5 +1,6 @@
"""Minimal implementation of CLIPVisionModel intended to be only used
within a vision language model."""
from array import array
from typing import Iterable, Optional, Tuple
import torch
@@ -17,7 +18,7 @@ from vllm.model_executor.layers.quantization import QuantizationConfig
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_clip_patch_grid_length(*, image_size: int, patch_size: int) -> int:
@@ -53,8 +54,10 @@ def dummy_seq_data_for_clip(
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)