^jdZddlZddlZddlZddlZddlmZddlmZddl m Z m Z m Z ddl Z ddlmZddlmZddlmZdd lmZdd lmZmZmZdd lmZmZmZmZmZm Z dd l!m"Z"m#Z#dd l$m%Z%ddl&m'Z'ddl(m)Z)m*Z*m+Z+m,Z,m-Z-ddl.m/Z/m0Z0ddl1m2Z2ddl3m4Z4m5Z5m6Z6ddl7m8Z8m9Z9ddl:m;Z;mZ>ddl?m@Z@ddlAmBZBmCZCe rddlDmEZEe6jeGZHe5eGdde,ZIde jdeKdeKfd ZLGd!d"ejZN d>d#ejd$e jd%e jd&e jd'e jdzd(eOdzd)eOd*e2e4fd+ZPGd,d-ejZQGd.d/e'ZRe5Gd0d1e0ZSGd2d3eSZTe5Gd4d5eSZUe5d67Gd8d9eSeZVe5d:7Gd;d unsqueezecatcossinviewzerosreget_default_dtype)r\rThalf_dimembs rArcz3MusicgenSinusoidalPositionalEmbedding.get_embedding{s !A%hhuoA.ii XU[[AGGISDPQll>=CCEOOPQRUXUbUbcdUeeii338a@EEnVXY 1  !))S%++na"@AqICvve--/00r@rBpast_key_values_lengthcr|j\}}}tj||zj|j}||j jdkDr|j ||j|j jd|jdjS)NrrF) sizer:rorerar^rYrT index_selectrudetach)rZrBrzbsz codebooksseq_len position_idss rAforwardz-MusicgenSinusoidalPositionalEmbedding.forwards"+.."2Y W-0FFJJ9K[K[\ T\\&&q) )   gt'9'9 :||((L,=,=b,ABIIKKr@)r)r5r6r7r8intrXrY staticmethodrcr:no_gradTensorr __classcell__r[s@rArRrRjsN8c8#8 G3GsG 1c 1# 1 1U]]_LLsLLr@rRmodulequerykeyvaluer3scalingdropoutkwargsc ||jddz}tj||jdd|z}|||z}tj j |d}tj j|||j}tj||} | jddj} | |fS)NrFr(r rjptrainingr,) r|r:matmulrInn functionalsoftmaxrr contiguous) rrrrr3rrr attn_weights attn_outputs rAeager_attention_forwardrs**R.D(<<s}}Q':;gEL!#n4 ==((2(>L==((6??([L,,|U3K''1-88:K  $$r@cXeZdZdZ ddedededzdedzdedzdedzd edzd edzffd Z dd e jd e jdzde dzde jdzdedzde e dee je jdzee jdzffdZxZS)MusicgenAttentionz=Multi-headed attention from 'Attention Is All You Need' paperN embed_dim num_headsr is_decoderbias is_causalconfig layer_idxc t |||_||_||_||z|_||_|j |z|jk7rtd|jd|d|j dz|_||_ ||_ ||_ tj||||_tj||||_tj||||_tj||||_y)Nz;embed_dim must be divisible by num_heads (got `embed_dim`: z and `num_heads`: z).rr)rWrXrrrhead_dimrrMrrrrrLineark_projv_projq_projout_proj) rZrrrrrrrrr[s rArXzMusicgenAttention.__init__s "" !Y.  MMI %$.. 8MdnnM]$YKr3 }}d* $""ii 94@ ii 94@ ii 94@  )YTB r@ hidden_stateskey_value_statespast_key_valuesr3output_attentionsrreturnc |du}|jdd}g|d|j} |j|j| j dd} d} |St |t rA|jj|j} |r |j} n|j} n|} |r|n|} |rK|I| rG j|jj}| j|jj}ng| jddd|j}|j| j|j dd}|j!| j|j dd}|K j#|||j\}}|r)t |t rd|j|j<t%j&|j(j*t,}||| |||f|j.sdn |j0|j2|d|\}}|j4g|dj7}|j9|}||fS) z#Input shape: Batch x Time x ChannelNrFr,r(FT)rrr)rKrrrurI isinstancer is_updatedgetrcross_attention_cacheself_attention_cachelayerskeysvaluesrrupdater get_interfacer_attn_implementationrrrrreshaperr)rZrrrr3rris_cross_attention input_shape hidden_shape query_statesrcurr_past_key_valuescurrent_states key_states value_stateskv_shapeattention_interfacerrs rArzMusicgenAttention.forwardsb.T9$))#2. 88b8$--8 {{=166|DNNqRST   &/+>?,77;;DNNK %+:+P+P(+:+O+O('6$-?)] /"=*-44T^^DIIJ/66t~~FMMLF--cr2FBF FH^499(CMMaQRSJ;;~6;;HEOOPQSTUL*+?+F+FzS_aeaoao+p( L%*_FY*ZAEO..t~~>(?(M(M KK , ,.E) %8      % $}}C$,,LL/ %  % ! \*k));;;;FFH mmK0 L((r@)rFTFNN)NNNF)r5r6r7r8rr>boolr-rXr:rr r rr9rrrs@rArrs:G !$"' !&(, $CCC C 4K C Tk C$;C%C:CH15(,.2). I)||I) ,,-I) I)  t+ I)  $; I)-.I) u||U\\D0% 2E2LL MI)r@rceZdZd deffd Z ddej dej dzdej dzdej dzdedzd edzd e e d ej fd Z xZ S)MusicgenDecoderLayerNrc *t||j|_t |j|j |j ddd|||_|j|_t|j|_ |j|_ tj|j|_t |j|j |j dd|||_tj|j|_tj$|j|j&d|_tj$|j&|jd|_tj|j|_y)NTF)rrrrrrrr)rrrrrr)rWrX hidden_sizerrnum_attention_headsattention_dropout self_attnrr activation_function activation_fnactivation_dropoutr LayerNormself_attn_layer_norm encoder_attnencoder_attn_layer_normrffn_dimfc1fc2final_layer_norm)rZrrr[s rArXzMusicgenDecoderLayer.__init__$s( ++*nn00,,  ~~ #F$>$>?"(";";$&LL$@!- NN  & &,, (*||DNN'C$99T^^V^^%H99V^^T^^%H " T^^ <r@rr3encoder_hidden_statesencoder_attention_maskr use_cacherrc |}|j|}|j|f||d|\}} tjj ||j|j }||z}|h|}|j |}|j|f|||d|\}} tjj ||j|j }||z}|}|j|}|j|j|}tjj ||j|j }|j|}tjj ||j|j }||z}|S)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. encoder_hidden_states (`torch.FloatTensor`): cross attention input to the layer of shape `(batch, seq_len, embed_dim)` encoder_attention_mask (`torch.FloatTensor`): encoder attention mask of size `(batch, 1, tgt_len, src_len)` where padding elements are indicated by very large negative values. past_key_values (`Cache`): cached past key and value projection states )rr3r)rr3r) rrrrrrrrrrrrr) rZrr3rrrrrresidual_s rArzMusicgenDecoderLayer.forwardEs*!11-@ *4>>  +)    q  --mt||VZVcVc-d  =0  ! ,$H 88GM0t00 !65 /      M1MM11-4<#> &+# ! _ as 'F8 F=NrBr3rrr inputs_embedsrrrc | | td|;|jdj|jdj\} } } n||ddddddfn td|r6|4t t j t j }||jnd} |tfdt