o i|@sddlmZddlmZmZddlZddlmZddlm Z ddl m Z m Z ddl mZddlmZdd lmZmZdd lmZdd lmZdd lmZmZmZmZd dlmZd dlm Z m!Z!m"Z"m#Z#m$Z$m%Z%m&Z&m'Z'ddl(m)Z)m*Z*ddl+m,Z,e-e.Z/eeddGdddeZ0Gddde%Z1Gddde&Z2Gddde#Z3Gddde Z4Gdd d e!Z5ed!deGd"d#d#eZ6eGd$d%d%e$e6Z7Gd&d'd'ej8Z9ed(dGd)d*d*e"eZ:Gd+d,d,ej8Z;eGd-d.d.e$ZdS)3) dataclass)OptionalUnionN)check_model_inputs)Cache DynamicCache)GenerationMixin)create_causal_mask)BaseModelOutputWithPastCausalLMOutputWithPast)PreTrainedModel)Unpack) ModelOutputauto_docstringcan_return_tuplelogging) AutoModel)LlamaAttentionLlamaDecoderLayerLlamaForCausalLMLlamaMLP LlamaModel LlamaRMSNormLlamaRotaryEmbeddingTransformersKwargs) CsmConfigCsmDepthDecoderConfig)CsmGenerationMixinz: Base class for the model autoregressive outputs. ) custom_introc@seZdZUdZdZeejed<dZ eejed<dZ ee ed<dZ ee ejdfed<dZee ejdfed<dZeejed <dZeejed <dZee ed <dZee ejdfed <dZee ejdfed <dZeejed<dS)CsmOutputWithPasta loss (`torch.FloatTensor` of shape `(1,)`, *optional*, returned when `labels` is provided): Language modeling loss (for next-token prediction). logits (`torch.FloatTensor` of shape `(batch_size, sequence_length, config.vocab_size)`): Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax). past_key_values (`Cache`, *optional*, returned when `use_cache=True` is passed or when `config.use_cache=True`): It is a [`~cache_utils.Cache`] instance. For more details, see our [kv cache guide](https://huggingface.co/docs/transformers/en/kv_cache). Contains pre-computed hidden-states (key and values in the self-attention blocks) that can be used (see `past_key_values` input) to speed up sequential decoding. depth_decoder_loss (`torch.FloatTensor` of shape `(1,)`, *optional*, returned when `labels` is provided): Language modeling loss (for next-token prediction) of the depth decoder model. depth_decoder_logits (`torch.FloatTensor` of shape `(batch_size, sequence_length, config.vocab_size)`): Prediction scores of the depth decoder (scores for each vocabulary token before SoftMax). depth_decoder_past_key_values (`Cache`, *optional*, returned when `use_cache=True` is passed or when `config.use_cache=True`): It is a [`~cache_utils.Cache`] instance. For more details, see our [kv cache guide](https://huggingface.co/docs/transformers/en/kv_cache). depth_decoder_hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`): Tuple of `torch.FloatTensor` (one for the output of the embeddings, if the model has an embedding layer, + one for the output of each layer) of shape `(batch_size, sequence_length, hidden_size)`. Hidden-states of the model at the output of each layer plus the optional initial embedding outputs. depth_decoder_attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`): Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length, sequence_length)`. backbone_loss (`torch.FloatTensor` of shape `(1,)`, *optional*, returned when `labels` is provided): Language modeling loss (for next-token prediction) of the backbone model. Nlosslogitspast_key_values. hidden_states attentionsdepth_decoder_lossdepth_decoder_logitsdepth_decoder_past_key_valuesdepth_decoder_hidden_statesdepth_decoder_attentions backbone_loss)__name__ __module__ __qualname____doc__r#rtorch FloatTensor__annotations__r$r%rr&tupler'r(r)r*r+r,r-r6r6W/home/ubuntu/.local/lib/python3.10/site-packages/transformers/models/csm/modular_csm.pyr"1s r"c@ eZdZdS) CsmRMSNormNr.r/r0r6r6r6r7r9br9c@r8)CsmRotaryEmbeddingNr:r6r6r6r7r<fr;r<c@r8)CsmMLPNr:r6r6r6r7r=jr;r=c@r8) CsmAttentionNr:r6r6r6r7r>nr;r>c@r8)CsmDecoderLayerNr:r6r6r6r7r?rr;r?z[ The bare Csm Model outputting raw hidden-states without any specific head on top. csTeZdZUeed<dZdZdgZdgZdZ dZ dZ dZ e edZfddZZS) CsmPreTrainedModelconfigmodelTr?r%)r&r'csPt|t|tr$|j}t|dD]}|jj|jd|j j dqdSdS)Nrg)meanstd) super _init_weights isinstanceCsmCodebooksHead num_codebooksrangeweightdatanormal_rAinitializer_range)selfmodulerIi __class__r6r7rFs  z CsmPreTrainedModel._init_weights)r.r/r0rr4base_model_prefixsupports_gradient_checkpointing_no_split_modules_skip_keys_device_placement_supports_flash_attn_supports_sdpa_can_compile_fullgraph_supports_attention_backendr?r>_can_record_outputsrF __classcell__r6r6rRr7r@vs r@cseZdZUeed<fddZee        ddee j dee j dee j dee j d ee d ee j d eed ee j d eedeeeffddZZS)CsmDepthDecoderModelrAcs>t|t|j|j|j|_tj|j|j dd|_ dSNF)bias) rE__init__nn EmbeddingrI vocab_sizebackbone_hidden_size embed_tokensLinear hidden_sizeinputs_embeds_projectorrOrArRr6r7ras zCsmDepthDecoderModel.__init__N input_idsbackbone_last_hidden_stateattention_mask position_idsr% inputs_embeds use_cachecache_positionkwargsreturnc  Ks|durtjstdd}|du|duArtd|r(|dur(t|jd}|durX|dur4|nd} |dur?|j dn|j d} |durK|j n|j } tj | | | | d}|durtj |ddd} | |j }|||}|ddk}|dur||dddf<n tjs|rtd ||}t|j|||||d }|}|d}|||}|jd|jjD]}||f||||||d | }q||}t||r|d Sdd S) aJ backbone_last_hidden_state (`torch.FloatTensor` of shape `(batch_size, backbone_hidden_size)`, *optional*): The last hidden state of the backbone model. Such input is required when the first codebook token (the one generated by the backbone model) is provided in the `input_ids` argument. NzCustom `position_ids` were provided but will be ignored. CSM depth decoder automatically determines position_ids from `cache_position` and as it requires them to be identical across the batch, the provided position_ids will be ignored.z;You must specify exactly one of input_ids or inputs_embeds.)rArrdevice)minzvWhen the first codebook token is provided, `backbone_last_hidden_state` should also be provided for correct inference.)rA input_embedsrmrqr%rn)rmrnr%rprqposition_embeddings)last_hidden_stater%)r2compiler is_compilinglogger warning_once ValueErrorrrAget_seq_lengthshaperuarangeclamprdrfwarningrir unsqueeze rotary_emblayersnum_hidden_layersnormr )rOrkrlrmrnr%rorprqrrpast_seen_tokensinputs_seq_lengthru codebook_idxsoffsetinput_ids_are_first_codebook causal_maskr&rx decoder_layerr6r6r7forwardsr        zCsmDepthDecoderModel.forward)NNNNNNNN)r.r/r0rr4rarrrr2 LongTensorr3Tensorrboolrrrr5r rr]r6r6rRr7r^sD     r^cs&eZdZfddZdddZZS)rHcs0t||_tt|jd|||_dS)Nr)rErarIrb Parameterr2emptyrK)rOrhrIrdrRr6r7ras  zCsmCodebooksHead.__init__Ncsf|durjd}|jt|n |d}|j|fddtjdDtjddS)Nrc s2g|]}tjdd|ddf|jqSN)rb functionallinearT).0 codebook_idxcodebook_weightr&r6r7 s$z,CsmCodebooksHead.forward..rdim)rrKr2rrJstack)rOr&rq seq_lengthrr6rr7rs    zCsmCodebooksHead.forwardrr.r/r0rarr]r6r6rRr7rHs rHa$ The CsmDepthDecoder Model transformer, with a [`CsmCodebooksHead`] on top, which can be seen a position-specific language modeling head, allowing to use a different linear layer for each codebook (e.g. position 0 is the first codebook and uses the first codebook head, etc.) cseZdZdZdZdZfddZ    ddejde e de ejde ej de ejf fd d Z e e          dde ejd e ej de ejd e ejde ee eej fde ej de ejde ede ejdeeejfdeedeeeffddZZS)CsmDepthDecoderForCausalLMNcs2t||`t|j|j|j|_t||_ dSr) rEralm_headrHrhrIrdcodebooks_headr^rBrjrRr6r7ras z#CsmDepthDecoderForCausalLM.