o iG@sddlmZmZddlmZddlmZmZddlm Z m Z ddl Z ddl mZddlmZddlmZmZmZmZddlmZd d lmZd d lmZd d lmZe dedZeeZ eGdddZ!Gdddee eZ"dS))ABCabstractmethod)Mapping) dataclassfield)GenericTypeVarN)Image)AudioDummyOptionsBaseDummyOptionsImageDummyOptionsVideoDummyOptions) init_logger)MultiModalDataDict)MultiModalDataItems)BaseProcessingInfo_I)boundc@s^eZdZUdZeeeBed<eed<e e dZ e ee fed<e e dZe ee fed<dS)ProcessorInputszq Represents the keyword arguments to [`vllm.multimodal.processing.BaseMultiModalProcessor.apply`][]. promptmm_items)default_factoryhf_processor_mm_kwargstokenization_kwargsN)__name__ __module__ __qualname____doc__strlistint__annotations__rrdictrrobjectrr&r&]/home/ubuntu/vllm_env/lib/python3.10/site-packages/vllm/multimodal/processing/dummy_inputs.pyrs rcs:eZdZdZdeddffdd Zedeee fdefdd Z e dd e deee fd eee fdBde fd d Z dd e deee fd eee fdBdefddZddde de dedBdeejfddZddde de de dedBdeejf ddZddde de de de dedBdeejf ddZZS) BaseDummyInputsBuilderz_ Abstract base class that constructs the dummy data to profile multi-modal models. inforeturnNcst||_dSN)super__init__r))selfr) __class__r&r'r-0s  zBaseDummyInputsBuilder.__init__ mm_countscCt)zD Build the text input corresponding to `mm_counts`. NotImplementedError)r.r1r&r&r'get_dummy_text5sz%BaseDummyInputsBuilder.get_dummy_textseq_len mm_optionscCr2)a Build the multimodal input which, after processing, results in the maximum possible number of placeholder tokens. Args: seq_len: Sequence length mm_counts: Count of items per modality mm_options: Configurable options per modality (optional). If None, use model defaults for backward compatibility. If provided, models can use these to customize dummy data generation. r3)r.r6r1r7r&r&r'get_dummy_mm_data<sz(BaseDummyInputsBuilder.get_dummy_mm_datacCs>||}||||}|jj|dd}ddi}t|||dS)a, Build the input which, after processing, results in the maximum possible number of placeholder tokens. Args: seq_len: Sequence length mm_counts: Count of items per modality mm_options: Configurable options per modality (optional) F)validate truncation)rrr)r5r8r) parse_mm_datar)r.r6r1r7 dummy_text dummy_mm_datadummy_mm_itemsrr&r&r'get_dummy_processor_inputsQs z1BaseDummyInputsBuilder.get_dummy_processor_inputs) overrideslength num_audiosr@cCsR|dkrgS|r|jr|j|krtd|j|t||j}t|f}|g|S)NrzOaudio.length override (%d) exceeds model's maximum length (%d), will be ignored)rAloggerwarningminnpzeros)r.rArBr@audior&r&r'_get_dummy_audiosls     z(BaseDummyInputsBuilder._get_dummy_audioswidthheight num_imagescCs|dkrgS|r4|jr|j|krtd|j|t||j}|jr4|j|kr.td|j|t||j}tjd||fdd}|g|S)NrzMimage.width override (%d) exceeds model's maximum width (%d), will be ignoredzOimage.height override (%d) exceeds model's maximum height (%d), will be ignoredRGB)color)rJrCrDrErKr new)r.rJrKrLr@imager&r&r'_get_dummy_imagess*     z(BaseDummyInputsBuilder._get_dummy_images num_frames num_videoscCs|dkrgS|rJ|jr|j|krtd|j|t||j}|jr4|j|kr.td|j|t||j}|jrJ|j|krDtd|j|t||j}tj|||dfdtjd}|g|S)Nrz]video.num_frames override (%d) exceeds model's maximum number of frames (%d), will be ignoredzMvideo.width override (%d) exceeds model's maximum width (%d), will be ignoredzOvideo.height override (%d) exceeds model's maximum height (%d), will be ignoredrN)dtype) rSrCrDrErJrKrFfulluint8)r.rJrKrSrTr@videor&r&r'_get_dummy_videoss:        z(BaseDummyInputsBuilder._get_dummy_videosr+)rrrrrr-rrr r"r5r rr8rr?r r!nptNDArrayrIr r rRr rZ __classcell__r&r&r/r'r(*s|       'r()#abcrrcollections.abcr dataclassesrrtypingrrnumpyrF numpy.typingr[PILr vllm.config.multimodalr r r r vllm.loggerrinputsrparsercontextrrrrCrr(r&r&r&r's"