D}tj |||}tj ||| }j| } || j|jz}tj j#|j"j$ }t'j(D]F\}}t+j,dd }j$r|j.kr8||||f|||d |}Hj1|}t3|| S)$ input_ids (`torch.LongTensor` of shape `(batch_size * num_codebooks, sequence_length)`): Indices of input sequence tokens in the vocabulary, corresponding to the sequence of audio codes. Indices can be obtained by encoding an audio prompt with an audio encoder model to predict audio codes, such as with the [`EncodecModel`]. See [`EncodecModel.encode`] for details. [What are input IDs?](../glossary#input-ids) The `input_ids` will automatically be converted from shape `(batch_size * num_codebooks, target_sequence_length)` to `(batch_size, num_codebooks, target_sequence_length)` in the forward pass. If you obtain audio codes from an audio encoding model, such as [`EncodecModel`], ensure that the number of frames is equal to 1, and that you reshape the audio codes from `(frames, batch_size, num_codebooks, target_sequence_length)` to `(batch_size * num_codebooks, target_sequence_length)` prior to passing them as `input_ids`. encoder_hidden_states (`torch.FloatTensor` of shape `(batch_size, encoder_sequence_length, hidden_size)`, *optional*): Sequence of hidden-states at the output of the last layer of the encoder. Used in the cross-attention of the decoder. encoder_attention_mask (`torch.LongTensor` of shape `(batch_size, encoder_sequence_length)`, *optional*): Mask to avoid performing cross-attention on padding tokens indices of encoder input_ids. Mask values selected in `[0, 1]`: - 1 for tokens that are **not masked**, - 0 for tokens that are **masked**. [What are attention masks?](../glossary#attention-mask) NzTYou cannot specify both decoder_input_ids and decoder_inputs_embeds at the same timerFzEYou have to specify either decoder_input_ids or decoder_inputs_embeds)rrc3VK|] }j|dd|f"ywrV)r).0codebookinputrZs rA z*MusicgenDecoder.forward..s.uT\ ; 1 1( ;E!X+st t  "%%b$*<*QRE*/++ 'C  &!!Q),Ede e 01,dkk2RT`hlhshsTtuOETE`!?!?!Afg  u`efs`tuuM+;;')+   ";;;'1"7 " ((0FG % ]5I5I(JJ  --mt||VZVcVc-d "+DKK"8  C"(..A"6 }}"5"F)%(> /# M    6 8++  r@NNNNNNN)r5r6r7r8rr&r_can_record_outputsr.rXr%r'r"r:r=rr;r rr r!r9rrrrs@rArrs$ .$%6aKX*+:>(,26!%d ##d*d  t+d  %0047 d !& 0 04 7 d  d ((4/d $;d +,d  : :d  d r@rc$eZdZdeffd ZdZdZeee dde jdzde jdzde jdzd e jdzd edzd e jdzd edzd eedeezfdZxZS) MusicgenModelrcdt||t||_|j yrV)rWrXrdecoderr )rZrr[s rArXzMusicgenModel.__init__(s&  &v.  r@c.|jjSrVr%rrZs rAget_input_embeddingsz"MusicgenModel.get_input_embeddings.s||(((r@c&||j_yrVr'rZrs rAset_input_embeddingsz"MusicgenModel.set_input_embeddings1s$) !r@NrBr3rrrrrrrc :|jd|||||||d|} | S)r)rBr3rrrrrr?)r%) rZrBr3rrrrrrdecoder_outputss rArzMusicgenModel.forward4sFZFRT\\ F )#9"7+' F  F r@r )r5r6r7r.rXr)r,r%r'r"r:r=rr;r rr r!r9rrrrs@rAr#r#&s4 )* .2.2:>:>(,26!%5##d*5 t+5 %0047 5 !& 0 04 7 5  5((4/5$;5+,5 : :5 5r@r#zK The MusicGen decoder model with a language modelling head on top. ) custom_introc8eZdZdZdeffd ZdZdZdZdZ e e e d!d e jdzd e jdzd e j dzd e jdzd edzde j dzde jdzdedzdeedeezfdZ d"dZd#d e jdededzfdZedZe j: d$de jdzdedzdedzde dzdedzde!df d Z"xZ#S)%MusicgenForCausalLMaudiorc Zt||t||_|j|_t j t|jDcgc].}t j|j|jd0c}|_ |jycc}w)NFr) rWrXr#rrrrrrrrlm_headsr )rZrrr[s rArXzMusicgenForCausalLM.__init__ws  "6* #11 SXY_YmYmSn oaRYYv))6+<+<5 I o   ps3B(cB|jjjSrVrr%rr(s rAr)z(MusicgenForCausalLM.get_input_embeddingsszz!!...r@c:||jj_yrVr7r+s rAr,z(MusicgenForCausalLM.set_input_embeddingss*/ 'r@c|jSrVr5r(s rAget_output_embeddingsz)MusicgenForCausalLM.get_output_embeddingss }}r@c||_yrVr:rZnew_embeddingss rAset_output_embeddingsz)MusicgenForCausalLM.set_output_embeddingss & r@NrBr3rrrrlabelsrrrc |9|7|5t||jj|jj}|j|f||||||d| } | j } t j|jD cgc] } | |  c} d} d}|| dddd|jd df}t}t jg|j}|j||jjk(d}t|jjD]f}|dd|fj!j#d|jd}|d|fj!j#d}||||z }h||jjz }| j$dg| jd d} t'|| | j(| j*| j,| j. Scc} w) a input_ids (`torch.LongTensor` of shape `(batch_size * num_codebooks, sequence_length)`): Indices of input sequence tokens in the vocabulary, corresponding to the sequence of audio codes. Indices can be obtained by encoding an audio prompt with an audio encoder model to predict audio codes, such as with the [`EncodecModel`]. See [`EncodecModel.encode`] for details. [What are input IDs?](../glossary#input-ids) The `input_ids` will automatically be converted from shape `(batch_size * num_codebooks, target_sequence_length)` to `(batch_size, num_codebooks, target_sequence_length)` in the