__init__rkr%rmrorqc  sHtj|||||fi|}|dddk}|s|d|d|S)Nrqrrlrn)rEprepare_inputs_for_generationpop) rOrkr%rmrorqrr model_inputsis_first_generation_steprRr6r7rs   z8CsmDepthDecoderForCausalLM.prepare_inputs_for_generationrrlrnlabelsrplogits_to_keeprrrsc  Ks|jd|||||||| d| } | d} t| tr+| dkr$tdd}n t| d}n| }|| dd|ddf| durA| |nd}|}d}|durg|dddf}|jd|d|jj|d| }t ||| j | j | j dS) a backbone_last_hidden_state (`torch.FloatTensor` of shape `(batch_size, backbone_hidden_size)`, *optional*): The last hidden state of the backbone model. Such input is required when the first codebook token (the one generated by the backbone model) is provided in the `input_ids` argument. labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*): Labels for computing the masked language modeling loss. Indices should either be in `[0, ..., config.vocab_size]` or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored (masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`. )rkrlrmrnr%rorprqrrN.)r$rrd shift_labels)r#r$r%r&r'r6) rBrGintslicer contiguous loss_functionrArdr r%r&r')rOrkrlrmrnr%rorrprqrrroutputsr& slice_indicesr$r#rr6r6r7r5sJ   & z"CsmDepthDecoderForCausalLM.forwardNNNN) NNNNNNNNNr)r.r/r0_tied_weights_keys_tp_plan_pp_planrar2rrrr3rrrrrlistrrrrr5r rr]r6r6rRr7r sr       rcs$eZdZfddZddZZS)CsmBackboneModelEmbeddingscsDtt|j|j|j|_|jdt |j|jdddS)Naudio_tokens_offsetsF) persistent) rErarbrcrIrdrhembed_audio_tokensregister_bufferr2rrjrRr6r7ra{s  z#CsmBackboneModelEmbeddings.__init__cCs |||j}|jdd}|S)Nrr)rrsum)rOrkrwr6r6r7rs z"CsmBackboneModelEmbeddings.forwardrr6r6rRr7rzs rcs0eZdZfddZeefddZZS)CsmBackboneModelcst|t||_dSr)rErarrfrjrRr6r7ras zCsmBackboneModel.__init__c stjdi|S)a& input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length, num_codebooks) or (batch_size, sequence_length)`): 1. (batch_size, sequence_length): corresponds to the input sequence prepared with the processor from the text prompt. Such input requires `input_values` to be provided so that audio can be encoded in codebook tokens and then merged with the text tokens. 2. (batch_size, sequence_length, num_codebooks): codebook tokens generated during the autoregressive decoding. Such input is not meant to be used by end users. Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and [`PreTrainedTokenizer.__call__`] for details. [What are input IDs?](../glossary#input-ids) Nr6)rEr)rO super_kwargsrRr6r7rszCsmBackboneModel.forward)r.r/r0rarrrr]r6r6rRr7rs  rz The Csm model consists of two llama-like auto-regressive transformer models: a backbone model that predicts the first codebook token and a depth decoder that predicts the other codebook tokens. cseZdZddgZfddZddZddZd d Zefd d Z fd dZ    d$de e j de e j de e j de e j de e j f ddZ    d$de jde ede e jde e jde e jf fdd Zee           d%de e jde e j de e j de e j de e jde eeee jfde e jde e jde ede e jd eee j fd!eedeeeffd"d#ZZS)&CsmForConditionalGenerationz5backbone_model.embed_tokens.embed_audio_tokens.weightz'depth_decoder.model.embed_tokens.weightcspt||j|_tj|j|jdd|_t|j|j|_ t ||_ t |j|_t|j|_|dSr_)rErardrbrgrhrrctext_vocab_sizeembed_text_tokensr _from_configbackbone_modelrdepth_decoder_config depth_decoderr from_config codec_config codec_model post_initrjrRr6r7ras   z$CsmForConditionalGeneration.