o -i @s2UddlmZmZmZddlmZmZmZmZm Z ddl Z ddl Z ddl mZddlmZddlmZmZddlmZddlmZmZmZddlmZdd lmZdd lmZdd l m!Z!dd l"m#Z#dd l$m%Z%m&Z&m'Z'ddl(m)Z)ddl*m+Z+ddl,m-Z-ddl.m/Z/ddl0m1Z1m2Z2m3Z3ddl4m5Z5m6Z6m7Z7m8Z8m9Z9ddl:m;Z;mZ>m?Z?m@Z@ddlAmBZBddlCmDZDddlEmFZFddlGmHZHmIZIddlJmKZKmLZLmMZMmNZNmOZOmPZPmQZQddlRmSZSmTZTmUZUmVZVmWZWddlXmYZYmZZZm[Z[ddl\m]Z]Gdddej^Z_Gd d!d!ej^Z`Gd"d#d#ej^ZaGd$d%d%ej^ZbGd&d'd'ZcGd(d)d)ej^ZdGd*d+d+eHZeGd,d-d-eHZfeeefBZgeehd.<Gd/d0d0ej^ZiGd1d2d2e=ZjGd3d4d4e;ejZkd5eele jmfd6enele2ffd7d8ZoGd9d:d:e9ZpGd;d<dd?d?ej^eUeVeTeWZsdS)@)IterableMappingSequence) AnnotatedAnyLiteral TypeAliascastN) BatchFeature) GlmAsrConfigGlmAsrProcessor)WhisperFeatureExtractor) ModelConfigSpeechToTextConfig VllmConfig)BaseDummyOptions)$get_tensor_model_parallel_world_size) PromptType) get_act_fn)MMEncoderAttention)ColumnParallelLinearQKVParallelLinearRowParallelLinear)QuantizationConfig)ApplyRotaryEmb)MultiModelKeys)MULTIMODAL_REGISTRY)MultiModalDataDictMultiModalFieldConfigMultiModalKwargsItems)DictEmbeddingItems ModalityDataModalityDataItemsMultiModalDataItemsMultiModalDataParser)BaseDummyInputsBuilderBaseMultiModalProcessorBaseProcessingInfoPromptReplacement PromptUpdatePromptUpdateDetails)IntermediateTensors)cached_tokenizer_from_config)cached_processor_from_config) TensorSchema TensorShape)DEFAULT_CONV_PARAMSDEFAULT_MAX_AUDIO_LEN_SDEFAULT_MERGE_FACTOR!_flatten_audio_features_by_length#_get_audio_output_lengths_for_tower_group_audio_embeddings_normalize_chunk_counts)MultiModalEmbeddings SupportsLoRASupportsMultiModal SupportsPPSupportsTranscription)AutoWeightsLoaderinit_vllm_registered_model maybe_prefix)ISO639_1_SUPPORTED_LANGScs6eZdZdZd fdd ZdedejfddZZ S) GlmAsrEncoderRotaryEmbeddinga$ Rotary Position Embedding for GLM-ASR encoder. Computes rotary position embeddings on-demand for efficiency. Only caches inv_freq as a buffer; cos/sin are computed during forward to avoid wasted computation during initialization and ensure correct device placement. returnNcstt|d|j|j}t|dr4|jr4|jdd}|jdd}t||}|jdd|_ n t|dd}|}d|_ ||_ ||_ d|t j d|d t jd |}|jd |d d dS)Nhead_dimrope_parameters rope_thetag@partial_rotary_factorg?attention_scalingrdtypeinv_freqF) persistent)super__init__getattr hidden_sizenum_attention_headshasattrrDgetintrGdimrCtorcharangefloatregister_buffer)selfconfigrCbaserFrUrK __class__^/home/ubuntu/veenaModal/venv/lib/python3.10/site-packages/vllm/model_executor/models/glmasr.pyrNTs(    z%GlmAsrEncoderRotaryEmbedding.__init__seq_lencCs0tj||jj|jjd}t||j}||jS)a9 Compute rotary position frequencies for given sequence length. Args: seq_len: The sequence length to compute embeddings for. Returns: Frequency tensor with shape [seq_len, dim/2]. Use .cos() and .sin() to get the rotary embedding components. )devicerJ)rVrWrKrbrJouterrG)rZraseqfreqsr_r_r`forwardrs   z$GlmAsrEncoderRotaryEmbedding.forward)rBN) __name__ __module__ __qualname____doc__rNrTrVTensorrf __classcell__r_r_r]r`rAJs rAcVeZdZdZ  ddedBdeffdd Zdejd ejd ejd ejfd d Z Z S)GlmAsrEncoderAttentiona$ Optimized Multi-headed Grouped Query Attention