T^(h:dZddlmZddlmZmZmZmZmZm Z ddl Z ddl m cm ZddlZ ddl m Z ddlmZddlmZdd lmZmZmZdd lmZdd lmZdd lmZdd lmZm Z ddl!m"Z"ddl#m$Z$m%Z%m&Z&m'Z'm(Z(ddl)m*Z*ddl+m,Z,ddl-m.Z.e&r ddl/m0Z0ddl1m2Z2e'jfe4Z5dZ6eGddeZ7eGddeZ8 d9dZ9gfdZ:Gdde jvZ<Gdd e jzZ>Gd!d"e j~Z@e"je@Gd#d$e jj~ZBd%ZCd:d&ZDGd'd(e j~ZEGd)d*e j~ZFGd+d,e j~ZGGd-d.e j~ZHd/ZIe$d0eIGd1d2e ZJd3ZKe$d0eIGd4d5eJZLGd6d7eJeZMgd8ZNy);zPyTorch Idefics model.) dataclass)AnyDictListOptionalTupleUnionN)nn)CrossEntropyLoss)ACT2FN)Cache DynamicCache StaticCache)GenerationMixin)AttentionMaskConverter) ModelOutput)PretrainedConfigPreTrainedModel)ALL_LAYERNORM_LAYERS)add_start_docstrings%add_start_docstrings_to_model_forwardis_torch_flex_attn_availableloggingreplace_return_docstrings) IdeficsConfig)IdeficsPerceiverResampler)IdeficsVisionTransformer) BlockMask)make_flex_block_causal_maskrceZdZUdZdZeejed<dZ ee e ejed<dZ ee ejed<dZ ee ejed<dZ ee ejed<y)IdeficsBaseModelOutputWithPasta Base class for Idefics model's outputs that may also contain a past key/values (to speed up sequential decoding). Args: last_hidden_state (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`): Sequence of hidden-states at the output of the last layer of the model. If `past_key_values` is used only the last hidden-state of the sequences of shape `(batch_size, 1, hidden_size)` is output. past_key_values (`tuple(tuple(torch.FloatTensor))`, *optional*, returned when `use_cache=True` is passed or when `config.use_cache=True`): Tuple of `tuple(torch.FloatTensor)` of length `config.n_layers`, with each tuple having 2 tensors of shape `(batch_size, num_heads, sequence_length, embed_size_per_head)`) and optionally if `config.is_encoder_decoder=True` 2 additional tensors of shape `(batch_size, num_heads, encoder_sequence_length, embed_size_per_head)`. Contains pre-computed hidden-states (key and values in the self-attention blocks and optionally if `config.is_encoder_decoder=True` in the cross-attention blocks) that can be used (see `past_key_values` input) to speed up sequential decoding. 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. 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)`. Attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads. image_hidden_states (`tuple(torch.FloatTensor)`, *optional*): Tuple of `torch.FloatTensor` (one for the output of the image embeddings, `(batch_size, num_images, sequence_length, hidden_size)`. image_hidden_states of the model produced by the vision encoder, and optionally by the perceiver Nlast_hidden_statepast_key_values hidden_states attentionsimage_hidden_states)__name__ __module__ __qualname____doc__r$rtorch FloatTensor__annotations__r%rr&r'r(j/var/www/skyplay_api_hub/venv/lib/python3.12/site-packages/transformers/models/idefics/modeling_idefics.pyr#r#=s"H6:x 1 129AEOXeE%*;*;$<=>E8B%(9(9":;Br1r#ceZdZUdZdZeejed<dZ eejed<dZ ee ejed<dZ ee ejed<dZee ejed<dZee ejed<y) IdeficsCausalLMOutputWithPasta Base class for Idefics causal language model (or autoregressive) outputs. Args: 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 (`tuple(tuple(torch.FloatTensor))`, *optional*, returned when `use_cache=True` is passed or when `config.use_cache=True`): Tuple of `tuple(torch.FloatTensor)` of length `config.n_layers`, with each tuple having 2 tensors of shape `(batch_size, num_heads, sequence_length, embed_size_per_head)`) 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. 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. 