forward pass. If you obtain audio codes from an audio encoding model, such as [`EncodecModel`], ensure that the number of frames is equal to 1, and that you reshape the audio codes from `(frames, batch_size, num_codebooks, target_sequence_length)` to `(batch_size * num_codebooks, target_sequence_length)` prior to passing them as `input_ids`. encoder_hidden_states (`torch.FloatTensor` of shape `(batch_size, encoder_sequence_length, hidden_size)`, *optional*): Sequence of hidden-states at the output of the last layer of the encoder. Used in the cross-attention of the decoder. encoder_attention_mask (`torch.LongTensor` of shape `(batch_size, encoder_sequence_length)`, *optional*): Mask to avoid performing cross-attention on padding tokens indices of encoder input_ids. Mask values selected in `[0, 1]`: - 1 for tokens that are **not masked**, - 0 for tokens that are **masked**. [What are attention masks?](../glossary#attention-mask) labels (`torch.LongTensor` of shape `(batch_size, sequence_length, num_codebooks)`, *optional*): Labels for language modeling. Note that the labels **are shifted** inside the model, i.e. you can set `labels = input_ids` Indices are selected in `[-100, 0, ..., config.vocab_size]` All labels set to `-100` are ignored (masked), the loss is only computed for labels in `[0, ..., config.vocab_size]` N)r3rrrrrr,rjrarHrF.r()losslogitsrrrr)rPrrC bos_token_idrrr:stackr5rKrrvra masked_fillrrrrurrrrrr)rZrBr3rrrrr@rroutputsrhead lm_logitsrCrDloss_fctrcodebook_logitscodebook_labelss rArzMusicgenForCausalLM.forwardsf  Y%6=;P*64;;3K3KT[[MeMefI=GTZZ  > )"7#9+' >  >  11 KK Om!4 OUVW   q!fll1o%5%778F')H;;r$++6D''$++2J2J(JDQF"$++";";< C"(H"5"@"@"B"G"GFLLY[L\"]"(h"7"B"B"D"I"I""M/BB C $++333D&I%%b?9??12+>? 0#33!//))$55   7!Ps?Hc  *|?|j||jj|jj\}}|j ||}|)|dkDr$|j d}||j d}| |ddddf}||||||dS)NrC max_lengthr,r(r,rF)rBr3rrrr)build_delay_pattern_maskgeneration_configrCrPapply_delay_pattern_maskrepeat) rZrBr3rrrrdelay_pattern_maskr4rs rAprepare_inputs_for_generationz1MusicgenForCausalLM.prepare_inputs_for_generations  %,0,I,I!33@@11<<-J- )I)11)=OP  %.1*<"((0I)!/!6!6v!>  &!!RS&)I#,%:&<."   r@rCrPcP|jd|j|jd}|j\}}}||n|jj}t j |||ft j|jdz}|jjdk(r|dzn|}|d|zdz kr*|j||zd|j||zdfSt|D]p} |jjdk(r|dd| f|dd| | || zf<3|ddd| zf|ddd| z| || zf<|ddd| zdzf|ddd| zdz| || zf<rt jt j ||ft j||z dz} | t jt j ||ft jz} |jjdk(r| jdd } | j!|j} | |z| |zz}|dddddf} | dk(j#dddf} t%| dkDr t'| }n|}|j||zd}|d d|fj||zd}||fS) aDBuild a delayed pattern mask to the input_ids. Each codebook is offset by the previous codebook by one, giving a delayed pattern mask at the start of sequence and end of sequence. Take the example where there are 4 codebooks and a max sequence length of 8, we have the delayed pattern mask of shape `(codebooks, seq_len)`: - [P, -1, -1, -1, -1, P, P, P] - [P, P, -1, -1, -1, -1, P, P] - [P, P, P, -1, -1, -1, -1, P] - [P, P, P, P, -1, -1, -1, -1] where P is the special padding token id and -1 indicates that the token is valid for prediction. If we include a prompt (decoder input ids), the -1 positions indicate where new tokens should be predicted. Otherwise, the mask is set to the value in the prompt: - [P, a, b, -1, -1, P, P, P] - [P, P, c, d, -1, -1, P, P] - [P, P, P, e, f, -1, -1, P] - [P, P, P, P, g, h, -1, -1] where a-h indicate the input prompt (decoder input ids) that are offset by 1. Now, we only override the -1 tokens in our prediction. rFNr_r(r,ri)diagonalrrj.)rrrKrSrPr:oneslongraraudio_channelsrtriurtrilrepeat_interleaverenonzerolenmin)rZrBrCrPrrrinput_ids_shiftedchannel_codebooksr delay_patternmaskfirst_codebook_ids start_idsfirst_start_id pattern_masks rArRz,MusicgenForCausalLM.build_delay_pattern_mask!s(%%b$*<*QR &/oo#]G#-#9Zt?U?U?`?` JJ]J7uzzR[RbRb cfh h 37++2L2LPQ2QMQ.Wd --1 1$$S=%8"=?P?X?XY\_lYlnp?qq q/0 wH{{))Q.PYZ[]eZePf!!Xx'H:L/L"LMU^^_abemam^mTn!!Q\8g>P3P"PQXabcefiqeqtueubuXv!!Q\A%5x'HBT7T"TU w JJ):6ejj IT^arTruvTv & 5::?PR\>]ejeoeo3p(qq ;; % % *);;A1;EM  !1!122,,u|/CC 'q!Qw/'2-668A> y>A  ^N%N!