for GLM-ASR encoder. Uses vLLM's QKVParallelLinear for fused projections, ApplyRotaryEmb for rotary position embeddings, and MMEncoderAttention for hardware-optimized attention computation with automatic backend selection. N quant_configprefixc st||_|j|_|j|_t|d|j|_|j|j|_t |_ |j|j |_ t d|j|j |_ t|j|j|j|jd||dd|_t|j|jd||dd|_t|dd}|re|dd }nt|dd }t|j||_tdd |_t|j |j|jd |j |d d |_dS)Nnum_key_value_headsr0Tz .qkv_projbiasrprqz.o_projrDrFg?)enforce_enablegz.attn) num_heads head_sizescale num_kv_headsrq)rMrNr[rPrQrvrOryrCrtp_sizenum_heads_per_rankmaxnum_kv_heads_per_rankrqkv_projro_projrSrT rotary_dimrapply_rotary_embrattn)rZr[rprq rope_paramsrFr]r_r`rNsP      zGlmAsrEncoderAttention.__init__ hidden_statesrotary_pos_emb_cosrotary_pos_emb_sinrBcCs |j\}}}||\}}|j|j}|j|j} |j|| | gdd\} } } | |||j|j} | |||j|j} | |||j|j} || dd|jf||| dd|jf<|| dd|jf||| dd|jf<| | | | } | ||d} | | \}}|S)@ Args: hidden_states: [batch_size, seq_len, hidden_size] rotary_pos_emb_cos: [seq_len, rotary_dim/2] - cosine of rotary embeddings rotary_pos_emb_sin: [seq_len, rotary_dim/2] - sine of rotary embeddings Returns: [batch_size, seq_len, hidden_size] rU.N) shaper~r{rCr}splitviewrrrr)rZrrr batch_sizera_qkvq_sizekv_sizeqkv attn_outputoutputr_r_r`rfs$   zGlmAsrEncoderAttention.forwardNro rgrhrirjrstrrNrVrkrfrlr_r_r]r`rns$ <rncsJeZdZdZ  d dedBdeffdd Zdejd ejfd d Z Z S) GlmAsrEncoderMLPzk Optimized MLP for GLM-ASR encoder. Uses vLLM's parallel linear layers for better performance. Nrorprqcslt||_|j|_|j|_t|j|jd||dd|_t|j|_ t |j|jd||dd|_ dS)NTz.fc1rsz.fc2) rMrNr[rPintermediate_sizerfc1r hidden_actact_fnrfc2rZr[rprqr]r_r`rNs&   zGlmAsrEncoderMLP.__init__rrBcCs*||\}}||}||\}}|SN)rrr)rZrrr_r_r`rf! zGlmAsrEncoderMLP.forwardrrr_r_r]r`rsrcrm)GlmAsrEncoderLayerz Optimized Transformer encoder layer for GLM-ASR. Combines attention and MLP with residual connections and layer norms. Nrorprqcsrt|j|_t|||dd|_t|||dd|_t|dd}tj |j|d|_ tj |j|d|_ dS)Nz .self_attnrprqz.mlplayer_norm_epsh㈵>eps) rMrNrPrn self_attnrmlprOnn LayerNorminput_layernormpost_attention_layernormrZr[rprqrr]r_r`rN.s"   zGlmAsrEncoderLayer.__init__rrrrBcCsJ|}||}|j|||d}||}|}||}||}||}|S)r)rrr)rrrr)rZrrrresidualr_r_r`rfIs   zGlmAsrEncoderLayer.forwardrrr_r_r]r`r(s$rc@s$eZdZdZdZdejfddZdS)_GlmAsrEncoderOutputa Simple output container compatible with transformers' BaseModelOutput. This lightweight container holds the encoder output and is compatible with the transformers library's output format while being more efficient than a full dataclass. Attributes: last_hidden_state: Final layer hidden states from the encoder. Shape: [batch_size, seq_len, hidden_size] last_hidden_statercCs ||_dSrr)rZrr_r_r`rNzs z_GlmAsrEncoderOutput.__init__N)rgrhrirj __slots__rVrkrNr_r_r_r`rks rcseZdZdZdgdiZ  ddedBdeffdd Zd ej d e ej ej ffd d Z dej d e fddZ dee eej fd eefddZZS) GlmAsrEncodera Optimized GLM-ASR Audio Encoder with vLLM native implementation. This encoder processes audio features through convolutional layers followed by transformer layers with rotary position embeddings. Optimized for performance with: - QKVParallelLinear for fused attention projections - Tensor parallelism support via ColumnParallelLinear/RowParallelLinear - Quantization support - Flash Attention (SDPA) r~q_projk_projv_projNrorprqcst|_tjjjddd|_tjjjdddd|_t fddt j D|_ t dd }tjj|d |_t|_dS) Nr0) kernel_sizepaddingrH)rstridercs$g|]}td|dqS)z.layers.r)r).0 layer_idxr[rqrpr_r` s z*GlmAsrEncoder.