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)`. Attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads. image_hidden_states (`tuple(torch.FloatTensor)`, *optional*): Tuple of `torch.FloatTensor` (one for the output of the image embeddings, `(batch_size, num_images, sequence_length, hidden_size)`. image_hidden_states of the model produced by the vision encoder, and optionally by the perceiver Nlosslogitsr%r&r'r()r)r*r+r,r5rr-r.r/r6r%rr&rr'r(r0r1r2r4r4js@)-D(5$$ %,*.FHU&& '.9=OXd5#4#456=8B%(9(9":;Br1r4c tj|jdjddj d|jdj |j }|jd|}|jdd|d<|jdd|d<|jdd|d<|jdd|d<d|vr|d}|jd||d<||jd||d <|d|djd||d<|d|djd||d<||fS|d|djd||d<||fS|d|djd||d<||fS) Nrr pixel_valuesimage_encoder_embeddingsperceiver_embeddingsimage_attention_masktoken_type_idsattention_mask) r-arangeshapeviewrepeattodevice index_selectget) input_ids expand_sizeis_encoder_decoderr>encoder_outputs model_kwargsexpanded_return_idxr=s r2expand_inputs_for_generationrMs  Y__Q'(--b!4;;A{KPPQSTWWXaXhXhi&&q*=>I#/#3#3ND#IL /;/?/?@Z\`/aL+,+7+;+;zfreeze_model..s$]qZ%:$]sTF)r rOrPrQmodulesanyrequires_grad_)modelmodule_exceptionsmappingmmodule_exceptions_mappedrWs @r2 freeze_modelras\\))\\G 5FFq FF--/) $]D\$]!]  ! !$ '  ! !% ( ) L Gs BcNeZdZdZ ddeeddffd ZdZdefdZ xZ S) IdeficsDecoupledEmbeddinga Implements a decoupling of parameters to allow freezing (or not) a subset of the embeddings. In practise, the regular `weight` can be trained or frozen (i.e. `partially_freeze=True`), and if `num_additional_embeddings` > 0, then it will create `num_additional_embeddings` additional parameters that are always trained. If `num_additional_embeddings=0`, then the module defaults back to the regular behavior of `nn.Embedding`. Npartially_freezereturnc B|||kDrtd|d|t |d|||||d|||_||_||_||_|r|jjd|j dkDr)tj|j ||||_ yy) a) Args: num_embeddings (`int`): Size of the dictionary of embeddings num_additional_embeddings (`int`): Number of additional embeddings. Only useful when you `partially_freeze=True`. embedding_dim (`int`): The size of each embedding vector partially_freeze: (`bool`, *optional*, defaults to `False`): If `True`, the regular `weight` will be frozen. `additional_weight` is never frozen. padding_idx (`int`, *optional*): The padding index (needs to be less than num_embeddings) Note: there are a lot of other parameters to initialize a standard `nn.Embedding` such as `padding_idx`, `max_norm` or `norm_type`. We are not supporting these. Nz/padding_idx must be within num_embeddings. Got z and )num_embeddings embedding_dimrDdtype padding_idxFr)rgrhrDrir0) ValueErrorsuper__init__rgrjnum_additional_embeddingsrdweightr[r rQadditional_embedding) selfrgrnrhrdrDrirjkwargs __class__s r2rmz"IdeficsDecoupledEmbedding.__init__s6  "{^'CN{m[`ao`pqr r  )'#    -&)B& 0  KK & &u -  ) )A -(* #==+ )D % .r1cb|jdk(r tj||jS|j }t j ||jk\}||}|j||jz }d||<tj||j}|||<|S)a we have 2 embeddings, with different indices - one pretrained self.weight and another self.additional_embedding.weight that is being trained. in order to make a lookup of the input ids, we: 1. find out the indices of the entries belonging to the 2nd embedding 2. extract those values while subtracting the size of the first embedding (num_embeddings), since the 2nd embedding starts from 0 and not num_embeddings 3. perform the 2nd embedding lookup 4. now we handle the 1st embedding, we overwrite indices belonging to the 2nd embedding with a padding index 5. perform the 1st embedding lookup 6. now we overwrite the values in the 1st embedding lookup with the values of the 2nd embedding lookup note: for the 1st embedding lookup we could have looked up only the low indices and not do the padding, but then we have to create a new tensor and populate it with 2 tensors that are spread out across various indices - i.e. not a simple concat - I haven't benchmarked the complex case if it's any faster, given that seqlens are usually relatively short it's