[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
+9 -14
View File
@@ -39,6 +39,7 @@ from vllm.model_executor.models.module_mapping import MultiModelKeys
from vllm.multimodal import MULTIMODAL_REGISTRY
from vllm.multimodal.inputs import (
MultiModalFieldConfig,
MultiModalKwargsItem,
MultiModalKwargsItems,
)
from vllm.multimodal.parse import ImageProcessorItems, MultiModalDataItems
@@ -837,20 +838,14 @@ class Idefics3ForConditionalGeneration(
tower_model="model.vision_model",
)
def get_num_mm_encoder_tokens(
def get_mm_lora_token_counts(
self,
num_image_tokens: int,
) -> int:
*,
modality: str,
mm_kwargs: MultiModalKwargsItem | None,
num_mm_embeds: int,
) -> tuple[int, int | None]:
del modality, mm_kwargs
hf_config = self.config
scale_factor = hf_config.scale_factor
return num_image_tokens * scale_factor**2
def get_num_mm_connector_tokens(
self,
num_vision_tokens: int,
) -> int:
hf_config = self.config
scale_factor = hf_config.scale_factor
return num_vision_tokens // scale_factor**2
return num_mm_embeds * scale_factor**2, num_mm_embeds