0i(UddlZddlmZddlZddlmZddlmZddlmZddl m Z ddl m Z dd l mZmZmZdd lmZmZdd lmZdd lmZmZmZmZdd lmZddlmZddlm Z m!Z!ddl"m#Z#m$Z$ejJe&Z' d+dejPdejRdejRdejRdejRdzde*dzde*fdZ+GddejPZ,Gdde Z-eGdd eZ.ed!"Gd#d$e.Z/Gd%d&ejPZ0ed'"Gd(d)e.e Z1gd*Z2y),N)Callable)nn)ACT2FN)Cache)GenerationMixin)GradientCheckpointingLayer)BaseModelOutputWithPastBaseModelOutputWithPoolingCausalLMOutputWithPast)ALL_ATTENTION_FUNCTIONSPreTrainedModel)Unpack)TransformersKwargsauto_docstringcan_return_tuplelogging)merge_with_config_defaults)capture_outputs) AutoModelAutoModelForCausalLM) VoxtralConfigVoxtralEncoderConfigmodulequerykeyvalueattention_maskscalingdropoutc ||jddz}tj||jdd|z}|||z}tj j |d}tj j|||j}tj||} | jddj} | |fS)Nrr)dimptrainingr) sizetorchmatmul transposer functionalsoftmaxr"r) contiguous) rrrrr r!r"kwargs attn_weights attn_outputs h/mnt/e/genesis-system/.venv/lib/python3.12/site-packages/transformers/models/voxtral/modeling_voxtral.pyeager_attention_forwardr5-s**R.D(<<s}}Q':;gEL!#n4 ==((2(>L==((6??([L,,|U3K''1-88:K  $$c,eZdZdZ ddedededededed edzd edzffd Zd e jd edefdZ dde jde jdzdede e je jdze e jdzffdZ xZS)VoxtralAttentionz=Multi-headed attention from 'Attention Is All You Need' paperN embed_dim num_headsr" is_decoderbias is_causal layer_idxconfigc zt |||_||_||_||z|_||_|j |z|jk7rtd|jd|d|j dz|_||_ ||_ |/|r-tjd|jjd||_t!j"||d|_t!j"||||_t!j"||||_t!j"||||_y) Nz;embed_dim must be divisible by num_heads (got `embed_dim`: z and `num_heads`: z).r%zInstantiating a decoder z without passing `layer_idx` is not recommended and will to errors during the forward call, if caching is used. Please make sure to provide a `layer_idx` when creating this class.Fr<)super__init__r9r:r"head_dimr? ValueErrorr!r;r=logger warning_once __class____name__r>rLineark_projv_projq_projout_proj) selfr9r:r"r;r<r=r>r?rHs r4rCzVoxtralAttention.__init__Js  "" !Y.  MMI %$.. 8MdnnM]$YKr3 }}d* $"     *4>>+B+B*CD,,  #ii 95A ii 94@ ii 94@  )YTB r6tensorseq_lenbszc|j|||j|jjddj S)Nrr)viewr:rDr-r0)rOrPrQrRs r4_shapezVoxtralAttention._shapers7{{3GQQRSUVWbbddr6 hidden_statesr output_attentionsreturnc 0|j\}}}|j|j||jz||}|j|j |d|} |j|j |d|} t j|jjt} | ||| | |f|jsdn |jd|d|\} } | j||dj} |j| } | | fS)z#Input shape: Batch x Time x Channelr$?)r"r!rW)r*rUrMr!rKrLr get_interfacer?_attn_implementationr5r)r"reshaper0rN)rOrVr rWr1rRtgt_len_ query_states key_states value_statesattention_interfacer3r2s r4forwardzVoxtralAttention.forwardus(,,.Wa{{4;;}#= #LgWZ[ [[]!;RE {{4;;}#=r3G (?(M(M KK , ,.E) %8      % $}}C$,,/ %  % ! \"))#w;FFH mmK0 L((r6)rZFTFNNNF)rI __module__ __qualname____doc__intfloatboolrrCr+TensorrUtuplere __classcell__rHs@r4r8r8Gs G   $'+&C&C&C &C  &C  &C&C:&C$&CPeU\\eCece /3"' ')||') t+') ') u||U\\D0% 2E2LL M ')r6r8c |eZdZdeffd Z ddej dej dedej fdZxZ S) VoxtralEncoderLayerr?cht||j|_t |j|j |j ||_tj|j|_ |j|_ t|j|_|j|_tj |j|j"|_tj |j"|j|_tj|j|_y)N)r9r:r"r?)rBrCd_modelr9r8encoder_attention_headsattention_dropout self_attnr LayerNormself_attn_layer_normr"ractivation_function activation_fnactivation_dropoutrJencoder_ffn_dimfc1fc2final_layer_normrOr?rHs r4rCzVoxtralEncoderLayer.