[VLM] Limit multimodal input cache by memory (#14805)

Signed-off-by: DarkLight1337 <tlleungac@connect.ust.hk>
This commit is contained in:
Cyrus Leung
2025-03-15 02:52:05 -07:00
committed by GitHub
parent 9ed6ee92d6
commit 3556a41434
13 changed files with 159 additions and 55 deletions
+2 -1
View File
@@ -18,6 +18,7 @@ from transformers.models.pixtral import PixtralProcessor
from vllm.config import VllmConfig
from vllm.inputs import InputProcessingContext
from vllm.jsontree import JSONTree, json_map_leaves
from vllm.model_executor.layers.activation import get_act_fn
from vllm.model_executor.layers.linear import (ColumnParallelLinear,
RowParallelLinear)
@@ -35,7 +36,7 @@ from vllm.multimodal.processing import (BaseMultiModalProcessor,
PromptReplacement, PromptUpdate)
from vllm.multimodal.profiling import BaseDummyInputsBuilder, ProcessorInputs
from vllm.sequence import IntermediateTensors
from vllm.utils import JSONTree, flatten_2d_lists, json_map_leaves
from vllm.utils import flatten_2d_lists
from .clip import CLIPVisionModel
from .interfaces import MultiModalEmbeddings, SupportsMultiModal, SupportsPP
+2 -1
View File
@@ -24,6 +24,7 @@ from vllm.distributed import (get_pp_group, get_tensor_model_parallel_rank,
get_tensor_model_parallel_world_size,
split_tensor_along_last_dim,
tensor_model_parallel_all_gather)
from vllm.jsontree import JSONTree, json_map_leaves
from vllm.model_executor import SamplingMetadata
from vllm.model_executor.layers.activation import (MulAndSilu, QuickGELU,
SiluAndMul)
@@ -50,7 +51,7 @@ from vllm.multimodal.processing import (BaseMultiModalProcessor,
PromptInsertion, PromptUpdate)
from vllm.multimodal.profiling import BaseDummyInputsBuilder, ProcessorInputs
from vllm.sequence import IntermediateTensors
from vllm.utils import JSONTree, flatten_2d_lists, json_map_leaves
from vllm.utils import flatten_2d_lists
from .interfaces import (MultiModalEmbeddings, SupportsLoRA,
SupportsMultiModal, SupportsPP, SupportsQuant)