__init__..rrr)rMrNr[rConv1d num_mel_binsrPconv1conv2 ModuleListrangenum_hidden_layerslayersrOrnormrA rotary_embrr]rr`rNs.   zGlmAsrEncoder.__init__ input_lengthsrBcCs0|dddd}|dddd}||fS)z Compute the output length after convolutions. Args: input_lengths: Input sequence lengths [batch_size] Returns: Tuple of (output after conv1, output after conv2) rHrr0r_)rZroutput_lengths_conv1output_lengths_conv2r_r_r` _get_feat_extract_output_lengthss z.GlmAsrEncoder._get_feat_extract_output_lengthsinput_featurescCstjj||}tjj||}|dd}|jd}||}| j |j d}| j |j d}|j D]}||||}q9||}t|dS)ag Forward pass through the encoder. Args: input_features: [batch_size, num_mel_bins, seq_len] Returns: _GlmAsrEncoderOutput: Object with .last_hidden_state attribute containing [batch_size, seq_len', hidden_size] where seq_len' is the sequence length after convolutions r0rHrIr)rVr functionalgelurr transposerrcostorJsinrrr)rZrroutput_seq_lenrotary_pos_embrr encoder_layerr_r_r`rfs       zGlmAsrEncoder.forwardweightsc Csddlm}gd}t|}t}|D]R\}}|D](\}} } | |vr%q|| |}|dr5||vr5q||} | j} | | || n|drN||vrNq||vrSq||} t| d|} | | || |q|S)zICustom weight loading to handle q_proj/k_proj/v_proj -> qkv_proj mapping.r)default_weight_loader))r~rr)r~rr)r~rrz.bias weight_loader) -vllm.model_executor.model_loader.weight_utilsrdictnamed_parameterssetreplaceendswithrrOadd) rZrrstacked_params_mapping params_dict loaded_paramsname loaded_weight param_name weight_nameshard_idparamrr_r_r` load_weightss0        zGlmAsrEncoder.load_weightsr)rgrhrirjpacked_modules_mappingrrrNrVrktuplerrrfrrrrlr_r_r]r`r~s$+ ,%rc@seZdZUdZeded<eeje ejBe ddddhdfed<eeje ejBe dddhdfed <eeje ejBe d fed <d S) GlmAsrFeatureInputsz Dimensions: - num_chunks: Number of audio chunks (flattened) - nmb: Number of mel bins - num_audios: Number of original audio files audio_featurestype num_chunksnmb chunk_length dynamic_dimsrfeature_attention_mask num_audios chunk_countsN) rgrhrirjr__annotations__rrVrklistr/r_r_r_r`rs$   rc@sFeZdZUdZdZeded<eee j e ddddhdfed<dS) GlmAsrEmbeddingInputsz Dimensions: - bn: Batch size - naf: Number of audio features - hs: Hidden size (must match the hidden size of language model backbone) audio_embedsrbnnafhsrN) rgrhrirjrrrrrrVrkr/r_r_r_r`r2s  r GlmAsrInputscsNeZdZdZ  d dededBdeffdd Zd ej d ej fd d Z Z S)GlmAsrMultiModalProjectora Projects audio encoder outputs to language model hidden space. This projector uses a two-layer MLP to map audio features from the encoder's intermediate size to the language model's hidden size. Uses vLLM's parallel linear layers for tensor parallelism support. Architecture: - Linear layer: intermediate_size -> hidden_size * 2 - Activation function (e.g., GELU) - Linear layer: hidden_size * 2 -> hidden_size Nror[rprqcsbtt|jj|jjd||dd|_t|j |_ t |jjd|jj||dd|_ dS)NrHz .linear_1) input_size output_sizerprqz .linear_2) rMrNr audio_configr text_configrPlinear_1rprojector_hidden_actactrlinear_2rr]r_r`rNSs     z"GlmAsrMultiModalProjector.