__init__s )nn44,,   %'LL$@!~~ #F$>$>?"(";";99T^^V-C-CD99V33T^^D " T^^ <r6rVr rWrXc|}|j|}|j|||\}}tjj ||j|j }||z}|}|j |}|j|j|}tjj ||j|j }|j|}tjj ||j|j }||z}|jtjk(rEtj|jjdz }tj || |}||fS)a Args: hidden_states (`torch.FloatTensor`): input to the layer of shape `(batch, seq_len, embed_dim)` attention_mask (`torch.FloatTensor`): attention mask of size `(batch, 1, tgt_len, src_len)` where padding elements are indicated by very large negative values. output_attentions (`bool`, *optional*): Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned tensors for more detail. )rVr rWr'i)minmax)ryrwrr.r"r)rr{r~r|rdtyper+float16finforclamp)rOrVr rWresidualr2 clamp_values r4rezVoxtralEncoderLayer.forwardsP!11-@ &*nn')/'5' # |  --mt||VZVcVc-d  =0  --m< **488M+BC  --mt?V?Vaeanan-o /  --mt||VZVcVc-d  =0   %-- /++m&9&9:>>EK!KK K<[YMl**r6)F) rIrgrhrrCr+rmrlrerorps@r4rrrrsK=}=,#( %+||%+ %+ %+  %+r6rrcDeZdZUeed<dZdZdZdZdZ dZ dZ dZ dZ dZdZy)VoxtralPreTrainedModelr?model)audiotextTNpast_key_values)rIrgrhr__annotations__base_model_prefixinput_modalitiessupports_gradient_checkpointing_no_split_modules_skip_keys_device_placement_supports_flash_attn_supports_sdpa_supports_flex_attn_supports_cache_class_supports_attention_backend_can_compile_fullgraphr6r4rrsI (&*#"3N "&!r6rz: The Voxtral encoder, which is a Whisper encoder.  custom_introceZdZUdZeed<dZdZdgZe e dZ deffd Z dZ d ejfd Zd ejfd Zee dd eed eezfdZdej4fdZxZS)VoxtralEncoderz Transformer encoder consisting of *config.encoder_layers* self attention layers. Each layer is a [`VoxtralEncoderLayer`]. Args: config: VoxtralEncoderConfig r?input_featuresrrr) attentionsrVcbt|||j|_|j|_|j }|j |_|j|_|jrtj|nd|_ tj|j |dd|_tj||ddd|_tj |j||_|j"j%dtj&t)|j*Dcgc] }t-|c}|_tj0|j |_tj4dd|_d|_|j;ycc}w) Nr[rr) kernel_sizepaddingr)rstriderF)r)rBrCr"encoder_layerdrop layerdroprt num_mel_binsmax_source_positionsscale_embeddingmathsqrt embed_scalerConv1dconv1conv2 Embeddingembed_positionsrequires_grad_ ModuleListrangeencoder_layersrrlayersrx layer_norm AvgPool1d avg_poolergradient_checkpointing post_init)rOr?r9r`rHs r4rCzVoxtralEncoder.__init__s6  ~~ 11NN "//$*$?$?!393I3I499Y/sYYt00)TUV YYy)1VWX !||D,E,EyQ ++E2mm%PVPePeJf$gQ%8%@$gh ,,v~~6,,q3&+# %hs6F,cJ|jD] }d|_ d|_yrf) parameters requires_grad_requires_grad)rOparams r4_freeze_parametersz!VoxtralEncoder._freeze_parameterss(__& (E"'E  (#r6rXc|jSNrrOs r4get_input_embeddingsz#VoxtralEncoder.get_input_embeddings s zzr6rc||_yrrrOrs r4set_input_embeddingsz#VoxtralEncoder.set_input_embeddings#s  r6r1c |jj|jjdz|jjdz}|j d|k7r"t d|d|j dd|d|j|jjj|jjj}tjj|j|}tjj|j |}|jddd }|jj}||zj|j}tjj!