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

[Core][VLM] Add precise multi-modal placeholder tracking (#8346)

Browse Source 
Signed-off-by: Peter Salas <peter@fixie.ai>
This commit is contained in:
Peter Salas
2024-11-01 16:21:10 -07:00
committed by GitHub
parent d151fde834
commit 6c0b7f548d
53 changed files with 913 additions and 281 deletions
Download Patch File Download Diff File Expand all files Collapse all files

8
vllm/model_executor/models/phi3v.py
View File

@@ -28,8 +28,8 @@ from transformers import CLIPVisionConfig, PretrainedConfig
from vllm.attention import AttentionMetadata
from vllm.config import (CacheConfig, ModelConfig, MultiModalConfig,
PoolerConfig)
from vllm.inputs import (INPUT_REGISTRY, DecoderOnlyInputs, InputContext,
token_inputs)
from vllm.inputs import (INPUT_REGISTRY, DecoderOnlyInputs, DummyData,
InputContext, token_inputs)
from vllm.logger import init_logger
from vllm.model_executor.layers.pooler import Pooler, PoolingType
from vllm.model_executor.layers.quantization import QuantizationConfig
@@ -380,7 +380,7 @@ def dummy_data_for_phi3v(ctx: InputContext,
image_feature_size = get_max_phi3v_image_tokens(ctx, num_crops=num_crops)
seq_data = dummy_seq_data_for_clip(
seq_data, ranges = dummy_seq_data_for_clip(
CLIP_VIT_LARGE_PATCH14_336_CONFIG,
seq_len,
num_images,
@@ -394,7 +394,7 @@ def dummy_data_for_phi3v(ctx: InputContext,
image_height_override=MAX_IMAGE_FEATURE_SIZE_HEIGHT,
)
return seq_data, mm_data
return DummyData(seq_data, mm_data, ranges)
@lru_cache