probably not faster or if faster not by much - but might be a good idea to measure. r) rnF embeddingrocloner-wherergrp)rqrGadditional_vocab_indicesinput_ids_additional_vocabadditional_embeddings full_vectors r2forwardz!IdeficsDecoupledEmbedding.forwards*  ) )Q .;;y$++6 6OO% #(;;yD 0, then it will create `out_additional_features * in_features` additional parameters that are always trained. If `out_additional_features=0`, then the module defaults back to the regular behavior of `nn.Linear`. N in_features out_featuresout_additional_featuresbiasrdrec"t||||||||_||_||_||_|r8|j jd|r|jjd|dkDr tj||||||_ yy)aG out_additional_features: int. Number of additional trainable dimensions. Only makes sense when `partially_freeze=True`. partially_freeze: bool. If True, the regular `weight` will be frozen and extra parameters (if any) will be trainable. If False, default to the regular behavior of nn.Linear. Fr)rrrrDriN) rlrmrrdrrror[rr rP additional_fc) rqrrrrrdrDrirss r2rmzIdeficsDecoupledLinear.__init__Gs lD&%H'>$ 0&(  KK & &u - ((/ "Q &!#'4 "D  'r1inputctj||j|j}|jdkDr)|j |}t j||fd}|S)Nrr8)rulinearrorrrr-cat)rqroutputadditional_featuress r2r}zIdeficsDecoupledLinear.forwardksV%dii8  ' '! +"&"4"4U"; YY(;s()!%""""% "  "  " "HU\\ell C r1rc,eZdZdfd ZdZdZxZS)IdeficsRMSNormct|tjt j ||_||_y)z= IdeficsRMSNorm is equivalent to T5LayerNorm N)rlrmr Parameterr-onesrovariance_epsilon)rq hidden_sizeepsrss r2rmzIdeficsRMSNorm.__init__s1 ll5::k#:; #r1c|jtjjdj dd}|tj ||j zz}|jjtjtjfvr%|j|jj}|j|zS)Nr8T)keepdim) rCr-float32powmeanrsqrtrrorifloat16bfloat16)rqr&variances r2r}zIdeficsRMSNorm.forwards ##EMM266q9>>r4>P% Ht?T?T4T(UU  ;;   ? ?),,T[[->->?M{{]**r1c^t|jjd|jS)Nz, eps=)tupleror@rrs r2rzIdeficsRMSNorm.extra_reprs* ))*+6$2G2G1HIIr1)gư>)r)r*r+rmr}rrrs@r2rrs$+Jr1rc.eZdZdfd ZdZddZxZS)IdeficsEmbeddingct|||_||_||_d|jt j d|jdt jj|t j|jz zz }|jd|d|j||jjt j y) Ng?rrrirDriinv_freqF persistentseq_lenrDri)rlrmdimmax_position_embeddingsbaser-r?int64rCfloatregister_buffer_set_cos_sin_cacherrDget_default_dtype)rqrrrrDrrss r2rmzIdeficsEmbedding.__init__s '>$  II Q!5;;?BB&X]XcXcBdgkgogoo q  ZeD +DMM4H4HPUPgPgPi  r1c||_tj|j|tjj |j }tj d||j }tj||fd}|jd|jj|d|jd|jj|dy) Nrzi,j->ijr8r cos_cachedFr sin_cached) max_seq_len_cachedr-r?rtype_asreinsumrrcosrCsin)rqrrDrirVfreqsembs r2rz#IdeficsEmbedding._set_cos_sin_caches") LL00u{{ S [ [\`\i\i j Y4==9iiB/ \3779<<+>5Q \3779<<+>5Qr1c||jkDr(|j||j|j|jd|j |j|j d|j |jfS)Nrr)rrrDrirrCr)rqxrs r2r}zIdeficsEmbedding.forwardsy T,, ,  # #GAHHAGG # T OOHW % ( (qww ( 7 OOHW % ( (qww ( 7  r1)ii'NrS)r)r*r+rmrr}rrs@r2rrs "R r1rc|dd|jddzf}|d|jddzdf}tj| |fdS)z*Rotates half the hidden dims of the input..Nr8rr)r@r-r)rx1x2s r2 rotate_halfrsZ 3"!''"+"" " #B 3 q " " #B 99rc2YB ''r1c||j|}||j|}||zt||zz}||zt||zz}||fS)anApplies Rotary Position Embedding to the query and key tensors. Args: q (`torch.Tensor`): The query tensor. k (`torch.Tensor`): The key tensor. cos (`torch.Tensor`): The cosine part of the rotary embedding. sin (`torch.Tensor`): The sine part of the rotary embedding. position_ids (`torch.Tensor`): The position indices of the tokens corresponding to the query and key tensors. For example, this can be used to pass offsetted position ids when working with a KV-cache. unsqueeze_dim (`int`, *optional*, defaults to 1): The 'unsqueeze_dim' argument specifies the dimension along which to unsqueeze cos[position_ids] and sin[position_ids] so that they can be properly broadcasted to the dimensions of q and k. For example, note that cos[position_ids] and sin[position_ids] have the shape [batch_size, seq_len, head_dim]. Then, if q and k have the shape [batch_size, heads, seq_len, head_dim], then setting unsqueeze_dim=1 makes