||j |j" }t%|j&D]\}} | || } | d}|j)|}t+| S) a Args: input_features (`torch.LongTensor` of shape `(batch_size, feature_size, sequence_length)`): Float values of mel features extracted from the raw speech waveform. Raw speech waveform can be obtained by loading a `.flac` or `.wav` audio file into an array of type `list[float]` or a `numpy.ndarray`, *e.g.* via the soundfile library (`pip install soundfile`). To prepare the array into `input_features`, the [`AutoFeatureExtractor`] should be used for extracting the mel features, padding and conversion into a tensor of type `torch.FloatTensor`. See [`~WhisperFeatureExtractor.__call__`] attention_mask (`torch.Tensor`)`, *optional*): Voxtral does not support masking of the `input_features`, this argument is preserved for compatibility, but it is not used. By default the silence in the input log mel spectrogram are ignored. rr$z7Voxtral expects the mel input features to be of length z , but found z-. Make sure to pad the input mel features to .)rdevicerrr')r )last_hidden_state)r?rrrrshaperEtoweightrrrr.gelupermuterr"r) enumeraterrr ) rOrr r1expected_seq_length inputs_embeds embed_posrVidx encoder_layer layer_outputss r4rezVoxtralEncoder.forward&s(#kk>>ARARSTAUUX\XbXbXiXijkXll    #': :IJ]I^^jkykkAClDkEErsFrGGHI (**1B1B1H1HQUQ[Q[QbQbQiQi*j **4::n+EF  **4::m+DE %--aA6 ((// &266}7J7JK  --mt||VZVcVc-d "+DKK"8 - C)-M*!,M  - 6 )+  r6 input_lengthsc6|dz dzdz}|dz dzdz}||fS)zs Computes the output length of the convolutional layers and the output length of the audio encoder rrr)rOroutput_lengthss r4 _get_feat_extract_output_lengthsz/VoxtralEncoder._get_feat_extract_output_lengthsWs7'*q014 '!+1A5n,,r6r)rIrgrhrirrmain_input_namerrr8rr_can_record_outputsrCrrModulerrrrrrrnr rer+ LongTensorrrorps@r4rrs  ! &O./&, 32$ bii")) , +, , + + ,  , ^-e>N>N-r6rc*eZdZdeffd ZdZxZS)VoxtralMultiModalProjectorr?cft|tj|jj |j jd|_t|j|_ tj|j j|j jd|_ y)NFrA) rBrCrrJ audio_configintermediate_size text_config hidden_sizelinear_1rprojector_hidden_actactlinear_2rs r4rCz#VoxtralMultiModalProjector.__init__asz  &"5"5"G"GI[I[IgIgnst &556 &"4"4"@"@&BTBTB`B`glm r6cl|j|}|j|}|j|}|Sr)rrr)rOaudio_featuresrVs r4rez"VoxtralMultiModalProjector.forwardgs2 n5 /  m4 r6)rIrgrhrrCrerorps@r4rr`sn}n r6rzs The Voxtral model, which consists of Whisper encoder, a multi-modal projector and a LLama language model. ceZdZdgZfdZdZdZdZdZdZ dZ e e d d e jd eed eezfdZe e dde j(dzd e jdzde j*dzde j(dzdedzde jdzde j(dzdedzde j(dzdee j*zd eed efdZfdZxZS)VoxtralForConditionalGenerationrc*t|||jj|_t j |j |_tj |j|_ t||_ |jyr) rBrCr vocab_sizer from_configr audio_towerrlanguage_modelrmulti_modal_projectorrrs r4rCz(VoxtralForConditionalGeneration.__init__vsn   ,,77$001D1DE2>>v?Q?QR%?