((}) Most generation-controlling parameters are set in `generation_config` which, if not passed, will be set to the model's default generation configuration. You can override any `generation_config` by passing the corresponding parameters to generate(), e.g. `.generate(inputs, num_beams=4, do_sample=True)`. For an overview of generation strategies and code examples, check out the [following guide](./generation_strategies). Parameters: inputs (`torch.Tensor` of varying shape depending on the modality, *optional*): The sequence used as a prompt for the generation or as model inputs to the encoder. If `None` the method initializes it with `bos_token_id` and a batch size of 1. For decoder-only models `inputs` should be in the format `input_ids`. For encoder-decoder models *inputs* can represent any of `input_ids`, `input_values`, `input_features`, or `pixel_values`. generation_config (`~generation.GenerationConfig`, *optional*): The generation configuration to be used as base parametrization for the generation call. `**kwargs` passed to generate matching the attributes of `generation_config` will override them. If `generation_config` is not provided, the default will be used, which had the following loading priority: 1) from the `generation_config.json` model file, if it exists; 2) from the model configuration. Please note that unspecified parameters will inherit [`~generation.GenerationConfig`]'s default values, whose documentation should be checked to parameterize generation. logits_processor (`LogitsProcessorList`, *optional*): Custom logits processors that complement the default logits processors built from arguments and generation config. If a logit processor is passed that is already created with the arguments or a generation config an error is thrown. This feature is intended for advanced users. stopping_criteria (`StoppingCriteriaList`, *optional*): Custom stopping criteria that complement the default stopping criteria built from arguments and a generation config. If a stopping criteria is passed that is already created with the arguments or a generation config an error is thrown. This feature is intended for advanced users. synced_gpus (`bool`, *optional*, defaults to `False`): Whether to continue running the while loop until max_length (needed to avoid deadlocking with `FullyShardedDataParallel` and DeepSpeed ZeRO Stage 3). streamer (`BaseStreamer`, *optional*): Streamer object that will be used to stream the generated sequences. Generated tokens are passed through `streamer.put(token_ids)` and the streamer is responsible for any further processing. kwargs (`dict[str, Any]`, *optional*): Ad hoc parametrization of `generate_config` and/or additional model-specific kwargs that will be forwarded to the `forward` function of the model. If the model is an encoder-decoder model, encoder specific kwargs should not be prefixed and decoder specific kwargs should be prefixed with *decoder_*. Return: [`~utils.ModelOutput`] or `torch.LongTensor`: A [`~utils.ModelOutput`] (if `return_dict_in_generate=True` or when `config.return_dict_in_generate=True`) or a `torch.FloatTensor`. If the model is *not* an encoder-decoder model (`model.config.is_encoder_decoder=False`), the possible [`~utils.ModelOutput`] types are: - [`~generation.GenerateDecoderOnlyOutput`], - [`~generation.GenerateBeamDecoderOnlyOutput`] If the model is an encoder-decoder model (`model.config.is_encoder_decoder=True`), the possible [`~utils.ModelOutput`] types are: - [`~generation.GenerateEncoderDecoderOutput`], - [`~generation.GenerateBeamEncoderDecoderOutput`] Nr2r3rrBrr4rFrP min_lengthrShas_default_max_lengthhas_default_min_lengthmodel_input_name inputs_tensorinput_ids_lengthr,r)generation_mode batch_sizemax_cache_lengthrOrVrSinput_ids_seq_lengthencoder_input_idsprefix_allowed_tokens_fnrorarSrp)rB expand_size)rorprSrqrrGot incompatible mode for generation, should be one of greedy or sampling. Ensure that beam search is de-activated by setting `num_beams=1`.r?).rScopydeepcopyrvalidate_validate_model_kwargsrrr_prepare_model_inputsrErKr_prepare_special_tokensrarr4&_prepare_attention_mask_for_generationrPru_prepare_generated_length_validate_generated_lengthris_encoder_decoder_prepare_cache_for_generationrR_decoder_start_token_tensorputcpuget_generation_modeappendr_get_logits_processor_get_stopping_criteriarSAMPLE GREEDY_SEARCH_expand_inputs_for_generationnum_return_sequences_samplerMreturn_dict_in_generate sequencesrT_pad_token_tensorr)rZrnrSrorprqrrr model_kwargsrequires_attention_maskkwargs_has_attention_maskrBryr}r{rwrxr~rVr|rH output_idss rAgeneratezMusicgenForCausalLM.generatessV  $ $ 6 6  MM*;</(//9&9 ""$ ##L$5$5$780@/K+QdQf1B1N-ThTj"3<"G$0$4$45Et$LTX$X!594N4N %22L5 1 #\__Q'4+=+==  $$%68QZcZjZj$k%6$?