o ۾i'( @sndZddlmZddlmZddlmZddlZddlZ ddl m Z m Z ddl mZddlmZmZmZdd lmZd d lmZeeZed eeBd efddZed eeBd efddZdZded efddZd eeBd eeeffddZ d eeBd efddZ!d ed edBfddZ"ded dfddZ#d ed ed efd!d"Z$ d(d#eeBd$ed%edBd efd&d'Z%dS))zGGUF utility functions.)cache)PathLike)PathN)KeysVisionProjectorType)GGMLQuantizationType) Gemma3ConfigPretrainedConfigSiglipVisionConfig) init_logger)list_filtered_repo_filesmodelreturnc Cst|}|s dS|jdkrdSz|d }|d}Wdn1s'wY|dkWStyJ}ztd||WYd}~dSd}~ww) z"Check if the file is a GGUF model.F.ggufTrbNsGGUFzError reading file %s: %s)ris_filesuffixopenread Exceptionloggerdebug)rfheadererV/home/ubuntu/.local/lib/python3.10/site-packages/vllm/transformers_utils/gguf_utils.pycheck_gguf_files    rcCs4d}t|}t||r|dd\}}t|SdS)z*Check if the model is a remote GGUF model.zG^[a-zA-Z0-9][a-zA-Z0-9._-]*/[a-zA-Z0-9][a-zA-Z0-9._-]*:[A-Za-z0-9_+-]+$:r F)strre fullmatchrsplitis_valid_gguf_quant_type)rpattern_ quant_typerrris_remote_gguf)s  r))_M_S_L_XL_XS_XXSgguf_quant_typecCsTtt|ddur dStD]}||r'|dt| }tt|ddur'dSq dS)zCheck if the quant type is a valid GGUF quant type. Supports both exact GGML quant types (e.g., Q4_K, IQ1_S) and extended naming conventions (e.g., Q4_K_M, Q3_K_S, Q5_K_L). NTF)getattrr_GGUF_QUANT_SUFFIXESendswithlen)r0r base_typerrrr%9s r%cCsHt|}t|r|dd}|d|dfStd|dtjdt)z,Split the model into repo_id and quant type.r r rz-Wrong GGUF model or invalid GGUF quant type: zI. - It should be in repo_id:quant_type format. - Valid base quant types: z% - Extended suffixes also supported: )r!r)r$ ValueErrorr_member_names_r2)rpartsrrrsplit_remote_ggufMs r9cCst|}t|r dSt|S)zCheck if the model is a GGUF model. Args: model: Model name, path, or Path object to check. Returns: True if the model is a GGUF model, False otherwise. T)r!rr))rrrris_gguf[s r:cCsx|dsdSz*t|}|sWdS|j}gd}|D]}t||}|r.|dWSqWdSty;YdSw)zCheck if GGUF model has multimodal projector file. Args: model: Model path string Returns: Path to mmproj file if found, None otherwise rN)z mmproj.ggufz mmproj-*.ggufz *mmproj*.ggufr)r3rrparentlistglobr)r model_path model_dirmmproj_patternsr& mmproj_filesrrrdetect_gguf_multimodalns"  rB mmproj_pathzSiglipVisionConfig | Nonec Csjtt|}d}|tjj}|r9z t|jd d}Wnt t fy8}z t d|WYd}~nd}~wwtjjdtftjjdtftjjdtftjjjdtftjjd tftjjd tftjjjd tfi}i}|D]!\}\}} ||} | durt d |dS| | jd||<qm|tjkrd |d<t dtdi|} |rt d|| St d| S)a[Extract vision config parameters from mmproj.gguf metadata. Reads vision encoder configuration from GGUF metadata fields using standardized GGUF constants. Automatically detects the projector type (e.g., gemma3, llama4) and applies model-specific parameters accordingly. The function extracts standard CLIP vision parameters from GGUF metadata and applies projector-type-specific customizations. For unknown projector types, it uses safe defaults from SiglipVisionConfig. Args: mmproj_path: Path to mmproj.gguf file (str or Path) Returns: SiglipVisionConfig if extraction succeeds, None if any required field is missing from the GGUF metadata Raises: Exception: Exceptions from GGUF reading (file not found, corrupted file, etc.) propagate directly from gguf.GGUFReader Nzutf-8z-Failed to decode projector type from GGUF: %s hidden_sizeintermediate_sizenum_hidden_layersnum_attention_heads image_size patch_sizelayer_norm_epsz8Missing required vision config field '%s' in mmproj.ggufFvision_use_headz8Detected Gemma3 projector, disabling vision pooling headz=Extracted vision config from mmproj.gguf (projector_type: %s)z1Extracted vision config from mmproj.gguf metadatar)gguf GGUFReaderr! get_fieldrClipPROJECTOR_TYPEbytesr8decodeAttributeErrorUnicodeDecodeErrorrwarning ClipVisionEMBEDDING_LENGTHintFEED_FORWARD_LENGTH BLOCK_COUNT Attention HEAD_COUNT IMAGE_SIZE PATCH_SIZE LAYERNORM_EPSfloatitemsrGEMMA3infor ) rCreaderprojector_typeprojector_type_fieldrVISION_CONFIG_FIELDS config_paramsgguf_key param_namedtypefieldconfigrrrextract_vision_config_from_ggufsP          ro hf_configcCsRt|}|dur'tt|}|}|jdv}|dur'|r't||dgd}|}|S)aCPatch HF config for GGUF models. Applies GGUF-specific patches to HuggingFace config: 1. For multimodal models: patches architecture and vision config 2. For all GGUF models: overrides vocab_size from embedding tensor This ensures compatibility with GGUF models that have extended vocabularies (e.g., Unsloth) where the GGUF file contains more tokens than the HuggingFace tokenizer config specifies. Args: model: Model path string hf_config: HuggingFace config to patch in-place Returns: Updated HuggingFace config N)gemma3 gemma3_textGemma3ForConditionalGeneration) text_config vision_config architectures)rBror!get_text_config model_typer)rrprCrurt is_gemma3 new_hf_configrrrmaybe_patch_hf_config_from_ggufs   r{repo_idr(revisioncCsvt|}d|dd|dd|dd|dg}t|||d}t|dkr-td|||jdd d |d}|S) asGet the GGUF file path from HuggingFace Hub based on repo_id and quant_type. Args: repo_id: The HuggingFace repository ID (e.g., "Qwen/Qwen3-0.6B") quant_type: The quantization type (e.g., "Q4_K_M", "F16") revision: Optional revision/branch name Returns: The path to the GGUF file on HuggingFace Hub (e.g., "filename.gguf"), z*-rz-*.ggufz*/*-)allow_patternsr}rz:Could not find GGUF file for repo %s with quantization %s.cSs|d|fS)N-)count)xrrr+sz,get_gguf_file_path_from_hf..)key)r!r r4r6sort)r|r(r} gguf_patternsmatching_files gguf_filenamerrrget_gguf_file_path_from_hfs(     r)N)&__doc__ functoolsrosrpathlibrrMregexr"gguf.constantsrr gguf.quantsr transformersrr r vllm.loggerr repo_utilsr __name__rr!boolrr)r2r%tupler9r:rBror{rrrrrsN       U +