3j:6SSKrSSKJr SSKJrJr SSKJrJ r J r J r SSK J r Jr SSKJrJrJr SSKJr SS KJr \"5(aS S KJrJr \R4"\5r"S S \ SS9r\\"SS9"SS\ 555rS/rg)N) AudioInput) ImageInputmake_nested_list_of_images)MultiModalDataProcessingKwargsProcessorMixinUnpack)PreTokenizedInput TextInput)auto_docstringis_vision_availablelogging)requires) VideoInput)Gemma4ImageProcessorKwargs get_aspect_ratio_preserving_sizec<\rSrSr%\\S'SSS.SS00SS0S.rSrg ) Gemma4ProcessorKwargs" images_kwargsT)paddingreturn_mm_token_type_idsdo_convert_rgbreturn_metadata) text_kwargsr audio_kwargs videos_kwargsN)__name__ __module__ __qualname____firstlineno__r__annotations__ _defaults__static_attributes__r f/home/wildlama/miniconda3/lib/python3.13/site-packages/transformers/models/gemma4/processing_gemma4.pyrr"s5--(, d +T2 Ir(rF)total)vision)backendsc ^\rSrSr\rSS\S\S\4U4SjjjrSS\S-S\ \ -\ \ -\ \ -S \ S \ 4U4S jjjrSS\\ \-S-S\ \ -\ \ -\ \ -S \ S \ S \\4 U4S jjjrS\S\S\4SjrS\S\S\4SjrS\S\S\4SjrS SjrS\S\4Sjr\U4Sj5r\S\ \4Sj5rSrU=r$)!Gemma4Processor1Nimage_seq_lengthaudio_seq_lengthaudio_ms_per_tokenc >X`lURUlURUlURUlURUlUR SS/05 SUlURUR 5UlXpl Xl [USS5Ul [USS5Ul [USS5Ul[USS5Ul[ T U]D"S UUUUUS.U D6 g) u image_seq_length (`int`, *optional*, defaults to 280): The number of soft tokens per image used for placeholder expansion. audio_seq_length (`int`, *optional*, defaults to 750): The maximum number of audio soft tokens per audio segment. Serves as an upper-bound cap when dynamic audio token counts are computed. audio_ms_per_token (`int`, *optional*, defaults to 40): Milliseconds of audio per output soft token. Used to dynamically compute the number of audio placeholder tokens as ``ceil(duration_ms / audio_ms_per_token)``. The default of 40 comes from the SSCP convolution's 4× time reduction on 10ms frames. additional_special_tokensz <|video|>audio_token_idN audio_token boa_token eoa_token)feature_extractorimage_processor tokenizervideo_processor chat_templater )r0image_token_id boi_token eoi_token image_tokenadd_special_tokens video_tokenconvert_tokens_to_idsvideo_token_idr1r2getattrr5r6r7r8super__init__) selfr9r:r;r<r=r0r1r2kwargs __class__s r)rHGemma4Processor.__init__6s.!1'66",,",,$00 $$&AK=%QR&'==d>N>NO !1#5%i1A4H"9mTB K> K>  /++'    r(imagestextvideosaudioc 8>[TU]"SXX4S.UD6upp4Ub [U5nU(a?U(d8UVs/sH+nSRUR/[ U5-5PM- nnU(a U(dUR /[ U5-nXX44$s snf)N)rMrNrOrP r )rGprepare_inputs_layoutrjoinrAlenr6)rIrMrNrOrPrJ image_listrKs r)rS%Gemma4Processor.prepare_inputs_layoutns',g&C' V' DJ' #f  /7F $U[\U[zCHHd../#j/ABU[D\ $$%E 2DV** ]s2BrJc >[T U]"SXS.UD6 UcUc [S5eUb URbURb UR c [S5eUbUVs/sHofR UR5PM nnUb[U5[U5:wa$[S[U5S[U5S35eUVs/sHn[U5PM n nXy:wa,[SURSUS URS U S 3 5egUc6[U5(a%[S [U5S URS 35egggs snfs snf)N)rMrNz+You must provide either `text` or `images`.zUAudio inputs were provided, but the tokenizer does not have an `audio_token` defined.z1Received inconsistently sized batches of images (z ) and text (z).zThe total number of zP tokens in the prompts should be the same as the number of images passed. Found rRz tokens and z images per sample.zFound z. tokens in the text but no images were passed.r ) rGvalidate_inputs ValueErrorr6r7r8countrArUanysum) rIrMrNrOrPrJsamplen_images_in_textsublistn_images_in_imagesrKs r)rYGemma4Processor.validate_inputss CvCFC T R!v;#d)+$KCPVK=XdehimendooqrCI%I&wc'l&"%I#9$.t/?/?.@A""2!31T5E5E4FlSeRffy{: C(8$9$9 S!1231T5E5E4FFtu%: R&Js !$E E image_inputs image_idxreturncdUSUnURURU-UR3$)Nnum_soft_tokens_per_image)r?rAr@)rIrcrdnum_soft_tokenss r)replace_image_token#Gemma4Processor.replace_image_tokens;&'BCIN..!