j,:SSKJr SSKJrJrJrJrJrJr SSK J r SSK J r J r JrJr SSKJr SSKJr SS KJr SS KJr \(a SSK J r SS KJr "S S \ 5r"SS\\5r"SS\\5r"SS\\5r"SS\\5r "SS\5r!\"S5"SS\!55r""SS\5r#g!\a "SS5rNf=f))Enum) TYPE_CHECKING AnnotatedAnyLiteralOptionalUnion) SegmentedPage)AnyUrl BaseModel ConfigDictField) deprecated)StoppingCriteriac\rSrSrSrg)rN__name__ __module__ __qualname____firstlineno____static_attributes__ri/home/mande/repo/quber/.venv/lib/python3.13/site-packages/docling/datamodel/pipeline_options_vlm_model.pyrrs rr)AcceleratorDevice)GenerationStopper)Pagec\rSrSr%Sr\\\"SS94\S'\\\"SS94\S'Sr \\ \"S S94\S 'S r \\ \ \"S S94\S 'Sr\\ \"SS94\S'S S.S\ SS\ SS\4SjjrS\S\4SjrSrg )BaseVlmOptionsz.Base configuration for vision-language models.zbType identifier for the VLM options. Used for discriminating between different VLM configurations. descriptionkindzbPrompt template for the vision-language model. Guides the model's output format and content focus.promptg@zScaling factor for image resolution before processing. Higher values provide more detail but increase processing time and memory usage. Range: 0.5-4.0 typical.scaleNzMaximum image dimension (width or height) in pixels. Images larger than this are resized while maintaining aspect ratio. If None, no size limit is enforced.max_sizegzSampling temperature for text generation. 0.0 uses greedy decoding (deterministic), higher values (e.g., 0.7-1.0) increase randomness. Recommended: 0.0 for consistent outputs. temperature)_internal_pagepager r)rreturncUR$)aBuild the prompt for VLM inference. Args: page: The parsed/segmented page to process. _internal_page: Internal parameter for experimental layout-aware pipelines. Do not rely on this in user code - subject to change. Returns: The formatted prompt string. )r%)selfr*r)s r build_promptBaseVlmOptions.build_promptRs {{rtextcU$)Nr)r-r0s rdecode_responseBaseVlmOptions.decode_responseds rr)rrrr__doc__rstrr__annotations__r&floatr'rintr(r.r2rrrrr r s8   8      ;   $  9  8     &  i  6     &   P     ,0 '!(  $CCrr c<\rSrSrSrSrSrSrSrSr Sr S r S r S r S rg )ResponseFormathdoctagsdoclangmarkdowndeepseekocr_markdownunlimited_ocr_markdownhtmlotsl plaintext chandra_html dots_jsonrN)rrrrDOCTAGSDOCLANGMARKDOWNDEEPSEEKOCR_MARKDOWNUNLIMITED_OCR_MARKDOWNHTMLOTSL PLAINTEXT CHANDRA_HTML DOTS_JSONrrrrr:r:hs6GGH15 D DI!LIrr:c \rSrSrSrSrSrSrg)InferenceFrameworkumlx transformersvllmrN)rrrrMLX TRANSFORMERSVLLMrrrrrQrQus C!L DrrQc \rSrSrSrSrSrSrg)TransformersModelType{ automodelzautomodel-causallmzautomodel-imagetexttotextrN)rrrr AUTOMODELAUTOMODEL_CAUSALLMAUTOMODEL_IMAGETEXTTOTEXTrrrrrZrZ{sI- ;rrZc \rSrSrSrSrSrSrg)TransformersPromptStylechatrawnonerN)rrrrCHATRAWNONErrrrraras D C Drrac\rSrSr%Sr\"SS9rSr\S\ S'\ \ \ "SSS /S 94\ S 'S r \ \ \ "S S S/S 94\ S'Sr\ \\ "SS94\ S'Sr\ \\ "SS94\ S'Sr\ \\ "SS94\ S'Sr\ \\ "SS94\ S'\ \\ "SS94\ S'\R,r\ \\ "SS94\ S'\R2r\ \\ "SS94\ S '\ \\ "S!S94\ S"'S#r\ \\ \ "S$S94\ S%'\R>\R@\RB\RD/r#\ \$\\ "S&S94\ S''/r%\ \$\ \ "S(S94\ S)'/r&\ \$\'\(\)4\ "S*S94\ S+'0r*\ \+\ \,4\ "S,S94\ S-'0r-\ \+\ \,4\ "S.S94\ S/'Sr.\ \\ "S0S94\ S1'S2r/\ \0\ "S3S94\ S4'Sr1\ \\ "S5S94\ S6'Sr2\ \\ "S7S94\ S8'\3S9\ 4S:j5r4S;r5g#)<InlineVlmOptionsz@Configuration for inline vision-language models running locally.Tarbitrary_types_allowedinline_model_optionsr$zHuggingFace model repository ID for the vision-language model. Must be a model capable of processing images and generating text.zQwen/Qwen2-VL-2B-Instructz!ibm-granite/granite-vision-3.3-2br#examplesrepo_idmainzGit revision (branch, tag, or commit hash) of the model repository. Allows pinning to specific model versions for reproducibility.zv1.0.0revisionFzAllow execution of custom code from the model repository. Required for some models with custom architectures. Enable only for trusted sources due to security implications.r"trust_remote_codezLoad model weights in 8-bit precision using bitsandbytes quantization. Reduces memory usage by ~50% with minimal accuracy loss. Requires bitsandbytes library and CUDA. load_in_8bitg@zThreshold for LLM.int8() quantization outlier detection. Values with magnitude above this threshold are kept in float16 for accuracy. Lower values increase quantization but may reduce quality.llm_int8_thresholdzIndicates if the model is pre-quantized (e.g., GGUF, AWQ). When True, skips runtime quantization. Use for models already quantized during training or conversion. quantizedzInference framework for running the VLM. Options: `transformers` (HuggingFace), `mlx` (Apple Silicon), `vllm` (high-throughput serving).inference_frameworkzHuggingFace Transformers model class to use. Options: `automodel` (auto-detect), `automodel-vision2seq` (vision-to-sequence), `automodel-causallm` (causal LM), `automodel-imagetexttotext` (image+text to