$? [!):)I)I %&   ,d 3 ;@W-1-X-X,l.L) * %??2.!'L!9T!A!nFWFbFbjnFn!'L!9T!A!nFWFbFbjnFn ::/#9#9-#- ;  ''(9;KMcd -77!;  " "1 %)9 9 O3KK22  0 6 6q 9 9  **   !- + )-(E(E *FF(33)F) % %   LL ).@ )*,??A  + + 7:>(,26*.!%b ##d*b  t+b  %0047 b !& 0 04 7 b  b ((4/b   4'b $;b +,b  2 2b  b N"#( TG'%2B2BG'RUG'cfimcmG'RU]]_'+597;9=#'-1Z t#Z,d2Z.4 Z 0$6 Z D[ Z>*ZZr@r1z_ The composite MusicGen model with a text encoder, audio encoder and Musicgen decoder, c eZdZUeed<dZdZdZdZ d;dedzde dzde dzd e dzffd Z d Z d Z d Ze dd edzfd!Z djA}d|vr td|jCy)a text_encoder (`PreTrainedModel`, *optional*): The text encoder model that encodes text into hidden states for conditioning. audio_encoder (`PreTrainedModel`, *optional*): The audio encoder model that encodes audio into hidden states for conditioning. decoder (`MusicgenForCausalLM`, *optional*): The decoder model that generates audio tokens based on conditioning signals. NzlEither a configuration has to be provided, or all three of text encoder, audio encoder and MusicGen decoder.rrr%zConfig: z has to be of type zIf `cross_attention_hidden_size` is specified in the MusicGen decoder's configuration, it has to be equal to the text encoder's `hidden_size`. Got z6 for `config.decoder.cross_attention_hidden_size` and z' for `config.text_encoder.hidden_size`.r()AutoModelForTextEncodingr*zConfig of the text_encoder: z/ is overwritten by shared text_encoder config: zConfig of the audio_encoder: z0 is overwritten by shared audio_encoder config: zConfig of the decoder: z* is overwritten by shared decoder config: z The encoder zB should not have a LM Head. Please use a model without and LM HeadrzThe selected decoder is not prepared for the encoder hidden states to be passed. Please see the following discussion on GitHub: https://github.com/huggingface/transformers/issues/23350)"rMr-rr config_classr%cross_attention_hidden_sizerrrWrXauto.modeling_autor from_configr+rr1 _from_configto_dictloggerwarningr[rrrenc_to_dec_projr;setinspect signaturer parametersrr ) rZrrrr%rr+decoder_signaturer[s rArXz)MusicgenForConditionalGeneration.__init__]sw >|3}7LPWP_~  >#)00 @T@T^e^l^lFfd&7&78 8F83FtGXGXFY!Z[[ >> 5 5 A~~99V=P=P=\=\\ AAGAkAk@lmIIOI\I\IhIhHij::    E3??@S@STL  61I11&2F2FGM ?)66v~~FG(*    # # + + -1I1I1Q1Q1S S NN.t/@/@/J/J.KLKK,,-/     $ $ , , .$++2K2K2S2S2U U NN/0B0B0L0L/MNKK--.0  <<   & & (DKK,?,?,G,G,I I NN)$,,*@*@)ABKK''(* 9=8I8I8P8P8e8e   59=9K9K9R9R9g9g !!637<<3F3F3[3[ 0#';;#;#; $(KK$=$=!"kk11     $ $ 0 0DLL4G4G4S4S S ##??G#%99T->->-E-E-Q-QSWS_S_SfSfSrSr#sD    2 2 4 @t0011st   1 1$,,2F2F G R R W W YZ "*; ;k  r@c6|jjSrV)rr)r(s rAr)z5MusicgenForConditionalGeneration.get_input_embeddingss  5577r@c6|jjSrV)r%r;r(s rAr;z6MusicgenForConditionalGeneration.get_output_embeddingss||1133r@c8|jj|SrV)r%r?r=s rAr?z6MusicgenForConditionalGeneration.set_output_embeddingss||11.AAr@*text_encoder_pretrained_model_name_or_path+audio_encoder_pretrained_model_name_or_path%decoder_pretrained_model_name_or_pathrc |jDcic]%\}}|jdr|tdd|'}}}|jDcic]%\}}|jdr|tdd|'} }}|jDcic]&\}}|jds|tdd|(} }}|D]} |d| z= | D]} |d| z= | D]} |d| z= |jdd} | | t dd|vret j |fi|dd i\} }| jd us| jd ur'tjd |d d | _d | _| |d<tj |g|i|} | jdd}|| t d d| vret j |fi| dd i\} } | jd us| jd ur'tjd |dd | _d | _| | d<tj |g|i| }| jdd}|| t dd| vrt j |fi| dd i\}} t|tr |j}|jd us|jd ur-tjd |d|d|dd |_d |_|| d<| djd us| djd urtjd|d|dt!j |fi| }td| j"|j"|j"d|}|| |||Scc}}wcc}}wcc}}w)a Instantiate a text encoder, an audio encoder, and a MusicGen decoder from one, two or three base classes of the library from pretrained model checkpoints. The model is set in evaluation mode by default using `model.eval()` (Dropout modules are deactivated). To train the model, you need to first set it back in training mode with `model.train()`. Params: text_encoder_pretrained_model_name_or_path (`str`, *optional*): Information necessary to initiate the text encoder. Can be either: - A string, the *model id* of a pretrained model hosted inside a model repo on huggingface.co. - A path to a *directory* containing model weights saved using [`~PreTrainedModel.save_pretrained`], e.g., `./my_model_directory/`. audio_encoder_pretrained_model_name_or_path (`str`, *optional*): Information necessary to initiate the audio encoder. Can be either: - A string, the *model id* of a pretrained model hosted inside a model repo on huggingface.co. - A path to a *directory* containing model weights saved using [`~PreTrainedModel.save_pretrained`], e.g., `./my_model_directory/`. decoder_pretrained_model_name_or_path (`str`, *optional*, defaults to `None`): Information necessary to initiate the decoder. Can be either: - A string, the *model id* of a pretrained model hosted inside a model repo on huggingface.co. - A path to a *directory* containing model weights saved using [`~PreTrainedModel.save_pretrained`], e.g., `./my_model_directory/`. model_args (remaining positional arguments, *optional*): All remaining positional arguments will be passed to the underlying model's `__init__` method. kwargs (remaining dictionary of keyword arguments, *optional*): Can be used to update the configuration object (after it being loaded) and initiate the model (e.g., `output_attentions=True`). - To update the text encoder configuration, use the prefix *text_encoder_* for each configuration parameter. - To update the audio encoder configuration, use the prefix *audio_encoder_* for each configuration parameter. - To update the decoder configuration, use the prefix *decoder_* for each configuration parameter. - To update the parent model configuration, do not use a prefix for each configuration parameter. Behaves differently depending on whether a `config` is provided or automatically loaded. Example: ```python >>> from transformers import MusicgenForConditionalGeneration >>> # initialize a musicgen model from a t5 text encoder, encodec audio encoder, and musicgen decoder >>> model = MusicgenForConditionalGeneration.from_sub_models_pretrained( ... text_encoder_pretrained_model_name_or_path="google-t5/t5-base", ... audio_encoder_pretrained_model_name_or_path="facebook/encodec_24khz", ... decoder_pretrained_model_name_or_path="facebook/musicgen-small", ... ) >>> # saving model after fine-tuning >>> model.save_pretrained("./musicgen-ft") >>> # load fine-tuned model >>> model = MusicgenForConditionalGeneration.from_pretrained("./musicgen-ft") ``` text_encoder_Naudio_encoder_decoder_rzxIf `text_encoder_model` is not defined as an argument, a `text_encoder_pretrained_model_name_or_path` has to be defined.rreturn_unused_kwargsTz Initializing z\ as a text_encoder model from a decoder model. Cross-attention and causal mask are disabled.Fz{If `audio_encoder_model` is not defined as an argument, an `audio_encoder_pretrained_model_name_or_path` has to be defined.z^ as an audio_encoder model from a decoder model. Cross-attention and causal mask are disabled.znIf `decoder_model` is not defined as an argument, a `decoder_pretrained_model_name_or_path` has to be defined.z9 as a decoder model. Cross attention layers are added to z and randomly initialized if z2's architecture allows for cross attention layers.zDecoder model z9 is not initialized as a decoder. In order to initialize z as a decoder, make sure that the attributes `is_decoder` and `add_cross_attention` of `decoder_config` passed to `.from_sub_models_pretrained(...)` are set to `True` or do not pass a `decoder_config` to `.from_sub_models_pretrained(...)`r)rrr%rr?)items startswithrapoprMr)from_pretrainedradd_cross_attentionrinfor+rr-r%rr1r)clsrrr model_argsrargumentrkwargs_text_encoderkwargs_audio_encoderkwargs_decoderrrencoder_configrr%decoder_configrs rAfrom_sub_models_pretrainedz;MusicgenForConditionalGeneration.from_sub_models_pretrainedsT$*<<> %""?3 S)+ ,e 3  $*<<> %""#34 S)*, -u 4  GMlln 3B8UX`XkXklvXwHS_& ' .  ' .C,- .' /C'#-. /! )CzC'( ) +..w=  9A % 226@6P6P>7BU7lp73 3",,48Z8Z^b8bKK'(R'ST^^16N-9>N60>#H-$44:=GK^L-00$?  :B % 337A7Q7Q?8CW8nr84 4",,48Z8Z^b8bKK'(S'TU^^16N-9>N61?$X.%55;>HL`M!$$Wd3 ?4< % ~-1;1K1K92=K2bf2.nn=%3%;%;N!,,59[9[_d9dKK'(M'NO00U/VWABBtv 15N-9=N6+9x(h'22e;~h?W?k?kot?t$%J$KL..S-TUMM*99:_rcqrG  %,,M The `decoder_input_ids` will automatically be converted from shape `(batch_size * num_codebooks, target_sequence_length)` to `(batch_size, num_codebooks, target_sequence_length)` in the forward pass. If you obtain audio codes from an audio encoding model, such as [`EncodecModel`], ensure that the number of frames is equal to 1, and that you reshape the audio codes from `(frames, batch_size, num_codebooks, target_sequence_length)` to `(batch_size * num_codebooks, target_sequence_length)` prior to passing them as `decoder_input_ids`. decoder_attention_mask (`torch.LongTensor` of shape `(batch_size, target_sequence_length)`, *optional*): Default behavior: generate a tensor that ignores pad tokens in `decoder_input_ids`. Causal