__init__rrBcCs*||\}}||}||\}}|Sr)rr r )rZrrrr_r_r`rfhrz!GlmAsrMultiModalProjector.forwardr) rgrhrirjr rrrNrVrkrfrlr_r_r]r`rEsrc@s\eZdZdZdefddZdedefddZdede fdd Z de e e d Bffd d Zd S) GlmAsrProcessingInfoz Processing information provider for GLM-ASR model. Provides access to model configuration, processor, and feature extractor needed for audio preprocessing and multimodal integration. rBcCs |jtSr)ctx get_hf_configr rZr_r_r`rws z"GlmAsrProcessingInfo.get_hf_configkwargscKs|jjtfi|Sr)r get_hf_processorr rZrr_r_r`rzsz%GlmAsrProcessingInfo.get_hf_processorcKs|jdi|jSNr_)rfeature_extractorrr_r_r`get_feature_extractor}sz*GlmAsrProcessingInfo.get_feature_extractorNcCsddiS)Naudior_rr_r_r`get_supported_mm_limitssz,GlmAsrProcessingInfo.get_supported_mm_limits)rgrhrirjr robjectr rr rrrrTrr_r_r_r`r os r c @s\eZdZdZdeeefdefddZ d dedeeefdeeefdBde fd d Z dS) GlmAsrDummyInputsBuilderz Builder for dummy inputs used in profiling and testing. Generates dummy text prompts and audio data that match the expected format for GLM-ASR model inputs. Used for memory profiling and performance benchmarking. mm_countsrBcCs |dd}|j}|j|S)Nrr)rSinfor audio_token)rZrr hf_processorr_r_r`get_dummy_texts   z'GlmAsrDummyInputsBuilder.get_dummy_textNra mm_optionsc Cs`|j}|j}|dd}|r|dnd}t|jdt}t||} d|j| ||diS)Nrr max_audio_len)lengthr overrides) rr sampling_raterSrOrr2rT_get_dummy_audios) rZrarrrr#raudio_overridesr  audio_lenr_r_r`get_dummy_mm_datas    z*GlmAsrDummyInputsBuilder.get_dummy_mm_datar) rgrhrirjrrrTrrrr'r_r_r_r`rs  r hf_inputsrBcCsn|d}|dur#ttdtjd|ddtjd|ddtddSttdtdtdtddS)a Configure multimodal field batching strategy for GLM-ASR. Determines how to batch audio inputs based on whether chunking is used. When chunk_counts is present, features are flattened across chunks; otherwise, they are batched normally. Args: hf_inputs: Dictionary of preprocessed inputs from HuggingFace processor. Returns: Dictionary mapping field names to MultiModalFieldConfig objects that specify batching behavior. rNrrr)rrrr)rSrrbatchedflat_from_sizes)r(rr_r_r`_glmasr_field_configs$  r+csHeZdZdZdeeejfee Bde e e fdBffdd Z Z S)GlmAsrMultiModalDataParserz Custom parser for GLM-ASR multimodal data. Extends the base parser to handle GLM-ASR specific audio data formats, including both pre-computed audio embeddings and raw audio features. datarBNcs(t|trt|ddhtdSt|S)Nrr)modalityrequired_fieldsfields_factory) isinstancerr r+rM_parse_audio_data)rZr-r]r_r`r2s  z,GlmAsrMultiModalDataParser._parse_audio_data) rgrhrirjrrrVrkr!rr"r2rlr_r_r]r`r,sr,c seZdZdZdefddZdeedede dee fdd Z d e d e e efd ee efd ee efdef fdd Zdedee efdee effddZdedee efdedeefddZZS)GlmAsrMultiModalProcessorz GLM-ASR processor that inherits directly from BaseMultiModalProcessor for better performance and cleaner