text).transformers_model_typezPrompt formatting style for Transformers models. Options: `chat` (chat template), `raw` (raw text), `none` (no formatting). Use `chat` for instruction-tuned models.transformers_prompt_stylezExpected output format from the VLM. Options: `doctags` (structured tags), `doclang` (Doclang XML), `markdown`, `html`, `otsl` (table structure), `plaintext`. Guides model output parsing.response_formatNzPyTorch data type for model weights. Options: `float32`, `float16`, `bfloat16`. Lower precision reduces memory and increases speed. If None, uses model default. torch_dtypezBList of hardware accelerators supported by this VLM configuration.supported_deviceszList of strings that trigger generation stopping when encountered. Used to prevent the model from generating beyond desired output boundaries. stop_stringszCustom stopping criteria objects for fine-grained control over generation termination. Allows implementing complex stopping logic beyond simple string matching.custom_stopping_criteriazAdditional generation configuration parameters passed to the model. Overrides or extends default generation settings (e.g., top_p, top_k, repetition_penalty).extra_generation_configzAdditional keyword arguments passed to the image processor. Used for model-specific preprocessing options not covered by standard parameters.extra_processor_kwargszEnable key-value caching for transformer attention. Significantly speeds up generation by caching attention computations. Disable only for debugging or memory-constrained scenarios. use_kv_cacheizMaximum number of tokens to generate. Limits output length to prevent runaway generation. Adjust based on expected output size and memory constraints.max_new_tokenszTrack and store generated tokens during inference. Useful for debugging, analysis, or implementing custom post-processing. Increases memory usage.track_generated_tokenszTrack and store the input prompt sent to the model. Useful for debugging, logging, or auditing. May contain sensitive information.track_input_promptr+c:URRSS5$)N/z--)rqreplace)r-s rrepo_cache_folder"InlineVlmOptions.repo_cache_folder^s||##C..rr)6rrrrr4r model_configr$rr6rr5rrsrtboolrurvr7rwrQrZr]ryrarfrzr:r|rrCPUCUDAMPSXPUr}listr~rr rrrdictrrrrr8rrpropertyrrrrrrjrjsJd;L,BD'( )B   #23VW     *  i  #h'     ( y  I    & )  I     (    &     ( y  ;    # -    ( ''Y D    (( $$y H    % "    (   @     $   y  T   , ) S  -     & i U#%66 78 @     & Y S#X 5     & I S#X '     ( )  0     ( I  /     & I  *    &        /3//rrjzUse InlineVlmOptions instead.c\rSrSrSrg)HuggingFaceVlmOptionsicrNrrrrrrcsrrc\rSrSr%Sr\"SS9rSr\S\ S'\ "S5r \ \ \ "SS 94\ S '0r\ \\\4\ "S S S 0/S94\ S'0r\ \\\4\ "SS 94\ S'Sr\ \\ "SS 94\ S'Sr\ \\ "SS 94\ S'\ \\ "SS 94\ S'/r\ \\\ "SS 94\ S'/r\ \\\ "SS 94\ S'Sr\ \\ "SS 94\ S 'S!rg")# ApiVlmOptionsihz;Configuration for API-based vision-language model services.Trlapi_model_optionsr$z*http://localhost:11434/v1/chat/completionszAPI endpoint URL for VLM service. Must be OpenAI-compatible chat completions endpoint. Default points to local Ollama server; update for cloud services or custom deployments.r"urlzoHTTP headers to include in API requests. Use for authentication or custom headers required by your API service. Authorizationz Bearer TOKENroheaderszAdditional query parameters to include in API requests. Service-specific parameters for customizing API behavior beyond standard options.paramsgN@zMaximum time in seconds to wait for API response before timing out. Increase for slow networks or complex vision tasks. Recommended: 30-120 seconds.timeoutzNumber of concurrent API requests allowed. Higher values improve throughput but may hit API rate limits. Adjust based on API service quotas and network capacity. concurrencyzExpected output format from the VLM API. Options: `doctags` (structured tags), `doclang` (Doclang XML), `markdown`, `html`, `otsl` (table structure), `plaintext`. Guides response parsing.r{zList of strings that trigger generation stopping when encountered. Sent to API to prevent the model from generating beyond desired output boundaries.r~zCustom stopping criteria objects for client-side generation control. Applied after receiving API responses for additional filtering or termination logic.rFzTrack and store the input prompt sent to the API. Useful for debugging, logging, or auditing. May contain sensitive information.rrN)rrrrr4r rr$rr6r rrrrrr5rrrr7rr8r:r~rrrrrrrrrrrhsEd;L)      $    ( ) S  4     & i  2     &        rrN)$enumrtypingrrrrrr docling_core.types.doc.pager pydanticr r r rtyping_extensionsrrTr ImportError%docling.datamodel.accelerator_optionsr%docling.models.utils.generation_utilsrdocling.datamodel.base_modelsrr r5r:rQrZrarjrrrrrrsJJ599( -DC92GYGT S$ d