o پi0@sddlmZmZddlmZzddlmZWn ey!edwddlmZGdddeZ Gdd d eZ Gd d d ed d Z GdddeZ e e e dS)) AutoProcessorPretrainedConfig)ProcessingKwargs)Qwen2_5_VLProcessorzOQwen2_5_VLProcessor can not be found. Please upgrade your transformers version.)DeepseekV2Configcs|eZdZUdZeed<           ddededededededededededeffdd Z Z S)DotsVisionConfigdots_vit model_type* h㈵>Fflash_attention_2{Gz?T embed_dim hidden_sizeintermediate_sizenum_hidden_layersnum_attention_heads num_channels patch_sizespatial_merge_sizetemporal_patch_size rms_norm_epsuse_biasc s|tjdi|||_||_||_||_||_||_||_||_ | |_ | |_ | |_ | |_ | |_||_||_||_||_dS)N)super__init__rrrrrrrrrrrattn_implementationinitializer_rangeinit_merger_std is_causal post_normgradient_checkpointing)selfrrrrrrrrrrrr#r$r%r&r'r(kwargs __class__r O/home/ubuntu/.local/lib/python3.10/site-packages/sglang/srt/configs/dots_vlm.pyr"s$ zDotsVisionConfig.__init__)r r r r r rrrrrFrrrFTF) __name__ __module__ __qualname__r str__annotations__intfloatboolr" __classcell__r r r+r-rsT     rcs eZdZdZfddZZS) DotsVLMConfigdots_vlmc sftjdi||di}|dd|_|dd|_tdi||_tdi||_dg|_ dS)N vision_configimage_token_idi/video_token_idiDDotsVLMForCausalLMr ) r!r"get im_span_id video_span_idrr9rlanguage_config architectures)r)r*r9r+r r-r"=s  zDotsVLMConfig.__init__)r.r/r0r r"r6r r r+r-r7:sr7c@seZdZdddiiZdS)DotsVLMProcessorKwargs text_kwargspaddingFN)r.r/r0 _defaultsr r r r-rBGs rBF)totalcs6eZdZdZddgZdgZdZ d fdd ZZS) DotsVLMProcessora! Constructs a DotsVLM processor which derives from Qwen2_5_VLProcessor, but overrides the image and video token ids. Besides, its tokenizer is a LlamaTokenizerFast instead of Qwen2TokenizerFast. [`DotsVLMProcessor`] offers all the functionalities of [`DotsVisionConfig`] and [`LlamaTokenizerFast`]. See the [`~DotsVLMProcessor.__call__`] and [`~DotsVLMProcessor.decode`] for more information. Args: image_processor ([`Qwen2VLImageProcessor`], *optional*): The image processor is a required input. tokenizer ([`LlamaTokenizerFast`], *optional*): The tokenizer is a required input. chat_template (`str`, *optional*): A Jinja template which will be used to convert lists of messages in a chat into a tokenizable string. image_processor tokenizer chat_template)LlamaTokenizerLlamaTokenizerFastNc stj|||dt|dsdn|j|_t|dsdn|j|_t|ds&dn|j|_t|ds1d n|j|_t|d dr>|jn| |jd |_t|d drS|j |_ dS| |jd |_ dS) N)rJ image_tokenz <|imgpad|> video_tokenz <|video_pad|> img_tokenz<|img|>endofimg_tokenz <|endofimg|>r:rr;) r!r"hasattrrMrNrOrPgetattrr:encoder;)r)rHrIrJr*r+r r-r"ds0  zDotsVLMProcessor.__init__)NNN) r.r/r0__doc__ attributes valid_kwargstokenizer_classr"r6r r r+r-rGOsrGN) transformersrrtransformers.processing_utilsrr ImportErrorsglang.srt.configs.deepseekvl2rrr7rBrGregisterr r r r-s   , 7