rh]-ddlmZmZddlZddlmZddlmZddlmZddl m Z ddl m Z m Z mZdd lmZdd lmZmZmZdd lmZd d lmZmZd dlmZmZmZmZddlm Z GddeZ!GddeZ"GddeZ#edGddeZ$GddejJZ&edGdde#e Z'gd Z(y)!)OptionalUnionN)nn)ACT2FN)Cache)GenerationMixin)BaseModelOutputBaseModelOutputWithPastCausalLMOutputWithPast)Unpack)TransformersKwargsauto_docstringcan_return_tuple)check_model_inputs) AutoModelAutoModelForCausalLM)Qwen2AudioAttentionQwen2AudioEncoderQwen2AudioEncoderLayerQwen2AudioPreTrainedModel) VoxtralConfigc eZdZy)VoxtralAttentionN__name__ __module__ __qualname__~/var/www/html/ai-insurance-compliance-backend/venv/lib/python3.12/site-packages/transformers/models/voxtral/modular_voxtral.pyrr&r"rc eZdZy)VoxtralEncoderLayerNrr!r"r#r&r&*r$r"r&c$eZdZdZdZdZdZdZdZy)VoxtralPreTrainedModelTN)rrr _supports_flex_attn_supports_cache_class_supports_attention_backend_can_compile_fullgraph_no_split_modulesr!r"r#r(r(.s& "&!"&r"r(z: The Voxtral encoder, which is a Whisper encoder. ) custom_introc6eZdZeedZe ddeefdZ y)VoxtralEncoder) attentions hidden_statesNkwargsc |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 } | 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. rz:Qwen2Audio expects the mel input features to be of length z , but found z-. Make sure to pad the input mel features to .)dtypedevicerr)ptrainingN)attention_masklayer_head_mask)last_hidden_state)configmax_source_positionsconv1strideconv2shape ValueErrortoweightr7r8r functionalgelupermuteembed_positionsdropoutr: enumeratelayers layer_normr ) selfinput_featuresr;r3expected_seq_length inputs_embeds embed_posr2idx encoder_layer layer_outputss r#forwardzVoxtralEncoder.forwardCs&#kk>>ARARSTAUUX\XbXbXiXijkXll    #': :LM`Laamn|oCoCDFoGnHHuvIuJJKL (**1B1B1H1HQUQ[Q[QbQbQiQi*j **4::n+EF  **4::m+DE %--aA6 ((// &266}7J7JK  --mt||VZVcVc-d "+DKK"8 - C)- $M *!,M  - 6 +  r"N) rrr rr&_can_record_outputsrr rrWr!r"r#r0r08s;', - +, - - r"r0c*eZdZdeffd ZdZxZS)VoxtralMultiModalProjectorr>cft|tj|jj |j jd|_t|j|_ tj|j j|j jd|_ y)NF)bias) super__init__rLinear audio_configintermediate_size text_config hidden_sizelinear_1rprojector_hidden_actactlinear_2rOr> __class__s r#r_z#VoxtralMultiModalProjector.__init__usz  &"5"5"G"GI[I[IgIgnst &556 &"4"4"@"@&BTBTB`B`glm r"cl|j|}|j|}|j|}|SrX)rergrh)rOaudio_featuresr2s r#rWz"VoxtralMultiModalProjector.forward{s2 n5 /  m4 r")rrr rr_rW __classcell__rjs@r#r[r[tsn}n r"r[zs The Voxtral model, which consists of Whisper encoder, a multi-modal projector and a LLama language model. ceZdZdgZddiZddgdgfiZdgZfdZdZd Z d Z d Z d Z d Z dejfdZee ddeej(deejdeej*deej(deedeejdeej(deedeej(deeej*fdeedefdZfdZxZS)VoxtralForConditionalGenerationzlm_head.weightlm_head colwise_repr2logitsrJc*t|||jj|_t j |j |_tj |j|_ t||_ |jyrX) r^r_rc vocab_sizer from_configra audio_towerrlanguage_modelr[multi_modal_projector post_initris r#r_z(VoxtralForConditionalGeneration.__init__sn   ,,77$001D1DE2>>v?Q?QR%?%G" r"c6|jjSrX)rxget_input_embeddingsrOs r#r|z4VoxtralForConditionalGeneration.get_input_embeddingss""7799r"c:|jj|yrX)rxset_input_embeddings)rOvalues r#rz4VoxtralForConditionalGeneration.set_input_embeddingss 007r"c6|jjSrX)rxget_output_embeddingsr}s r#rz5VoxtralForConditionalGeneration.get_output_embeddingss""88::r"c:|jj|yrX)rxset_output_embeddings)rOnew_embeddingss r#rz5VoxtralForConditionalGeneration.set_output_embeddingss 11.Ar"c:|jj|yrX)rx set_decoder)rOdecoders r#rz+VoxtralForConditionalGeneration.set_decoders ''0r"c6|jjSrX)rx get_decoderr}s r#rz+VoxtralForConditionalGeneration.get_decoders""..00r"rPc|j|}|j}|jd|jjj }|j |}|S)a This 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. Args: 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__`] Returns: `torch.FloatTensor`: The audio embeddings. r5)rwr=reshaper>rarbry)rOrP audio_outputsaudio_hidden_states audio_embedss r#get_audio_embedsz0VoxtralForConditionalGeneration.get_audio_embedssZ((8 +==199"dkk>V>V>h>hi112EF r" input_idsr; position_idspast_key_valuesrRlabels use_cachecache_positionlogits_to_keepr3returnc ||j|}|/|j|} ||jjk(} | || <|jd||||||| | d| }|S)ap 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, torch_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."] ```)r;rrrRrrrrr!)r|rr>audio_token_idrx)rOrrPr;rrrRrrrrr3raudio_token_maskoutputss r#rWz'VoxtralForConditionalGeneration.forwardsb  7D557 BM  %00@L )DKK,F,FF .:M* ++>4+>+> , )%+')) ,  , r"c|jdd}|jd}t| |i|}| |ddk(r||d<|S)NrPrr)popgetr^prepare_inputs_for_generation)rOargsr3rPr model_inputsrjs r#rz=VoxtralForConditionalGeneration.prepare_inputs_for_generations\ $4d;$45wD!!1!12Dc5<</0D+,D DDL  r"rp)r(r0rp))typingrrrr activationsr cache_utilsr generationr modeling_outputsr r r processing_utilsr utilsrrr utils.genericrautorr qwen2_audio.modeling_qwen2_audiorrrrconfiguration_voxtralrrr&r(r0Moduler[rp__all__r!r"r#rs # ! )``&II/2 1 *  0 6 4 &4  4 n   K&