diff --git a/vllm/config/lora.py b/vllm/config/lora.py index e5c1d14f6d..f32b89ae59 100644 --- a/vllm/config/lora.py +++ b/vllm/config/lora.py @@ -25,7 +25,6 @@ logger = init_logger(__name__) LoRADType = Literal["auto", "float16", "bfloat16"] MaxLoRARanks = Literal[1, 8, 16, 32, 64, 128, 256, 320, 512] -LoRAExtraVocabSize = Literal[256, 512] @config(config=ConfigDict(arbitrary_types_allowed=True)) diff --git a/vllm/lora/model_manager.py b/vllm/lora/model_manager.py index adc2353b3e..802e5b9233 100644 --- a/vllm/lora/model_manager.py +++ b/vllm/lora/model_manager.py @@ -328,8 +328,8 @@ class LoRAModelManager: self.punica_wrapper_mapping[prefix] = connector_punica_wrapper else: logger.warning_once( - "Connector LoRA support disabled: model does not implement " - "get_num_mm_connector_tokens(). This method is required to " + "Connector LoRA support disabled: get_mm_lora_token_counts() " + "returned no connector token counts, which are required to " "determine the connector's token budget for LoRA operations." ) diff --git a/vllm/model_executor/models/bailing_moe_v3_vl.py b/vllm/model_executor/models/bailing_moe_v3_vl.py index a236d0722b..e596c95ddf 100644 --- a/vllm/model_executor/models/bailing_moe_v3_vl.py +++ b/vllm/model_executor/models/bailing_moe_v3_vl.py @@ -26,6 +26,7 @@ from vllm.multimodal import MULTIMODAL_REGISTRY from vllm.multimodal.inputs import ( MultiModalFeatureSpec, MultiModalFieldConfig, + MultiModalKwargsItem, MultiModalKwargsItems, ) from vllm.multimodal.parse import MultiModalDataItems @@ -520,8 +521,12 @@ class BailingMoeV3VLForConditionalGeneration( tower_model="visual.", ) - def get_num_mm_encoder_tokens(self, num_image_tokens: int) -> int: - return num_image_tokens * self.visual.spatial_merge_size**2 - - def get_num_mm_connector_tokens(self, num_vision_tokens: int) -> int: - return num_vision_tokens // self.visual.spatial_merge_size**2 + 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 + return num_mm_embeds * self.visual.spatial_merge_size**2, num_mm_embeds diff --git a/vllm/model_executor/models/cohere_compass.py b/vllm/model_executor/models/cohere_compass.py index 8593bedfc0..c8902045bb 100644 --- a/vllm/model_executor/models/cohere_compass.py +++ b/vllm/model_executor/models/cohere_compass.py @@ -82,6 +82,7 @@ from vllm.multimodal.inputs import ( ImageItem, MultiModalFeatureSpec, MultiModalFieldConfig, + MultiModalKwargsItem, MultiModalKwargsItems, ) from vllm.multimodal.parse import ( @@ -2085,24 +2086,16 @@ class CohereCompassForConditionalGeneration( tower_model="visual.", ) - def get_num_mm_encoder_tokens( + def get_mm_lora_token_counts( self, - num_image_tokens: int, - ) -> int: - hf_config = self.config - vision_config = hf_config.vision_config - merge_size = vision_config.spatial_merge_size - - return num_image_tokens * merge_size**2 - - def get_num_mm_connector_tokens( - self, - num_vision_tokens: int, - ) -> int: - hf_config = self.config - vision_config = hf_config.vision_config - merge_size = vision_config.spatial_merge_size - return num_vision_tokens // merge_size**2 + *, + modality: str, + mm_kwargs: MultiModalKwargsItem | None, + num_mm_embeds: int, + ) -> tuple[int, int | None]: + del modality, mm_kwargs + merge_size = self.config.vision_config.spatial_merge_size + return num_mm_embeds * merge_size**2, num_mm_embeds @lru_cache diff --git a/vllm/model_executor/models/interfaces.py b/vllm/model_executor/models/interfaces.py index e28cc9b806..433c373c8e 100644 --- a/vllm/model_executor/models/interfaces.py +++ b/vllm/model_executor/models/interfaces.py @@ -31,7 +31,7 @@ import torch import torch.nn as nn from torch import Tensor from transformers.models.whisper.tokenization_whisper import LANGUAGES -from typing_extensions import Self, TypeIs, deprecated +from typing_extensions import Self, TypeIs from vllm.logger import init_logger from vllm.model_executor.layers.quantization import QuantizationConfig @@ -422,17 +422,6 @@ class SupportsMultiModal(SupportsMultiModalEmbeddings, Protocol): yield - @deprecated( - "get_num_mm_encoder_tokens is deprecated; use get_mm_lora_token_counts instead." - ) - def get_num_mm_encoder_tokens(self, num_image_tokens: int) -> int: ... - - @deprecated( - "get_num_mm_connector_tokens is deprecated; use " - "get_mm_lora_token_counts instead." - ) - def get_num_mm_connector_tokens(self, num_vision_tokens: int) -> int: ... - def get_mm_lora_token_counts( self, *, @@ -446,12 +435,9 @@ class SupportsMultiModal(SupportsMultiModalEmbeddings, Protocol): connector forwards. Models with multiple modalities can override this when each modality has different encoder padding or pooling behavior. """ - del modality, mm_kwargs - num_encoder_tokens = self.get_num_mm_encoder_tokens(num_mm_embeds) - num_connector_tokens = self.get_num_mm_connector_tokens(num_encoder_tokens) - return ( - num_encoder_tokens, - num_connector_tokens if isinstance(num_connector_tokens, int) else None, + raise NotImplementedError( + f"{type(self).__name__} must implement get_mm_lora_token_counts " + "to support tower/connector LoRA." ) def _embed_text_input_ids( diff --git a/vllm/platforms/interface.py b/vllm/platforms/interface.py index dc7178cf84..57a7a82f21 100644 --- a/vllm/platforms/interface.py +++ b/vllm/platforms/interface.py @@ -1111,11 +1111,6 @@ class Platform: """Checks if the platform allows inplace memory updates.""" return True - @classmethod - def get_lora_vocab_padding_size(cls) -> int: - """Returns how much padding the LoRA logits need for kernels.""" - return 256 - @classmethod def get_device_communicator_cls(cls) -> str: """Get device specific communicator class for distributed communication.""" diff --git a/vllm/v1/worker/gpu_model_runner.py b/vllm/v1/worker/gpu_model_runner.py index 047fdc5ff0..df7dc96516 100644 --- a/vllm/v1/worker/gpu_model_runner.py +++ b/vllm/v1/worker/gpu_model_runner.py @@ -3022,10 +3022,8 @@ class GPUModelRunner( ) self.lora_manager.set_active_adapters(lora_requests, tower_mapping) - # Only set connector mapping if the model actually has a connector. - # Some multimodal models inherit a stub `get_num_mm_connector_tokens` - # from `SupportsMultiModal`, which returns None and should not be - # treated as a signal that connector LoRA is supported. + # Only set connector mapping if the model actually has a connector + # and reports connector token counts. mm_mapping = ( self.model.get_mm_mapping() # type: ignore[attr-defined] if hasattr(self.model, "get_mm_mapping")