mask will also be used by default. labels (`torch.LongTensor` of shape `(batch_size, sequence_length, num_codebooks)`, *optional*): Labels for language modeling. Note that the labels **are shifted** inside the model, i.e. you can set `labels = input_ids` Indices are selected in `[-100, 0, ..., config.vocab_size]` All labels set to `-100` are ignored (masked), the loss is only computed for labels in `[0, ..., config.vocab_size]` Examples: ```python >>> from transformers import AutoProcessor, MusicgenForConditionalGeneration >>> import torch >>> processor = AutoProcessor.from_pretrained("facebook/musicgen-small") >>> model = MusicgenForConditionalGeneration.from_pretrained("facebook/musicgen-small") >>> inputs = processor( ... text=["80s pop track with bassy drums and synth", "90s rock song with loud guitars and heavy drums"], ... padding=True, ... return_tensors="pt", ... ) >>> pad_token_id = model.generation_config.pad_token_id >>> decoder_input_ids = ( ... torch.ones((inputs.input_ids.shape[0] * model.decoder.num_codebooks, 1), dtype=torch.long) ... * pad_token_id ... ) >>> logits = model(**inputs, decoder_input_ids=decoder_input_ids).logits >>> logits.shape # (bsz * num_codebooks, tgt_len, vocab_size) torch.Size([8, 1, 2048]) ```rNrr)rBr3rr).N)rrr,0Expected 1 frame in the audio code outputs, got Y frames. Ensure chunking is disabled by setting `chunk_length=None` in the audio encoder.r(rjr.)rBr3rrrrrr@) rCrDrdecoder_hidden_statesdecoder_attentionsrencoder_last_hidden_staterencoder_attentionsr?)rrrarrr9rrrr%rrrPrCrDr audio_codesrKrMr\rr_rrrCrDrrrrr)rZrBr3rrrrr2rrrr@rrrrr common_kwargsrrraudio_encoder_outputsrframesrrrr.s rArz(MusicgenForConditionalGeneration.forwardsTX!!  ,,. +JC~~o.CH#CO(<(>$?@ 01EJ$S-=)>)@%AB +9>s3z?#456%* c" +  "/d//#-+&   O /-?O / 2    $ $ 0 0DLL4G4G4S4S S ##??G$($8$89N$O !  %$9N9CXC` 2 ++88$++:M:M:d:d!  &+@+H$6D$6$6%))%'% ! 0;;K.9.?.? +FCG{ FvhOTT {{""11Q6;;L;LQ;OSWS_S_SmSmqrSr;r);;A1;E +F 3 ; ;C$,,B\B\<\^e f >JT\\ > '1"7#1/+ >  >  >  %%"))+;;"1"?"?.99,==&5&G&G"1"?"?.99  r@next_sequence_lengthc  ^|I|jj||jj|jj\}}|jj ||}| )| dkDr$|j d}||j d}|| |dd| dfn|}d||||||dS)N)rPr,rQ)rBr2rrr3rr)r%rRrSrCrPrTrU) rZrrrr3rrr2decoder_delay_pattern_maskr4rs rArWz>MusicgenForConditionalGeneration.prepare_inputs_for_generation6s & -<@LL[[F JJ T\\%?%??C5::^d e$ %  $ 7 !,.. '+AA F F H M M O % +BDU*V\^ _ '<7)56N)O&).__%;.I!x**1-Irrroutput_hidden_states return_dictr,rrjr3rr2r?) get_encoderrdrio_same_deviceranyrrrrrrrr4rmain_input_namerr: concatenate zeros_liker)rZrzrryrSencoderirrelevant_prefixrrencoder_kwargsencoder_signatureencoder_accepts_wildcardr4rs ` rA+_prepare_text_encoder_kwargs_for_generationzLMusicgenForConditionalGeneration._prepare_text_encoder_kwargs_for_generations""$ 7J '.2G   +D$0#5#5#7  %I7HII eO    1 1'// B M MN#+/@#@#gNVgDg '7E7K7K7M$3HeQY]nQn%N/@.Q.Q*+1B1W1W-.*990@/K+QUQbQbQrQr(,}%+8'(#5n5GG  %.1*< % 1 13DeFVFVWhFi2jpq r </161B1B!"23U5E5ElScFd5eflm2 -.+:L]*^ &'? s!F.< F4 F4c|jd}t|drd|j_gd}|j Dcic]\}t fd|Ds|}}}t tj|jj} d| vxsd| v} | s(|j Dcic] \}}|| vs ||}}}||n|jj}d|d <|jjjd k(rB|||<|j di|} | j"} | j$} | j&\}}}}n|j&d d k7rt)d |j&d d |dddd ddf||<|j di|}|j"}|j$}|ddd dddf||<|j di|}|j"}|j$}|j&\}}}}|j+||d |z|f} || ddddddd ddf<|| ddddd dd ddf<|dgk7s|dgk7rt-j.||gd } ndg|z} |d k7rt)d|d| dj1||jj2z|}||d<| |d<|Scc}}wcc}}w)Nr3)modalityrTrc3@K|]}j|ywrVrrs rArz`MusicgenForConditionalGeneration._prepare_audio_encoder_kwargs_for_generation..rrrrrr,r(z3Expected stereo audio (2-channels) but example has z channel.rjrrrr audio_scalesr?)rrdrrrrrrrrrrrr%rr\encoderrrKrMnew_onesr:rFrr)rZrrryrrrrrrr rrrrrrraudio_encoder_outputs_leftaudio_codes_leftaudio_scales_leftaudio_encoder_outputs_rightaudio_codes_rightaudio_scales_rightrs ` rA,_prepare_audio_encoder_kwargs_for_generationzMMusicgenForConditionalGeneration._prepare_audio_encoder_kwargs_for_generationsO""G"4 7J '.2G   +D$0#5#5#7  %I7HII eO    1 1'// B M MN#+/@#@#gNVgDg '7E7K7K7M$3HeQY]nQn%N 0@/K+QUQcQcQsQs(,}% <<   - - 2/;N+ ,$2GNN$D^$D !/;;K0==L.9.?.? +FCG!!!$) I,J\J\]^J_I``ij0*J ' ; G G !.D.D +FCG*33VS!i-QX4YZK(8K1cc1 %):K1addA & TF*.@TF.J${{,=?Q+RXYZ $v| Q;B6(KPP  (/77dll>X>X8XZab,= ()'3 ^$} s !K> K Kct||jjj|jjjSrV)rPrr%rCrE)rZr@s rA%prepare_decoder_input_ids_from_labelszFMusicgenForConditionalGeneration.prepare_decoder_input_ids_from_labelss1!