cos[position_ids] and sin[position_ids] broadcastable to the shapes of q and k. Similarly, if q and k have the shape [batch_size, seq_len, heads, head_dim], then set unsqueeze_dim=2. Returns: `tuple(torch.Tensor)` comprising of the query and key tensors rotated using the Rotary Position Embedding. ) unsqueezer)qkrr position_ids unsqueeze_dimq_embedk_embeds r2apply_rotary_pos_embrsg* l  % %m 4C l  % %m 4C3w;q>C/0G3w;q>C/0G G r1c2eZdZdededeffd ZdZxZS) IdeficsMLPrintermediate_size hidden_actct|tj||d|_tj||d|_tj||d|_t||_y)NFr) rlrmr rP gate_proj down_projup_projr act_fn)rqrrrrss r2rmzIdeficsMLP.__init__s[ ;0AN#4kNyy.?eL Z( r1c|j|j|j||j|zSrS)rrrr)rqrs r2r}zIdeficsMLP.forwards0~~dkk$..*;rpast_key_valueoutput_attentions use_cachecache_positionrec |jxs|du} |j\} } } |j|j| | |j|j j dd} | s|j|j| | |j|j j dd}|j|j| | |j|j j dd}n|j\} }} |j|j| ||j|j j dd}|j|j| ||j|j j dd}|jd}|||dz }| s2|j|t|| \}}t| ||||\} }|%d|i}|j|||j|\}}|jr"|j!| } |j#|}|}||ddddddd|jdf}| j$j&dk(r2|0| j)} |j)}|j)}|j*r || dkDrdnd }t,j.j0j3| ||||j4r |j6nd | }|j| |j| |j fk7r7t9d | |j| |j fd |j|j dd}|j;| | |j<}|j?|}d}|rt@jCd|||fS)Nrrr)rrcudaTF) attn_mask dropout_prz `attn_output` should be of size z , but is z^attn_weights are not extracted in scaled_dot_product_attention. The model returns None instead)"rsizerrArrrrrr@rmaxrupdaterrrrrDtyperrr-r rrtrainingrrkreshaperrrr)rqr&rr>rrrrrrrq_len_ query_states key_states value_stateskv_len kv_seq_lenrr cache_kwargs causal_maskr attn_output attn_weightss r2r}zIdeficsAttention.forwardMs"44T8HPT8T%**, UA{{=166sE4>>SWS`S`akklmopq !]388eT^^UYUbUbcmmnoqrsJ;;}5::3t~~W[WdWdeoopqstuL+002LAvq%56;;CY]YfYfgqqrsuvwJ ,-223PTP]P]^hhijlmn  %%b)  % .+ +J!|SU=STHC';L*VY[^`l'm $L*  %,n=L'5'<'Q5RR SW9r1rcJeZdZd dedeeffd Z ddejdeejdeejdee ejdee d ee d eejd e ejee ejejfffd Z xZS)IdeficsDecoderLayerrrct||j|_t|j|j|j |||_t|j|j|j|_ t|j|j|_ t|j|j|_|j |_y)N)rrrrrrrrr)rlrmrrnum_attention_headsr self_attnrrrmlprrinput_layernormpost_attention_layernormrqrrrss r2rmzIdeficsDecoderLayer.__init__s !--)((00NN  (($66((  .f.@.@fFYFYZ(6v7I7IvObOb(c%~~ r1r&r>rrrrrrec |}|j|}|j|||||||\}} } tjj ||j|j }||z}|}|j |}|j|}tjj ||j|j }||z}|f} |r| | fz } |r| | fz } | S)a] Args: hidden_states (`torch.FloatTensor`): input to the layer of shape `(batch, seq_len, embed_dim)` attention_mask (`torch.FloatTensor`, *optional*): 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. use_cache (`bool`, *optional*): If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding (see `past_key_values`). past_key_value (`Tuple(torch.FloatTensor)`, *optional*): cached past key and value projection states )r&r>rrrrrpr)r)r'r rrrr*r() rqr&r>rrrrrresidualself_attn_weightspresent_key_valueoutputss r2r}zIdeficsDecoderLayer.forwards0!,,]; ?Cnn')%)/)?M? ; (*; --mt||VZVcVc-d  =0 !55mD /  --mt||VZVcVc-d  =0  "  )+ +G  )+ +Gr1rS)NNNFFN)r)r*r+rrrrmr-rr!rrr.r}rrs@r2r#r#s&}&#&,26378<,1$)598||8!.8u//0 8 !u||!45 8 $D> 8D>8!!1!128 u  (51B1BEDUDU1U+V"WW X8r1r#cjeZdZddedeeffd Z ddejdeejdeejdeejdeejd ee d ee d ee ejd e ejee ejejfffd Z xZ S)IdeficsGatedCrossAttentionLayerrrc  t||j|_t|j|jd|j ||j ||_t|j|j|j|_ t|j|j|_t|j|j|_|j |_t#j$|_t#j$|_|j*dk(r|j,dk(rtt#j.t1j2dd|j|_t#j.t1j2dd|j|_n|j,dk(r\t#j.t1j2d|_t#j.t1j2d|_nt9d |j,d |j*d k(r|j,dk(rtt#j.t1j:dd|j|_t#j.t1j:dd|j|_n|j,dk(r\t#j.t1j:d|_t#j.t1j:d|_n|t9d |j,d |j*d vr;|j,dk(rt#j.t1j<d |j>dd|jf|_t#j.t1j<d |j>dd|jf|_n|j,dk(rut#j.t1j<d |j>d|_t#j.t1j<d |j>d|_n2t9d |j,d tAd|j*dtC|dr tC|ds t9dy)NT)rrrrrrrr%rzerosvectorrrz Unknown value for `alpha_type` ()r>normalrandomgaussianr )rstdrzAlpha initialization scheme z not yet implemented!alpha_cross_attn alpha_densez+Alpha parameters not initialized correctly!)"rlrmrrr&rr cross_attnrrrr(rrr)r*rr Tanhact_cross_attn act_densealpha_initializer alpha_typerr-r6r=r>rkrr9alphas_initializer_rangeNotImplementedErrorrr+s r2rmz(IdeficsGatedCrossAttentionLayer.