vllm.model_executor.models.jina_vl
JinaVLForSequenceClassification
¶
Bases: Qwen2VLForConditionalGeneration
, SupportsCrossEncoding
, SupportsMultiModal
, SupportsScoreTemplate
Source code in vllm/model_executor/models/jina_vl.py
_pooler
instance-attribute
¶
_pooler = from_config_with_defaults(
pooler_config,
pooling_type=LAST,
normalize=False,
softmax=True,
)
weight_mapper
class-attribute
instance-attribute
¶
weight_mapper = WeightsMapper(
orig_to_new_prefix={
"score.0.": "score.dense.",
"score.2.": "score.out_proj.",
"model.language_model.": "language_model.model.",
"visual.": "visual.",
"lm_head.": "language_model.lm_head.",
"model.": "language_model.model.",
}
)
__init__
¶
__init__(*, vllm_config: VllmConfig, prefix: str = '')
Source code in vllm/model_executor/models/jina_vl.py
forward
¶
forward(
input_ids: Tensor,
positions: Tensor,
intermediate_tensors: Optional[
IntermediateTensors
] = None,
inputs_embeds: Optional[Tensor] = None,
**kwargs: object,
) -> Tensor
Source code in vllm/model_executor/models/jina_vl.py
get_placeholder_str
classmethod
¶
get_score_template
classmethod
¶
load_weights
¶
pooler
¶
pooler(
hidden_states: Tensor, pooling_metadata: PoolingMetadata
) -> Optional[PoolerOutput]
post_process_tokens
classmethod
¶
post_process_tokens(prompt: TokensPrompt) -> None
JinaVLMultiModalProcessor
¶
Bases: Qwen2VLMultiModalProcessor
Source code in vllm/model_executor/models/jina_vl.py
_call_hf_processor
¶
_call_hf_processor(
prompt: str,
mm_data: Mapping[str, object],
mm_kwargs: Mapping[str, object],
tok_kwargs: Mapping[str, object],
) -> BatchFeature
Source code in vllm/model_executor/models/jina_vl.py
JinaVLScorer
¶
Bases: Module