[LoRA][Refactor] Unify multimodal LoRA token count hooks (#55071)
Signed-off-by: linitra24 <renshuang.zhou@daocloud.io>
This commit is contained in:
@@ -28,6 +28,7 @@ from vllm.multimodal import MULTIMODAL_REGISTRY
|
||||
from vllm.multimodal.cache import BaseMultiModalProcessorCache
|
||||
from vllm.multimodal.inputs import (
|
||||
MultiModalFieldConfig,
|
||||
MultiModalKwargsItem,
|
||||
MultiModalKwargsItems,
|
||||
)
|
||||
from vllm.multimodal.parse import (
|
||||
@@ -739,18 +740,12 @@ class LlavaForConditionalGeneration(
|
||||
tower_model="vision_tower",
|
||||
)
|
||||
|
||||
def get_num_mm_encoder_tokens(
|
||||
def get_mm_lora_token_counts(
|
||||
self,
|
||||
num_image_tokens: int,
|
||||
) -> int:
|
||||
# LLaVA's vision encoder outputs one token per patch without
|
||||
# spatial merging or pixel shuffle
|
||||
return num_image_tokens
|
||||
|
||||
def get_num_mm_connector_tokens(
|
||||
self,
|
||||
num_vision_tokens: int,
|
||||
) -> int:
|
||||
# LLaVA's MLP projector outputs the same number of tokens
|
||||
# as it receives from the vision encoder (1:1 mapping)
|
||||
return num_vision_tokens
|
||||
*,
|
||||
modality: str,
|
||||
mm_kwargs: MultiModalKwargsItem | None,
|
||||
num_mm_embeds: int,
|
||||
) -> tuple[int, int | None]:
|
||||
del modality, mm_kwargs
|
||||
return num_mm_embeds, num_mm_embeds
|
||||
|
||||
Reference in New Issue
Block a user