implementation. rBcCs|j}t|jdS)N) target_sr)rrr,r#)rZrr_r_r`_get_data_parsers  z*GlmAsrMultiModalProcessor._get_data_parser audio_listr processorc Cs|j}|j}t|dt}t||}t||}g} |D]#} t| tr't| n| jd} t d| |d|} | t | |q| S)Nr rr0) r#rrOr2rTr1rlenrr|appendmin) rZr6rr7r#rr  window_size max_windowsrr n_samplesn_chunksr_r_r`_calculate_chunk_countss   z1GlmAsrMultiModalProcessor._calculate_chunk_countspromptmm_data mm_kwargs tok_kwargscsFd|vr |d|d<|dg}|rt|ts|gn|}|s6|j|}||}tt |gdddS|jj di|}t di|d|j i}t j ||||d} d| vra| d| d <n%d | vrd | vr| d } t| tjrtj| jd | jd tjd } | | d <|jjdi|} ||| j| } tj| tjd | d<| S)Naudiosr) input_idspt) tensor_typer#)r@rArBrCinput_feature_maskrrrrrIrr_)poprSr1rr get_tokenizerencode_apply_hf_processor_tokens_onlyr rrr#rM_call_hf_processorrVrkonesrlongrr?rtensor)rZr@rArBrCrr6 prompt_idsroutputsrmaskr7rr]r_r`rMsJ    z,GlmAsrMultiModalProcessor._call_hf_processorr(hf_processor_mm_kwargscCst|Sr)r+)rZr(rTr_r_r`_get_mm_fields_config?sz/GlmAsrMultiModalProcessor._get_mm_fields_configmm_items out_mm_kwargscs|jjdi|}|j}|}|j}t|dd}||dur)|jt|dt} t|dt } | d} d} g| durddl m } m }| durd }| | D])}||}| ||}t|trqt|}||| | }t||}q[n/tt| D](}| ||d}t|trt|d }||| | }t|qd tffd d }td ||dgS)Nr<|pad|> merge_factor conv_paramsrrr0)_as_list_chunk_counts#_get_audio_output_lengths_from_maskritem_idxcshr|}nd}|dur||}|jd}ntd|dkr&tdgt|}tj|dS)Nrrz>Either feature_attention_mask or audio_embeds must be providedzAudio is too short)embed_token_id)rSr ValueErrorrTr*select_token_id)r] num_featuresrembed audio_tokensaudio_output_lengthsaudio_token_id out_mm_datar_r`get_replacement_glmasr}s    zMGlmAsrMultiModalProcessor._get_prompt_updates..get_replacement_glmasrr)r.target replacementr_)rrrJ get_vocabrrOrSrfr3r1get_data glmasr_utilsr[r\r1rrVstackr9rTsumitemrr8rP unsqueezer()rZrVrTrWr7 tokenizervocabr[rrYrZrrr[r\ start_idxcountend_idxrSlengthsidxrhr_rdr`_get_prompt_updatesFsX              z-GlmAsrMultiModalProcessor._get_prompt_updates)rgrhrirjr$r5rrr r rTr?rrrrr rMrrUr#rrr)ryrlr_r_r]r`r3sN    >    r3)r dummy_inputscseZdZeZgdddgdZdddedeffd d Ze d ed e d edBfddZ d e fddZ ded edBfddZded ejeejdfBfddZded efddZ  d6dejdejdedBdejdBded ejeBf d d!Zd"ejd ejdBfd#d$Zd%eeeejfd eefd&d'Ze d(ed efd)d*Ze d(ed+ed efd,d-Z e d.e!j"d(ed/ed0edBd+e#d1d2ed3edBd e$fd4d5Z%Z&S)7GlmAsrForConditionalGenerationr gate_projup_proj)r~ gate_up_projro)rq vllm_configrqcst|jj}|j}|jj}||_||_||_||dt|j |t |dd|_ t ||t |dd|_ Wdn1sAwY||t||jt |ddgd|_Wdn1sdwY|jj|_dS)Nr audio_towerrmulti_modal_projectorlanguage_modelLlamaForCausalLM)r hf_configrq architectures)rMrN model_configrrpmultimodal_configr[_mark_tower_modelrrr?rrr_mark_language_modelr>rrmake_empty_intermediate_tensors)rZrrqr[rprr]r_r`rNs:     z'GlmAsrForConditionalGeneration.