__init__cCs|jjSrrrfrOr6r6r7get_input_embeddingssz0CsmForConditionalGeneration.get_input_embeddingscCs ||j_dSrr)rOvaluer6r6r7set_input_embeddingss z0CsmForConditionalGeneration.set_input_embeddingscCs(|jjr||jjj|jjjdSdSr)rAtie_codebooks_embeddings_tie_or_clone_weightsrrfrrrBrr6r6r7 _tie_weightss z(CsmForConditionalGeneration._tie_weightscs|ddrtj|i|\}}n tj|i|}dtfddt|jD}t|jjddi||D] }t |j|q?d|vrR||fS|S)Noutput_loading_infoFdepth_decoder_cs(i|]\}}|r|d|qSr) startswith)rattrrprefix prefix_lenr6r7 s  z?CsmForConditionalGeneration.from_pretrained.._from_model_config) getrEfrom_pretrainedlenvarsgeneration_configitemsrupdatedelattr)clsargsrrrB loading_infodepth_decoder_attrsrrRrr7rs   z+CsmForConditionalGeneration.from_pretrainedcsVd}|jj}|dd|D] \}}t|j|||qtj|i|dS)Nrtransformers_version)rr to_diff_dictrrsetattrrEsave_pretrained)rOrrrrrrrrRr6r7rs   z+CsmForConditionalGeneration.save_pretrainedNrk input_valuesinput_values_cutoffsrrscsF||}|durtj|d}||dk}||dk}tj||jd t |d}|| dk}t jg}t ||D]?\} } | | dk} t| jddD]+} | | } | | d} | d| | f}|j| d}|jdd}||dqUqBtdd |Dtfd d |D}|j|}Wdn1swY|jj}||k}|j|}||||<tjdd|jjf|jtjd |jj}|j|d}||jj k}|!|"d||<|dur| d!dd|jj}||||<|||<|d kj#dd}d||d|dddf<|}||dS)a Merges the input_ids and input_values to produce a single inputs_embeds tensor: 1 - Infers the codec model on the input_values to retrieve codebook token. 2 - Embeds codebook tokens and places them at the correct positions in the inputs_embeds tensor. 3 - If labels are provided, expands them to match codebook dimensions and position the target codebook tokens in the inputs_embeds tensor. Args: input_ids (`torch.Tensor` of shape `(batch_size, sequence_length)`): The input ids to embed. input_values (`torch.Tensor` of shape `(batch_size, channels, audio_sequence_length)`): The audio input values to embed. input_values_cutoffs (`torch.Tensor` of shape `(batch_size, max_num_audio)`): The cutoffs of the audio input values relative to its batch index, padded with -1 when no audio. Nrrrrtr.css|]}|jdVqdS)rN)rrelr6r6r7 szQCsmForConditionalGeneration._merge_input_ids_with_input_values..c s,g|]}tj|ddd|jdfqS)r)rbrpadrrmax_audio_framesr6r7rs,zRCsmForConditionalGeneration._merge_input_ids_with_input_values..)rudtypeiTas_tuple)ror)$rrbrrdiffr2rmaxruexpandrrno_gradziprJrrencode audio_codes transposeappendrget_audio_codes_maskrAaudio_token_idrrfonesrIlongcodebook_eos_token_idsqueezeaudio_eos_token_idrepeatrnonzero)rOrkrrrro audio_lengthsinput_values_maskaudio_tokens_listbatch_input_valuesbatch_input_values_cutoffsrQ start_idxend_idx audio_batch codec_outputs codebook_idsbatched_audio_token_idsaudio_codes_maskraudio_token_mask audio_embedsaudio_eos_frame_idsaudio_eos_embedsaudio_eos_token_masklabels_expanded depth_decoder_ignore_frames_idxsr6rr7"_merge_input_ids_with_input_valuess\            z>CsmForConditionalGeneration._merge_input_ids_with_input_valuesr%rmrorqc  stjd |||||d|}|dur>|jdkr>|ddur>|j||d|d|dd}||d|ddd|S) N)rkr%rmrorqrrorrr)rkrrr)rorrkr6)rErndimrr r) rOrkr%rmrorqrrr merged_inputsrRr6r7r9s(  z9CsmForConditionalGeneration.prepare_inputs_for_generationrrnrprrrc  Ks|dur|jdkr|||||} | d}| d}d}|jd|||||| | d| }|d}t| tr:t| dn| }||dd|ddf}d}d}d}d}|dur|dddddf}|jd|||jj d| }|ddddddfd kj d d }||d d|jj df}t j j|d dd}|jdd}||d|ddddf}||}|jd||| d|d| }|j}||}t|||||j|j|j|dur|jnd|dur|jnd|dur|jnd|dur|jd Sdd S)a input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length, num_codebooks) or (batch_size, sequence_length)`): 1. (batch_size, sequence_length): corresponds to the input sequence prepared with the processor from the text prompt. Such input requires `input_values` to be provided so that audio can be encoded in codebook tokens and then merged with the text tokens. 2. (batch_size, sequence_length, num_codebooks): codebook tokens generated during the autoregressive decoding. Such input is not meant to be used by end users. Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and [`PreTrainedTokenizer.__call__`] for details. [What are input IDs?](../glossary#input-ids) input_values_cutoffs (`torch.Tensor` of shape `(batch_size, max_num_audio)`, *optional*): Specify the end positions of audio segments within each batch entry, relative to the concatenated audio input. If a batch entry has fewer segments than the maximum, it is padded with -1. For example, in a batch of 2 sequences where the first contains 2 audio segments of length l1, and the second contains 1 audio segment of length l2, the input_values_cutoffs would be: [[l1, 2 * l1], [l2, -1]]. labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*): Labels for computing the masked language modeling loss. Indices should be in `[config.audio_token_id, -100, -101]`. Requires targeted `input_values` to be provided as audio tokens will be inferred from it using the `codec_model`. - `config.audio_token_id` indicates an audio frames (considering sequence length elements as frames) - `-100` will be ignored in the loss computation - `-101` indicates the audio frame will be used only for the backbone model (using the first codebook token as labels) Such labels can be prepared using `output_labels=True` when calling [`CsmProcessor`]. logits_to_keep (`int` or `torch.Tensor`, *optional*): Kept for compatibility. Does not support another value than: 1. `0`, which is equivalent to keeping all logits, used in the training regime 2. `1`, which is equivalent to keeping only the last logit, used in the generation regime Example: ```python >>> import torch >>> from transformers import CsmForConditionalGeneration, AutoProcessor >>> from datasets import load_dataset, Audio >>> model_id = "sesame/csm-1b" >>> torch_device = "cuda" if torch.cuda.is_available() else "cpu" >>> processor = AutoProcessor.from_pretrained(model_id) >>> ds = load_dataset("hf-internal-testing/dailytalk-dummy", split="train") >>> # ensure the audio is 24kHz >>> ds = ds.cast_column("audio", Audio(sampling_rate=24000)) >>> conversation = [] >>> # prepare a conversation with text and corresponding audio >>> for text, audio, speaker_id in zip(ds[:4]["text"], ds[:4]["audio"], ds[:4]["speaker_id"]): ... conversation.append( ... { ... "role": f"{speaker_id}", ... "content": [{"type": "text", "text": text}, {"type": "audio", "path": audio["array"]}], ... } ... ) >>> inputs = processor.apply_chat_template( ... conversation, ... tokenize=True, ... return_dict=True, ... output_labels=True, ... ).to(torch_device) >>> model = CsmForConditionalGeneration.from_pretrained(model_id, device_map=torch_device) >>> output = model(**inputs) >>> output.loss.backward() ```Nrror)rkrmrnr%rorprqr)r$rrdrrrr.r)rTr)rkrlrp return_dictr) r#r-r(r$r%r&r'r)r*r+r,r6)r!r rrGrrrrrArdallrIrbrrr rr#r"r%r&r'r$)rOrkrrmrrnr%rorrprqrrrr"backbone_outputsbackbone_hidden_statesrbackbone_logitsr#r-r(depth_decoder_outputsbackbone_labels train_maskdepth_decoder_input_ids train_idxsbackbone_last_hidden_statesdepth_decoder_labelsr6r6r7rXsS  (   z#CsmForConditionalGeneration.forwardr) NNNNNNNNNNr)r.r/r0rrarrr classmethodrrrr2rr rrr3rrrrrrrrrr5r"rr]r6r6rRr7rs     U      r)r@rr^rr)? dataclassesrtypingrrr2torch.nnrbtransformers.utils.genericr cache_utilsrr generationr masking_utilsr modeling_outputsr r modeling_utilsr processing_utilsrutilsrrrrautorllama.modeling_llamarrrrrrrrconfiguration_csmrrgeneration_csmr get_loggerr.r|r"r9r<r=r>r?r@r^ModulerHrrrr__all__r6r6r6r7sb        (  +_f U