$"2"2_"D!EdnnEUVVr( video_inputs video_idxc USUnUSUnURc[RS5 URcSO URUlURVs/sH$n[ US-5SS[ US-5S3PM& nnSR UVs/sH.owSUR URU-UR3PM0 sn5nU$s snfs snf) Nnum_soft_tokens_per_videovideo_metadataaGemma4 requires frame timestamps to construct prompts, but the `fps` of the input video could not be inferred. Probably `video_metadata` was missing from inputs and you passed pre-sampled frames. Defaulting to `fps=24`. Please provide `video_metadata` for more accurate results.<02d:rR) fpslogger warning_once timestampsintrTr?rCr@) rIrkrlrhmetadataseconds timestamp_strtvideo_replacements r)replace_video_token#Gemma4Processor.replace_video_tokens&'BCIN 01)< <<    e  &\\1rx|| ]e\o\op\oQXC2 .s31S25Fs4KL\o pHHbo pbo]^s!DNN#D$4$4$F#GGW Xbo p ! q ps !+C5C  audio_inputs audio_idxcUSUn[U5n[S5H'nUS-S- S-S-nUSSS2SUn[U5nM) URUR[ UR 55-UR 3$)Ninput_features_maskrr)rUranger7r6rxr]r8)rIrrmaskr|_t_outs r)replace_audio_token#Gemma4Processor.replace_audio_tokens129= IqAUQY1$q(E!9Ve$DD A ..!$"2"2S_"D!EdnnEUVVr(c D[RRS05nURU5 URSS5=(d URR nURSS5=(d URR nURSS5=(d URRnXvS--n0n Ubd/n UH9n [U SU SUUUS 9upX-nX-nU RX-US--5 M; S/[U5-nU RU US .5 Ub`[URS S 5nUVs/sH)nUR[R"U5U5PM+ nnU RS U05 [!S0U D6$s snf)a Computes the number of placeholder tokens needed for multimodal inputs with the given sizes. Args: image_sizes (`list[list[int]]`, *optional*): The input sizes formatted as (height, width) per each image. audio_lengths (`list[int]`, *optional*): The lengths of audio inputs in number of samples. Used to dynamically compute per-audio token counts. Returns: `MultiModalData`: A `MultiModalData` object holding number of tokens per each of the provided input modalities, along with other useful data. r patch_sizeNpooling_kernel_sizemax_soft_tokensrrr)heightwidthr max_patchesr)num_image_tokensnum_image_patches sampling_ratei>num_audio_tokensr )rr&getupdater:rrrrappendrUrFr9_compute_audio_num_tokensnpzerosr)rI image_sizes audio_lengthsrJrrrrr vision_datar image_sizetarget_htarget_w patch_height patch_widthrrlengthrs r)_get_num_multimodal_tokens*Gemma4Processor._get_num_multimodal_tokenss .77;;ORP V$"&&|T:]d>R>R>]>]   3T : fd>R>R>f>f (++,=tDlH\H\HlHl%Q(>>   "! ) %E%a=$Q-) +(; &" (5 &4  '' (BFY[\F\(\]*"#c+&6 6    4D[lm n  $$D$:$:OVTM^k ^kTZ..rxx/?O^k      24DE F, ,,  s 0FrcN[U5nURRS-nURRS-nX5-U- URR-S-nUS::agSnSupn Un [ U5Hn U SU --U- U -S-n M [ XR 5$)aPNumber of audio soft tokens, replicating the encoder's seq-length arithmetic. Mirrors Gemma4AudioFeatureExtractor mel framing + the two stride-2 Conv2d subsampling layers in Gemma4AudioSubSampleConvProjection, capped at ``audio_seq_length``. Must match ``audio_mask.sum()`` from the audio tower or vLLM's ``_merge_multimodal_embeddings`` will raise on a length mismatch. Args: audio_waveform: A 1-D array or list containing the raw audio samples. sampling_rate: The sampling rate of the audio waveform in Hz. Returns: The number of audio soft tokens to insert as placeholders. rrr)rrr)rUr9 frame_length hop_lengthrminr1) rIaudio_waveformr num_samplesframe_size_for_unfoldpad_leftnum_mel_framessscp_num_layers sscp_kernel sscp_stride sscp_paddingr|rs r)r)Gemma4Processor._compute_audio_num_tokenss.) !% 6 6 C Ca G))66!;%03HHTMcMcMnMnnqrr Q 18. , 'AQ%% 3 CaGA( 1++,,r(c >[TU]S/-$)Nmm_token_type_ids)rGmodel_input_names)rIrKs r)r!Gemma4Processor.model_input_names,sw(,?+@@@r(c SS/$)Nrgrnr )rIs r)unused_input_names"Gemma4Processor.unused_input_names0s+-HIIr() r2r1r6r5r7r?r8r@r0rAr>rCrE)Nii()NNNN)NN)r!r"r#r$rvalid_processor_kwargsrxrHrr r listrrrSr rrYdictstrrir~rrrpropertyrrr' __classcell__)rKs@r)r.r.1s3 # #"$6 6 6  6 6 t%)Z^! +T!+++d9o=EV@WW+ +  ++48rs %AXXCAA*%e   H % ,E  ;JnJ JD  r(