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

[Core] Simplify and unify mm uuid handling & auto-generated mm hash overrides processing. (#24271)

Browse Source 
Signed-off-by: Chenheli Hua <huachenheli@outlook.com>
This commit is contained in:
Chenheli Hua
2025-09-09 21:36:09 -07:00
committed by GitHub
parent 0efdb5c3ba
commit 009d689b0c
13 changed files with 110 additions and 127 deletions
Download Patch File Download Diff File Expand all files Collapse all files

9
vllm/model_executor/models/deepseek_vl2.py
View File

@@ -21,7 +21,8 @@ from vllm.model_executor.model_loader.utils import set_default_torch_dtype
from vllm.model_executor.models.transformers import replace_linear_class
from vllm.multimodal import MULTIMODAL_REGISTRY
from vllm.multimodal.inputs import (MultiModalDataDict, MultiModalFieldConfig,
MultiModalKwargsItems, NestedTensors)
MultiModalKwargsItems, MultiModalUUIDDict,
NestedTensors)
from vllm.multimodal.parse import (ImageEmbeddingItems, ImageProcessorItems,
ImageSize, MultiModalDataItems)
from vllm.multimodal.processing import (BaseMultiModalProcessor,
@@ -290,7 +291,7 @@ class DeepseekVL2MultiModalProcessor(
mm_data_items: MultiModalDataItems,
hf_processor_mm_kwargs: Mapping[str, object],
tokenization_kwargs: Mapping[str, object],
mm_hash_overrides: Optional[dict[str, list[str]]] = None,
mm_uuids: Optional[MultiModalUUIDDict] = None,
) -> tuple[list[int], MultiModalProcessingInfo, bool]:
# The processor logic is different for len(images) <= 2 vs > 2
# Since the processing cache assumes that the processor output is
@@ -302,7 +303,7 @@ class DeepseekVL2MultiModalProcessor(
mm_data_items=mm_data_items,
hf_processor_mm_kwargs=hf_processor_mm_kwargs,
tokenization_kwargs=tokenization_kwargs,
mm_hash_overrides=mm_hash_overrides,
mm_uuids=mm_uuids,
)
return super()._cached_apply_hf_processor(
@@ -310,7 +311,7 @@ class DeepseekVL2MultiModalProcessor(
mm_data_items=mm_data_items,
hf_processor_mm_kwargs=hf_processor_mm_kwargs,
tokenization_kwargs=tokenization_kwargs,
mm_hash_overrides=mm_hash_overrides,
mm_uuids=mm_uuids,
)