__init__s !--*((00#NN!00 (($66((  .f.@.@fFYFYZ(6v7I7IvObOb(c%nn  ggi  # #w .  H,(* U[[AtGWGW5X(Y%#%<< Aq$BRBR0S#T ""g-(* U[[^(D%#%<< A#?  #CFDUDUCVVW!XYY  % % /  H,(* UZZ1dFVFV5W(X%#%<< 1aAQAQ0R#S ""g-(* UZZ](C%#%<< 1 #>  #CFDUDUCVVW!XYY  % %)I I  H,(* LLcv/N/NVWYZ\`\l\lUmn)%$&<<LLcv/N/NVWYZ\`\l\lUmn$ ""g-(* LLcv/N/NVWY)%$&<< #6KjKjrs0u#v  #CFDUDUCVVW!XYY&(DVE]E]D^^s&tu u01gdM6RJK K7Sr1r&r>r(r<cross_attention_gaterrrrec | td| td| td|} |j|}|j||||\}} } tj j ||j|j}|j|dk(dddddfd}| |j|j|zz}|} |j|}|j|}tj j ||j|j}| |j|j|zz}|f} |r| | fz } |r| | fz } | S) a Args: hidden_states (`torch.FloatTensor`): input to the layer of shape `(batch, seq_len, embed_dim)` attention_mask (`torch.FloatTensor`, *optional*): attention mask of size `(batch, 1, tgt_len, src_len)` where padding elements are indicated by very large negative values. image_attention_mask (`torch.FloatTensor`, *optional*): image attention mask of size `(batch, 1, tgt_len, src_len)` where padding elements are indicated by very large negative values. cross_attention_gate (`torch.FloatTensor`, *optional*): gate of size `(batch, seq_len)` used to zero-out cross-attention output for tokens attending no images. output_attentions (`bool`, *optional*): Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned tensors for more detail. use_cache (`bool`, *optional*): If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding (see `past_key_values`). past_key_value (`Tuple(torch.FloatTensor)`, *optional*): cached past key and value projection states Nzt`image_hidden_states` is required for Idefics cross attention module which are visual features to be conditioned on.z`cross_attention_gate` is required for Idefics cross attention module to zero-out the cross-attention hidden_states attending to no images.zMPast key value states are not implemented for Idefics cross attention module.)r&rr>rr-rr )rkrFr)r?r rrrr masked_fillrAr=r*r(rBr>) rqr&r>r(r<rGrrrr/r0r1r2s r2r}z'IdeficsGatedCrossAttentionLayer.forward:s8  &#  '^   %%&uv v ,,]; ?Coo'0// ?N? ; (*;  --mt{{UYUbUb-c %113G13LaQRTXj2Y[^_  4#6#6t7L7L#MP]#]] !55mD /  --mt{{UYUbUb-c  4>>$2B2B#Cm#SS  "  )+ +G  )+ +Gr1rSr )r)r*r+rrrrmr-rrrr.r}rrs@r2r4r4s@L}@L#@LJ266:7;7;,1$)8<I||I!.I&ell3 I 'u||4 I 'u||4 I$D>ID>I!u||!45I u  (51B1BEDUDU1U+V"WW XIr1r4aL This model inherits from [`PreTrainedModel`]. Check the superclass documentation for the generic methods the library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads etc.) This model is also a PyTorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#torch.nn.Module) subclass. Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage and behavior. Parameters: config ([`IdeficsConfig`]): Model configuration class with all the parameters of the model. Initializing with a config file does not load the weights associated with the model, only the configuration. Check out the [`~PreTrainedModel.from_pretrained`] method to load the model weights. zSThe bare LLaMA Model outputting raw hidden-states without any specific head on top.c2eZdZeZdZdZddgZdZdZ dZ dZ y)IdeficsPreTrainedModelr\Tr#r4Fc|jj}t|tjrY|j j jd||j%|jj jyyt|tjrf|j j jd||j2|j j |jjyyy)Nr )rr<) rinitializer_rangerTr rProdatanormal_rzero_rQrj)rqrWr<s r2 _init_weightsz$IdeficsPreTrainedModel._init_weightsskk++ fbii ( MM   & &CS & 9{{&   &&('  - MM   & &CS & 9!!- ""6#5#56<<>..r1N) r)r*r+r config_classbase_model_prefixsupports_gradient_checkpointing_no_split_modules_supports_sdpa_supports_cache_class_supports_static_cacherQr0r1r2rKrKs7 !L&*#.0QRN " ?r1rKa Args: input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`): Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you provide it. Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and [`PreTrainedTokenizer.__call__`] for details. [What are input IDs?](../glossary#input-ids) attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*): Mask to avoid performing attention on padding token indices. 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) Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and [`PreTrainedTokenizer.__call__`] for details. If `past_key_values` is used, optionally only the last `decoder_input_ids` have to be input (see `past_key_values`). If you want to change padding behavior, you should read [`modeling_opt._prepare_decoder_attention_mask`] and modify to your needs. See diagram 1 in [the paper](https://arxiv.org/abs/1910.13461) for more information on the default strategy. - 1 indicates the head is **not masked**, - 0 indicates the head is **masked**. position_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*): Indices of positions of each input sequence tokens in the position embeddings. Selected in the range `[0, config.n_positions - 1]`. [What are position IDs?](../glossary#position-ids) past_key_values (`tuple(tuple(torch.FloatTensor))`, *optional*, returned when `use_cache=True` is passed or when `config.use_cache=True`): Tuple of `tuple(torch.FloatTensor)` of length `config.n_layers`, with each tuple having 2 tensors of shape `(batch_size, num_heads, sequence_length, embed_size_per_head)`) and 2 additional tensors of shape `(batch_size, num_heads, encoder_sequence_length, embed_size_per_head)`. Contains pre-computed hidden-states (key and values in the self-attention blocks and in the cross-attention blocks) that can be used (see `past_key_values` input) to speed up sequential decoding. If `past_key_values` are used, the user can optionally input only the last `decoder_input_ids` (those that don't have their past key value states given to this model) of shape `(batch_size, 1)` instead of all `decoder_input_ids` of shape `(batch_size, sequence_length)`. inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*): Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation. This is useful if you want more control over how to convert `input_ids` indices into associated vectors than the model's internal embedding lookup matrix. use_cache (`bool`, *optional*): If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding (see `past_key_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. output_hidden_states (`bool`, *optional*): Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for more detail. return_dict (`bool`, *optional*): Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple. cache_position (`torch.LongTensor` of shape `(sequence_length)`, *optional*): Indices depicting the position of the input sequence tokens in the sequence. Contrarily to `position_ids`, this tensor is not affected by padding. It is used to update the cache in the correct position and to infer the complete sequence length. c$eZdZdZdeffd Zd"dZgfdZgfdZdZ dZ e e  d#d e ejd e ej d e ejd e eej$d e ej$de ej$de ej$de ej$de ej de ede ede ede ede ede ejdeeeff dZ d$d ej dej dej d edef dZed ej dededej8dej:dej d efd!ZxZS)% IdeficsModelz Transformer decoder consisting of `config.num_hidden_layers` layers. Each layer is a [`IdeficsDecoderLayer`] Args: config: IdeficsConfig rc jt||||_|j|_|j |_t |j |j|j|j|j|_ |jj|_ |j|_ t|j|_|jr]|j }t#||jj$|j&|j(|j*|j,|_t1j2t5|j6Dcgc]}t9||c}|_|j<|_|j6|j<z}t1j2t5|Dcgc]}t?