+ UV-j^RIHt^RIHt^RIHt^RIHt^RI H t ^RI H t ^RI Ht^RIHt^RIHt]P&!R ]4!R R ]P.4tR# ]dt]!R ] 24RtAL.RtAii;i) )OptionalN)InputEmbeddingsFeatures ModelConfig) LanguageModel)Glm46VProcessor) VisionModel) AutoProcessorrz#Error registering glm4v processor: caa]tRt^toV3RlV3RlltR V3RlRllt]R4t]R4t R V3RlRllt R t R t Vt V;t#) Modelc <V^8dQhRS[/#)configr)format __classdict__s"k/Volumes/fast/ai/experiments/ui-tars-smoke/.venv/lib/python3.14/site-packages/mlx_vlm/models/glm4v/glm4v.py __annotate__Model.__annotate__sHH{Hc<\SV`4Wn\VP4Vn\ VPV4VnR#N) super__init__rr vision_config vision_towerr text_configlanguage_model)selfr __class__s&&rrModel.__init__s;  '(<(<=+F,>,>Grcn<V^8dQhRS[S[P,RS[S[P,/#)r input_ids pixel_values)rmxarray)rrs"rrrs5<J<JBHH%<Jrxx(<Jrc VPRR4pVPRR4pVPRR4pVeTMTpVfARVPn\ VPP P V4R7#VPPPPPpVPV4pVPP P V4p VPRR4p V eT p MVPW'RR7p VPR 4VPP^,,P4p \ P"!W^,\%V R,4.^R 7p \ P&!V ^R 7PV ^,P4p VP)VP*P,VP*P.V V V4p VfVe?VPP1WWV4wrWPnWPn\ V R7#) image_grid_thwNvideo_grid_thwmask) inputs_embedscached_image_featuresF)output_hidden_states:NrNaxis)popgetr _position_idsrmodel embed_tokensr patch_embedprojweightdtypeastypeprodspatial_merge_sizetolistr$splitsum concatenate#merge_input_ids_with_image_featuresrimage_token_idvideo_token_idget_rope_index _rope_deltas)rr"r#kwargsr'r(r)grid_thwr8r*cached hidden_states split_sizesfinal_inputs_embeds position_ids rope_deltass&&&, rget_input_embeddingsModel.get_input_embeddingss  $4d;$4d;zz&$'%3%?>^  04D   -*"1177DDYO !!--2299??#**51 ++11>>yI 3T:  "M!--U.M ##B'4+<+<+O+OQR+RRfh HHAKO0DEAMNN=qAHHa &&M #FF KK & & KK & &       %)C(,(;(;(J(J>) %L1=   -/:   ,&5HIIrc  W@8Hp\P!V4^8XdWA8HpVPwrg.p^p \V4EF0p WZ,p \P!V 4P 4p V ^8dVWV ,p V P^,V 8wd&\ RV RV P^, RV 24h\P !V P\P44p\P!W^, ^4pW,p\P!V RR7p\P!VVW:,4pW, p MW:,pVPV4EK3 \P!V^R7#)a,Merge image features into input embeddings at image token positions. Args: image_token_id: The token ID for image placeholders video_token_id: The token ID for video placeholders (fallback) image_features: Vision features from the vision tower [num_features, hidden_dim] inputs_embeds: Input embeddings [batch_size, seq_len, hidden_dim] input_ids: Input token IDs [batch_size, seq_len] grid_thw: Grid dimensions for each image (optional, not used in simple case) Returns: Updated input embeddings with image features inserted z!Number of image token positions (z+) does not match number of image features (z ) for batch r-r/) r$r>shaperangeitem ValueErrorcumsumr9int32where expand_dimsappendstack)rArBimage_featuresr*r"image_positions batch_sizeseq_len batch_outputsfeature_start_idx batch_idx image_mask num_positionsbatch_featuresrTfeature_indicesgathered_featuresimage_mask_expanded batch_outputs&&&&& rr@)Model.merge_input_ids_with_image_features[sx,$5 66/ "a ''9O(oo  z*I(3JFF:.335Mq !/%M(I" "''*m;$;M?K55C5I5I!5L4M\ZcYdf :#4#4RXX#>?"$((:z1"E%3$C!')nnZb&I#!xx'):M.transform_keysu3!,++h3;;HnUC%,kk"8:PQC7G(H(Hkk)-EFJr)items)rweightsr{kvs&& rsanitizeModel.sanitizes3 18 @ a !#@@@s/)rrr)NN)NNN)__name__ __module__ __qualname____firstlineno__rrM staticmethodr@propertyrjrqr__static_attributes____classdictcell__ __classcell__)rrs@@rr r s_HH <J<J|G/G/R00. A Arr )typingrmlx.corecorer$mlx.nnnnbaserrrlanguager processingrvisionr transformersr register ExceptioneprintModuler rrrsq*#'5*,o> uABIIuA 5 /s 3445sA""A>) A99A>