j7% %SrSSKrSSKrSSKJr SSKJrJrJ r J r J r J r SSK rSSKJr SSKJr SSKJr SSKJr SS KJr \R2"\5r\R8S \R:S \R<S 0r\ \\ 4\!S 'S\S\ \ 4Sjr"S\S\ \ SS4Sjr#S\S\ \ S\$4Sjr%S\ \R$\RL4S\R$4Sjr'S\ \ \R$\RL4S\ \R$4Sjr(S\ \ \4S\ \4Sjr)S\ S\ S\ \ \\ 4S\\ \ /\4S\4 Sjr*S#S\ S \S!\S\ \ S\ \ 4 S"jjr+g)$zInternal utilities for VLM runtimes. This module contains shared utility functions used across different VLM runtime implementations to avoid code duplication and ensure consistency. N)Path)AnyCallableDictListOptionalUnion)version)Image)TransformersPromptStyle) VlmEngineType)GenerationStopperzmlx-vlm transformersvllm_ENGINE_PACKAGES engine_typereturnc[RU5nUcg[RR U5$![RR a gf=f)zInstalled version of an engine's backing library. Returns None when the engine has no known backing package or the package is not installed, in which case the version requirement cannot be evaluated. N)rget importlibmetadatar PackageNotFoundError)rpackages h/home/mande/repo/quber/.venv/lib/python3.13/site-packages/docling/models/inference_engines/vlm/_utils.py_installed_engine_versionr"sV "";/G!!))'22    2 2s:AA min_versionc Ucg[U5nUcg[R"U5[R"U5:a![Un[ SUSUSUSUS3 5eg)aRaise when the engine's backing library is older than the model requires. Why: an outdated backing library typically fails deep inside the vendor code (e.g. a TypeError in mlx-vlm's Qwen3-VL vision tower) rather than at load time, which is hard to act on. NzThis model requires z>=z, but z is installed. Upgrade z, or select a different engine.)rr parser RuntimeErrorrr installedrs rcheck_min_engine_versionr"1s)+6I }}Y'-- "<<";/"7)2k]& L"")*I K  =c Ucg[U5nUcg[R"U5[R"U5:ag[Un[R SUR UUUUU5 g)zNon-raising variant of :func:`check_min_engine_version` for auto-selection. Logs why an engine was skipped so the fallback is not silent. Tzb%s engine not selected: %s %s is installed but this model requires >=%s. Install %s>=%s to use it.F)rr rr_logwarningvaluer s rengine_version_satisfiedr(Ksu)+6I}}Y7==#=={+GLL *  r#imagec[U[R5(aURS:XaGURSS;a4[ R "UR[R55nO[URS:Xa3[ R "UR[R5SS9nO[SUR35eURS:waURS5nU$)zConvert any image format to RGB PIL Image. Args: image: Input image as PIL Image or numpy array Returns: RGB PIL Image Raises: ValueError: If numpy array has unsupported shape )r+L)modezUnsupported numpy array shape: RGB) isinstancenpndarrayndimshaper fromarrayastypeuint8 ValueErrorr/convert)r)s rnormalize_image_to_pilr;js%$$ ::?u{{1~7OOELL$:;E ZZ1_OOELL$:EE>u{{mLM M zzU e$ Lr#imagescDUVs/sHn[U5PM sn$s snf)zPreprocess a batch of images to RGB PIL Images. Args: images: List of images as PIL Images or numpy arrays Returns: List of RGB PIL Images )r;)r<imgs rpreprocess_image_batchr?s"4: :6C "3 '6 :: :s extra_configc/nURS/5nUHon[U[5(aURU5 M+[U[5(dMB[ U[5(dMYURU"55 Mq U$)aExtract and instantiate GenerationStopper instances from config. This handles both GenerationStopper instances and classes, instantiating classes as needed. Args: extra_config: Extra generation configuration dictionary Returns: List of GenerationStopper instances custom_stopping_criteria)rr1rappendtype issubclass)r@stopperscustom_criteriacriterias rextract_generation_stoppersrIso)+H"&&'A2FO# h 1 2 2 OOH % $ ' 'JxAR,S,S OOHJ ' $ Or#repo_idrevisionartifacts_path download_fnc URSS5nUcUO [U5nUcU"X5$X$- R5(aX$- $/nUR5(a@UR5Vs/sH%ofR 5(dMUR PM' nn[ SUSX$- SUSU(aSRU5OSS US 3 5es snf) aDResolve the path to model artifacts, downloading if needed. This standardizes the logic for finding or downloading model artifacts across different runtimes. Args: repo_id: HuggingFace repository ID (e.g., "microsoft/Phi-3.5-vision-instruct") revision: Model revision (e.g., "main") artifacts_path: Optional path to cached artifacts directory download_fn: Function to download models, takes (repo_id, revision) and returns Path Returns: Path to the model artifacts directory Raises: FileNotFoundError: If artifacts_path is provided but model not found /z--zModel 'z2' not found in artifacts_path. Expected location: z Available models in : , nonezT To fix this issue: 1. Download the model: docling-tools models download-hf-repo zy 2. Or remove --artifacts-path to enable auto-download 3. Or use a different model that exists in your artifacts_path)replacerexistsiterdiris_dirnameFileNotFoundErrorjoin)rJrKrLrMrepo_cache_folderavailable_modelsps rresolve_model_artifacts_pathr]s. T2'5'=^4CWN7--  , 4 4 6 611  " "0>0F0F0HW0H1HHJ0H WgY""0"D!EF##1"2".>tyy)*FKLNNUYWO  P   Xs 2CCprompt processor prompt_stylec U[R:XaU$U[R:XagUS:XaC[R S5 SnSnSnUSUUU3n[R SUS U35 U$U[R :XaS S S 0S US./S./nUR USS9$[SUSSRS[5535e)aFormat a prompt according to the specified style. This centralizes prompt formatting logic that was previously duplicated across different model implementations. Args: prompt: User prompt text processor: Model processor with apply_chat_template method prompt_style: Style of prompt formatting to use repo_id: Optional model repository ID for model-specific formatting Returns: Formatted prompt string, or None if prompt_style is NONE Nz#microsoft/Phi-4-multimodal-instructz"Using specialized prompt for Phi-4z<|user|>z <|assistant|>z<|end|>z <|image_1|>zFormatted prompt for rPuserrDr)text)rDrc)rolecontentT)add_generation_promptzUnknown prompt style: z. Valid values are rQc38# UHoRv M g7fN)r').0ss r (format_prompt_for_vlm..s)S;Ra'';Rs) r RAWNONEr%debugCHATapply_chat_templater9rY) r^r_r`rJuser_prompt_prefixassistant_prompt prompt_suffix formattedmessagess rformat_prompt_for_vlmrws(.222 055 5 9 9 78'*! !"+fXm_EUDV W  *7)2i[AB 055 5W%#V4  ,,XT,RR$\N3 $ )S;R)S ST V  r#rh),__doc__importlib.metadatarloggingpathlibrtypingrrrrrr numpyr2 packagingr PILr ,docling.datamodel.pipeline_options_vlm_modelr )docling.models.inference_engines.vlm.baser %docling.models.utils.generation_utilsr getLogger__name__r%MLX TRANSFORMERSVLLMrstr__annotations__rr"boolr(r3r;r?rIr]rwr#rrs ==PCC" y.$}c)* = Xc]   -5c]   4-5c] >% RZZ(?"@U[[< ; u{{BJJ./ 0 ; %++ ;sCx. :0 0 0 U49-.0 3*d*+ 0  0 n" 3 3 3 *3 c] 3 c] 3 r#