&$++*=*=*J*JDKKL_L_LlLlmmr@ctd)NzResizing the embedding layers via the EncoderDecoderModel directly is not supported. Please use the respective methods of the wrapped objects (model.encoder.resize_token_embeddings(...) or model.decoder.resize_token_embeddings(...)))NotImplementedError)rZargsrs rAresize_token_embeddingsz8MusicgenForConditionalGeneration.resize_token_embeddingss! ;  r@cr|jjD] }d|_ d|j_y)z3 Freeze the audio encoder weights. FN)rr requires_grad_requires_gradrZparams rAfreeze_audio_encoderz5MusicgenForConditionalGeneration.freeze_audio_encoders8''224 (E"'E  (,1)r@cr|jjD] }d|_ d|j_y)z2 Freeze the text encoder weights. FN)rrrr r!s rAfreeze_text_encoderz4MusicgenForConditionalGeneration.freeze_text_encoder"s8&&113 (E"'E  (+0(r@rnc||S|jd}|I|djdd}tj|tj|j dzS| t dd}|jD]-}t|tjs|jd}ntj|dftj|j |zS) z3Initializes input ids for generation, if necessary.Nr2rrFr_rHzB`bos_token_id` has to be defined when no `input_ids` are provided.r,) rr|r:rZr[rarMrrrrK)rZrnrErr2rKr}rs rA*_maybe_initialize_input_ids_for_generationzKMusicgenForConditionalGeneration._maybe_initialize_input_ids_for_generation*s  M&**+<=  &#A&++-cr2E::e5::dkkJTQ Q  ab b !((* E%."[[^  zz:q/DKKPS___r@c||n|jj}||n|jj}||S||Std)Nz\`decoder_start_token_id` or `bos_token_id` has to be defined for encoder-decoder generation.)rSrDrErM)rZrDrEs rArz>  (4'?|TE[E[EhEh ! -) )  %  j  r@rorprqrrr/c ~ |jd|ddd}|j|fi|\}} |j} | tjtj fvr t d|j| j|j| ||| jd)t| dturt| dd| d<||n t}||n t}d| v} | jdddu} |j!||j"| \} }} | j$d}|j'|| | j(|j*| d <|j,| d <| jdd| r|j/| || | d<d| vr|j1| | ||} d | vrd | vr|j3| d | } |j5||| |j6|j8| j( \}} |j$d}|jdduxr|j:du}|jdduxr|j<du}|j?||||| |}|j@jC||j6|j:\}}|| d<||jE|jG|j} |j,:|j,dkDr+|jItK|j,d|_|jM||| d||j(}|jO||}|jPd||jR|jTjVd| \}} |jY||| |d<|jZ|f|||d|| }|j\r |j^}n|}|j@ja|| d}|||jbk7je||j@jfd}|d}| jd}|dg|z}|j@jTjhdk(r(|jjjm||jn}n|jjjm|dddddddddf|}|jn}|jjjm|dddddddddf|}|jn}tqjr||gd}|j\r ||_/|S|S) rtNFrr2rrr3rBrr4rr)r}ryrrDrErarFrPrurvrOrr,rr)rBrrprefill_outputs)rorprS)N.r)rr(rjr?):_extract_generation_mode_kwargs_prepare_generation_configrrrrrMrr_validate_generation_modertyper9rrrrrErKrrarr4rr rrr_bos_token_tensorrPrurr%rRrrrrrrrrrr_prefillrrrrTrrrr\rdecode audio_valuesr:rr)rZrnrSrorprqrrrgeneration_mode_kwargsrr|rrrzryr}rBr{rwrxrrHrr output_valuescodec_outputs_leftoutput_values_leftcodec_outputs_rightoutput_values_rights rArz)MusicgenForConditionalGeneration.generateXsV"&!E!EdFTY[_ae!f*I$*I*IJ[*f_e*f'<+??A >#8#8.:V:V"W WT  ##L$5$5$78 &&8IKab   - . :tLQbDc?dhm?m.=P\]nPopqPr.sL* +0@/K+QdQf1B1N-ThTj"3<"G$0$4$45Et$LTX$X!9=8R8R %22L9 5 '#((+  $$%68QZgZnZn$o%6$?$? [!):)I)I %&   ,d 3 ;@W-1-X-X0,.L) * L 0KK|-=?PL l 2~7ULL^,L #'"P"P!-%#4#P#P*<< '' #Q#  <%??2.!'L!9T!A!nFWFbFbjnFn!'L!9T!A!nFWFbFbjnFn ::/#9#9-'- ; 15 0U0U *FF(331V1 - - 6P 12   LL ),??A  + + 7>#{{==#  #  <59MM)M^`l4m01$,,  -//  %      4 4 **J J\\:::|TpGqr  .?.Q.Q QRZZ  22B  * #''7   6J.L <<   - - 2 ..55)6l  "&!3!3!:!::aCaCQRl;Sbn!:!o !3!@!@ "&"4"4";";Jq!QTPQTST}>> from transformers import MusicgenForConditionalGeneration >>> model = MusicgenForConditionalGeneration.from_pretrained("facebook/musicgen-small") >>> # get the unconditional (or 'null') inputs for the model >>> unconditional_inputs = model.get_unconditional_inputs(num_samples=1) >>> audio_samples = model.generate(**unconditional_inputs, max_new_tokens=256) ```r,)rar`r)r2r3r4) r:rvrrrrar`r[r1)rZ num_samplesrr3s rAget_unconditional_inputsz9MusicgenForConditionalGeneration.get_unconditional_inputsNsy,"KK !T[[55AA B4;;^b^h^h k1%5dkkQVQ[Q[\).0)  r@)NNNN)NNN) NNNNNNNNNNNNrrV)NNr)r,)7r5r6r7r-r<rrrrrr1rXr)r;r? classmethodstrrr$r"r:r= BoolTensorr;r9r rr r!rrrrWdictrrarrrr rrrr#r%r'listrrrrrrr;rrs@rArrQs ")!O&*#)-/304.2 e%e&,e'- e %t+ eN84BBFBF<@ Bk47$JBk694ZBk03Tz Bk BkBkH.226150459:>;?(,26:>*.!%` ##d*` ((4/` ''$. ` &&- ` !++d2 ` !& 0 04 7` u001D8` ` ((4/`  %0047`   4'` $;` +,`   ` ` J,0##', "Dj, f.2#'&*-/-/-/3 ,- -/ !$d -/ Dj -/ t#-/ uc5<<&7!88 9-/^/||/* / , / c3h /dJNJ*s!s!j  r@r)rr1r#r)Nr)Yr8rrrlrcollections.abcr dataclassesrtypingrrrr:torch.nnrrr r activationsr cache_utilsr r r generationrrrrrr masking_utilsrrmodeling_flash_attention_utilsrmodeling_layersrmodeling_outputsrrrrrmodeling_utilsrrprocessing_utilsr utilsr!r"r# utils.genericr$r%utils.output_capturingr&r'auto.configuration_autor)rr+configuration_musicgenr-r.generation.streamersr/ get_loggerr5rr1rrrPModulerRr>rrrrrr#r1r__all__r?r@rArXs2  $!// %&!CCK:G&@@IE0*I4   H % (( ("%,,c[^()LBII)Lf!% II% <<% % << % LL4' % T\ %%'(%8m) m)`\5\~4o44"O -O dE+EEP Z1?Z  Zz X '>X  X v rr@