__init__r.irBNcCs|drdStd)Nrz)<|begin_of_audio|><|pad|><|end_of_audio|>z Only audio modality is supported) startswithr_)clsr.rr_r_r`get_placeholder_strs z2GlmAsrForConditionalGeneration.get_placeholder_strcCstjddddS)Nzlanguage_model.zmulti_modal_projector.z audio_tower.)r connector tower_model)rfrom_string_fieldrr_r_r`get_mm_mappings z-GlmAsrForConditionalGeneration.get_mm_mappingrcKsX|dd}|durtd|dS|dd}|durdStd||dd|dddS)Nr)rrrrrr)rrrr)rIrr)rZrrrr_r_r`_parse_and_validate_audio_inputs     z>GlmAsrForConditionalGeneration._parse_and_validate_audio_input audio_input.cCsP|ddkr t|dS|d}|d}t|tr'tj|dd}tj|dd}|jd}t|d|d}|j|j j j j d }| |j }|jjj}|jjj}||} |jd } | | | } | | kro|ddd| ddf}||d |}||} t|jd t} t|jd t}t|j |d | |}t| |}t||}t||S)Nrrrrrrr)rrIr0rrYrZ)rr1rrVcatrr7rSrrrweightrJrr[rrPrreshaperrOr3r1r5ror4rflattentolistr6)rZrrrrraudio_hidden_statesrPr merge_ratioraseq_len_truncatedrrYrZremasked_audio_featureschunk_embeddingsr_r_r`_process_audio_inputsR             z3GlmAsrForConditionalGeneration._process_audio_inputcKs*|jdi|}|durgS||}|Sr)rr)rZrrrr_r_r`embed_multimodal+s  z/GlmAsrForConditionalGeneration.embed_multimodalrE positionsintermediate_tensors inputs_embedscKs$|durd}|jj||||d}|S)N)r)rmodel)rZrErrrrrr_r_r`rf4sz&GlmAsrForConditionalGeneration.forwardrcCs |j|Sr)rcompute_logits)rZrr_r_r`rGs z-GlmAsrForConditionalGeneration.compute_logitsrcCsdg}t||d}||S)Nzaudio_tower.embed_positions) skip_prefixes)r=r)rZrrloaderr_r_r`rMs  z+GlmAsrForConditionalGeneration.load_weightsrcCst|}t|ddS)znGet the audio token from processor. Similar to get_placeholder_str but returns single token. rrX)r-rO)rrr7r_r_r`_get_audio_tokenRs z/GlmAsrForConditionalGeneration._get_audio_token task_typecCs(t|}|j}t|dt}t||jdS)Nr )max_audio_clip_s sample_rate)r-rrOr2rr#)rrrr7rrr_r_r`get_speech_to_text_config[s z8GlmAsrForConditionalGeneration.get_speech_to_text_configr stt_configlanguage) transcribe translaterequest_prompt to_languagecCst|}||} |dkr|j||} | d| } n|dkr&| d} ntd|d| dg} |j| dd d } || }|d |id }tt|S) z@Get the generation prompt to be used for transcription requests.rztranslate the speech to rz4can you transcribe the speech into a written format?zUnsupported task type user)rolecontentFT)tokenizeadd_generation_promptr)prompt_token_idsmulti_modal_data) r,rsupported_languagesrSr_apply_chat_templaterKr r)rrrrrrrrrrrfull_lang_name_to user_contentmessagesr@r prompt_dictr_r_r`get_generation_promptgs$    z4GlmAsrForConditionalGeneration.get_generation_prompt)NN)'rgrhrir@rrrrrN classmethodrTrrrrrrrVrkrrr8rr+rfrrrrrrrrnpndarrayrrrrlr_r_r]r`r{s! @   $   r{)tcollections.abcrrrtypingrrrrr numpyrrVtorch.nnr transformersr transformers.models.glmasrr r transformers.models.whisperr vllm.configrrrvllm.config.multimodalrvllm.distributed.parallel_statervllm.inputs.datar%vllm.model_executor.layers.activationr9vllm.model_executor.layers.attention.mm_encoder_attentionr!vllm.model_executor.layers.linearrrr'vllm.model_executor.layers.quantizationr2vllm.model_executor.layers.rotary_embedding.commonr)vllm.model_executor.models.module_mappingrvllm.multimodalrvllm.multimodal.inputsrrrvllm.multimodal.parser r!r"r#r$vllm.multimodal.processingr%r&r'r(r)r* vllm.sequencer+vllm.tokenizersr,!vllm.transformers_utils.processorr-vllm.utils.tensor_schemar.r/rmr1r2r3r4r5r6r7 interfacesr8r9r:r;r<utilsr=r>r?whisperr@ModulerArnrrrrrrrrrr rrrkrr+r,r3register_processorr{r_r_r_r`sv                $  ;y*C*%   %: