[LoRA][Refactor] Unify multimodal LoRA token count hooks (#55071)

Signed-off-by: linitra24 <renshuang.zhou@daocloud.io>
This commit is contained in:
linitra24
2026-09-14 13:19:15 +00:00
committed by GitHub
parent 6623fe5b4e
commit 9d4d9aa5bc
23 changed files with 276 additions and 295 deletions
+10 -6
View File
@@ -54,6 +54,7 @@ from vllm.model_executor.models.utils import (
)
from vllm.model_executor.models.vision import get_vit_attn_backend
from vllm.multimodal import MULTIMODAL_REGISTRY
from vllm.multimodal.inputs import MultiModalKwargsItem
from vllm.sequence import IntermediateTensors
from vllm.transformers_utils.configs.dotsocr import DotsOCRConfig, DotsVisionConfig
from vllm.utils.tensor_schema import TensorSchema, TensorShape
@@ -718,13 +719,16 @@ class DotsOCRForCausalLM(nn.Module, SupportsMultiModal, SupportsPP, SupportsLoRA
return image_embeds.split(sizes)
def get_num_mm_encoder_tokens(self, num_image_tokens: int) -> int:
def get_mm_lora_token_counts(
self,
*,
modality: str,
mm_kwargs: MultiModalKwargsItem | None,
num_mm_embeds: int,
) -> tuple[int, int | None]:
del modality, mm_kwargs
merge_size = self.vision_tower.spatial_merge_size
return num_image_tokens * (merge_size**2)
def get_num_mm_connector_tokens(self, num_vision_tokens: int) -> int:
merge_size = self.vision_tower.spatial_merge_size
return num_vision_tokens // (merge_size**2)
return num_mm_embeds * (merge_size**2), num_mm_embeds
def embed_multimodal(self, **kwargs: object) -> MultiModalEmbeddings:
image_input = self._parse_and_validate_image_input(**kwargs)