||c}|_ d|_!tE|j|jF|_$|jK|jM|ycc}wcc}w)N)rgrnrhrdrj)rFr)'rlrmr pad_token_idrj vocab_sizercadditional_vocab_sizerfreeze_text_layers embed_tokensr image_sizer vision_model use_resamplerperceiver_configrrresampler_depthresampler_n_headsresampler_head_dimresampler_n_latentsperceiver_resamplerr ModuleListrangenum_hidden_layersr#layerscross_layer_intervalr4gated_cross_attn_layersgradient_checkpointingrrnorm post_initfreeze_relevant_params)rqrrdinum_cross_layersrss r2rmzIdeficsModel.__init__s   !.. ++5!,,&,&B&B ,,#66((  !..99#114V5I5IJ   %66 '@$$.. 00 22 33 44 (D $mm?DVE]E]?^ _! 1 5 _  %+$?$?!!33t7P7PP')}}KPQaKb ca ,Vq A c( $',#"6#5#56;N;NO   ##F+ ` ds H+<H0c| |j}|jr|j|j|jr"t |j |j yyN)r])rr_freeze_text_module_exceptionsfreeze_vision_layersrarbfreeze_vision_module_exceptions)rqrs r2rsz#IdeficsModel.freeze_relevant_params1sQ >[[F  $ $  # #F$H$H I  & & **f>d>d e 'r1cX|j|jfD]}t||yrw)rmrqra)rqr]rWs r2r_zIdeficsModel.freeze_text_layers;s+{{DII. FF 3D E Fr1c2t|j|yrw)rarb)rqr]s r2ryz!IdeficsModel.freeze_vision_layers?sT&&:KLr1c|jSrSr`rs r2get_input_embeddingsz!IdeficsModel.get_input_embeddingsBs   r1c||_yrSr~rqvalues r2set_input_embeddingsz!IdeficsModel.set_input_embeddingsEs !r1rGr>rr% inputs_embedsr9r:r;r<rroutput_hidden_statesinterpolate_pos_encoding return_dictrrec | |jn |j}| | n|jj} | | n|jj} | | n|jj} ||n|jj }|du|duz r t d|jr%|jr| rtjdd} ||j|}d}| rIt|ts9d}| t}n*tj|}tjd|j \}}}||j#nd}||z}|2t%j&|||j dz|j}|F|D|j)j+d dz }|j-|dk(d|dd| df}n||j/d}t1|||fDcgc]}|duc}d k7r t d |~|j3|j4| }|j dd \}}|j7j8||zg|j d d}|j;|| j<}nJ|H|j?\}}}}|j3|j4| }|j9||z||}|jj@rN|4|jC}|j?d|j?d }}n|j?\}}}}|}n0|#j?d|j?d }}n t d|j9||z|}| j?d}| j/d } | jEddd|} | j9||||z} |C|j?\}}}||f}| t%jF||} |jI| } nd} | dk(jKd j3|j4jMdj3|} |2t%jF||ft$jN|j }|jQ||||| }|}!| rdnd}"| rdnd}#d}$tS|jTD]\}%}&| r|"|!fz }"d}'|jr[|jrOd}| rtjdd} |jW|'|&|!||||| | | | |%|jX|jZ|}(n*|'|&|!||||| | | | |%|jX|jZ|}(|(d}!| r |(| rd nd}$| s|#|(dfz }#|j]|!}!| r|"|!fz }"| r|$nd})|r|)j_})|j9||||}|stad|!|)|"|#|fDStc|!|)|"|#|Scc}w)Nz:You must specify exactly one of input_ids or inputs_embedszX`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`.FTzWe detected that you are passing `past_key_values` as a tuple of tuples. This is deprecated and will be removed in v4.47. Please convert your cache or use an appropriate `Cache` class (https://huggingface.co/docs/transformers/kv_cache#legacy-cache-format)rrrDr8rz_Exactly 1 of pixel_values, image_encoder_embeddings or perceiver_embeddings has to be not-None.)rirD)r9rzBIf `perceiver_embeddings` are passed, use_resampler should be Truer rrr0c n| | zdk(r| | | z}|||||||| d}|d}||||||| | }|S)Nr)r>r(r<rGrrr)r>rrrrrr0) main_blockr&r>rrr(r<rGrrrrnrorxblockr2 layer_outputss r2vblockz$IdeficsModel.forward..vblocksv"33q84YBV5VWF$%'5,?-A-A*;"+'+ G%,AJM *!#1!-#1&7'#1! %$r1zZ`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`...) r>rrr(r<rGrrrrnrorc3$K|]}|| ywrSr0)rUvs r2rXz'IdeficsModel.forward..Is=s)r$r%r&r'r()2rDrrrruse_return_dictrkrprrrr`rTrrfrom_legacy_cacher@get_seq_lengthr-r?longcumsum masked_fill_rsumrCrirrArbr$rrcrirBrinvert_attention_maskrZsqueezer_update_causal_mask enumeraterm_gradient_checkpointing_funcrnrorqto_legacy_cacherr#)*rqrGr>rr%rr9r:r;r<rrrrrrrDreturn_legacy_cache batch_size seq_lengthrpast_key_values_lengthseq_length_with_pastr num_imagesr( image_seq_lenimage_hidden_size text_seq_lenimage_batch_sizeimage_sequence_lengthimage_hidden_shaperGr&all_hidden_statesall_self_attnsnext_decoder_cacheidx decoder_layerrr next_caches* r2r}zIdeficsModel.forwardHs&&/%:!! @T@T1B1N-TXT_T_TqTq$8$D $++JjJj "+!6IDKKATATUVAW(W`m`t`tN  %,*>)..077;a?L  % %n&91 ='J;<8L  !)33A6L L2JL`#abaT b cgh hq  %'??F?KL%1%7%7%; "J 9<22499*z:QkT`TfTfghgiTjkL#'"3"3)D\#4# & 1G_GdGdGf DJ M3D":"="=DJJW]"="^ "5":":: ;RTact"u  ;; $ $#+'+'?'?