%G" r6c6|jjSr)rrrs r4rz4VoxtralForConditionalGeneration.get_input_embeddingss""7799r6c:|jj|yr)rrrs r4rz4VoxtralForConditionalGeneration.set_input_embeddingss 007r6c6|jjSr)rget_output_embeddingsrs r4rz5VoxtralForConditionalGeneration.get_output_embeddingss""88::r6c:|jj|yr)rset_output_embeddings)rOnew_embeddingss r4rz5VoxtralForConditionalGeneration.set_output_embeddingss 11.Ar6c:|jj|yr)r set_decoder)rOdecoders r4rz+VoxtralForConditionalGeneration.set_decoders ''0r6c6|jjSr)r get_decoderrs r4rz+VoxtralForConditionalGeneration.get_decoders""..00r6zThis method is used to get the audio embeddings from input features (a log mel spectrogram), meaning inferring the audio encoder and the multi-modal projector.rrr1rXc |j|fddi|}|j}|jd|jjj }|j |}||_|S)aa input_features (`torch.FloatTensor`): Float values of mel features extracted from the raw speech waveform. Raw speech waveform can be obtained by loading a `.flac` or `.wav` audio file into an array of type `list[float]` or a `numpy.ndarray`, *e.g.* via the soundfile library (`pip install soundfile`). To prepare the array into `input_features`, the [`AutoFeatureExtractor`] should be used for extracting the mel features, padding and conversion into a tensor of type `torch.FloatTensor`. See [`~WhisperFeatureExtractor.__call__`] return_dictTr$)rrr^r?rrr pooler_output)rOrr1 audio_outputsaudio_hidden_states audio_embedss r4get_audio_featuresz2VoxtralForConditionalGeneration.get_audio_featuressn)((TTTVT +==199"dkk>V>V>h>hi112EF &2 #r6N input_idsr position_idsrrlabels use_cachecache_positionlogits_to_keepc ||j|}|||j|dj} ||jjk(j d} |j | j|j| j|j}|jd||||||| | d| }|S)aj Example: ```python >>> from transformers import VoxtralForConditionalGeneration, AutoProcessor >>> import torch >>> device = "cuda" if torch.cuda.is_available() else "cpu" >>> repo_id = "mistralai/Voxtral-Mini-3B-2507" >>> processor = AutoProcessor.from_pretrained(repo_id) >>> model = VoxtralForConditionalGeneration.from_pretrained(repo_id, dtype=torch.bfloat16, device_map=device) >>> conversation = [ { "role": "user", "content": [ { "type": "audio", "url": "https://huggingface.co/datasets/hf-internal-testing/dummy-audio-samples/resolve/main/dude_where_is_my_car.wav", }, {"type": "text", "text": "What can you tell me about this audio?"}, ], } ] >>> inputs = processor.apply_chat_template(conversation) >>> inputs = inputs.to(device, dtype=torch.bfloat16) >>> outputs = model.generate(**inputs, max_new_tokens=30) >>> processor.batch_decode(outputs[:, inputs.input_ids.shape[1]:], skip_special_tokens=True) ["This audio is a humorous conversation between two friends, likely in English, where one of them is trying to figure out what the other's tattoo says."] ```T)rr$)r rrrrrrrr) rr rr?audio_token_id unsqueezemasked_scatterrrr)rOr rr rrrrrrrr1r audio_token_maskoutputss r4rez'VoxtralForConditionalGeneration.forwardsb  7D557 BM  %)*?22>t2TbbL!*T[[-G-G GRRSUV )88 ##M$8$89+> , )%+')) ,  , r6c|jdd}|jdd}t| |i|}|s|jdds||d<|S)Nris_first_iterationFrT)popgetrBprepare_inputs_for_generation)rOargsr1rr model_inputsrHs r4rz=VoxtralForConditionalGeneration.prepare_inputs_for_generations] $4d;#ZZ(r1sg, $ ! )9kkF&RR752F   H %!% II% <<% % << % LL4' % T\ %%4U)ryyU)p8+48+v "_ " " n-+n-  n-b   L&