@S'T$3G3L3LQ3OQeQjQjklQm0 K_KdKdKfH J 7H"6  ! )/B/G/G/JL_LdLdefLg,Mab b166z:P]C]_pq,003 3==bA3::1aMR388\S]`mSmn  *9L9Q9Q9S 6 3Q"24I!J #+',zz2DV'T$#'#=#=>R#S #' $83#>"C"C"C"K!O!OVZV`V`!O!a j jop j quu    !"ZZ12%**]MaMaN11 M>?L] & #7BD0d!"+DKK"8` 6 C#!m%55!) %V**t}}"&''t!&I $ A A!!" #'((%--00"! $!'!!#1!-#2(;)=)=&7'!)-)B)B,0,H,H#1! "*!,M%28I1q%Q" =#3"55A` 6D -0  -!1 1 +4'$ #335J166z:}^op'5FXkl  .+&+% 3   }cs, W= input_tensorc |jjdk(r||dk(jr|Sy|jjdk(r7t|tj r t |}t|tr|S||jnd}t|t}|jjdk(r(|s&|s$tj||||jry|j|j} }|jd} |r|j!} n1t|tj r|jdn|| zdz} |j#|| | || ||jd } |jjdk(rQ|O|jj$d vr7|s5t j&|j(} tj*| | } | S) Nflash_attention_2r flex_attentionrsdpa)rr is_trainingrr8)sequence_length target_lengthrirDrr)r xpu)r_attn_implementationrZrTr-rr!r rrr_ignore_causal_mask_sdparrirDr@get_max_cache_shape5_prepare_4d_causal_attention_mask_with_cache_positionrfinfomin_unmask_unattended)rqr>rrr%rpast_seen_tokensusing_static_cacherirDrrr min_dtypes r2rz IdeficsModel._update_causal_maskWs ;; + +/B B)~/D.I.I.K%% ;; + +/? ?.%,,7!<^!L.)4%% @O?Z?99;`a'E ;; + +v 5>PYj%>>*'7 MM  $**L,?,?v&,,Q/ +??AMnell;$$R(%7!; PP +')#))!,Q  KK , , 6*%%**o=%  E*..I0CCKQZ[Kr1rrrirDrc ||jdk(r|}|Stj|j} tj||f| ||}|dk7rtj |d}|tj |||jddkDz}|ddddddfj|ddd}||j}|jd} |ddddddd| f|ddddddfj|jz} | dk(} |ddddddd| fj| | |ddddddd| f<|S) a Creates a causal 4D mask of shape `(batch_size, 1, query_length, key_value_length)` from a 2D mask of shape `(batch_size, key_value_length)`, or if the input `attention_mask` is already 4D, do nothing. Args: attention_mask (`torch.Tensor`): A 2D attention mask of shape `(batch_size, key_value_length)` or a 4D attention mask of shape `(batch_size, 1, query_length, key_value_length)`. sequence_length (`int`): The sequence length being processed. target_length (`int`): The target length: when generating with static cache, the mask should be as long as the static cache, to account for the 0 padding, the part of the cache that is not filled yet. dtype (`torch.dtype`): The dtype to use for the 4D attention mask. device (`torch.device`): The device to place the 4D attention mask on. cache_position (`torch.Tensor`): Indices depicting the position of the input sequence tokens in the sequence. batch_size (`torch.Tensor`): Batch size. N) fill_valuerirDr)diagonalrr8r)rr-rrfulltriur?rexpandrwr@rCrDrI) r>rrrirDrrrrrr mask_length padding_masks r2rzBIdeficsModel._prepare_4d_causal_attention_mask_with_cache_positionsyD  %.*<*<*>!*C(K*' E*..I** -0Ye\bK!##jjqA 5<< fEH^H^_acdHee eK%dD!Q&67>>z1bRTUK))//1 ,2226 *1aL[L+@ANSTVZ\`bcScDdDgDg&&E   ,q0 5@Aq,;,AV5W5c5c )6 Aq!\k\12r1rS)NNNNNNNNNNNNFNNF) r)r*r+r,rrmrsr_ryrrrLLAMA_INPUTS_DOCSTRINGrr-r!rrr.rr rr#r}rr staticmethodrrirDrrrs@r2rZrZsw .,}.,`f46F68M!"++AB151537=A5948@D<@7;$(,0/338&*59!K E,,-K !.K u//0 K "$u'8'8"9: K   1 12 K u001K #+5+<+<"=K 'u'8'89K 'u||4K D>K $D>K 'tnK #+4.K d^K !!1!12!K " u44 5#K CK h#( D DllD D  D  DL7 777{{ 7  7  777r1rZc'eZdZdgZddgZd$fd ZdZdZdZdZ dZ d Z d Z e eeee  d%d eej(d eej*deej(deeej.deej.deej.deej.deej.deej*deej(deedeedeedeedeedeej(deeeff"dZ d&fd Z d'ded eee fd!edeee fffd" Z!e"d#Z#xZ$S)(IdeficsForVisionText2Textzlm_head.weightzmodel.embed_tokens.weightct||t||_t |j |j |jd|j|_ |jy)NF)rrrrrd) rlrmrZr\rrr]r^freeze_lm_headlm_headrr)rqrrbrss r2rmz"IdeficsForVisionText2Text.__init__s[  !&) -****$*$@$@#22    r1c.|jjSrSr\r`rs r2rz.IdeficsForVisionText2Text.get_input_embeddingsszz&&&r1c&||j_yrSrrs r2rz.IdeficsForVisionText2Text.set_input_embeddingss"' r1c|jSrSrrs r2get_output_embeddingsz/IdeficsForVisionText2Text.get_output_embeddingss ||r1c||_yrSr)rqnew_embeddingss r2set_output_embeddingsz/IdeficsForVisionText2Text.set_output_embeddingss % r1c||_yrSr\)rqdecoders r2 set_decoderz%IdeficsForVisionText2Text.set_decoders  r1c|jSrSrrs r2 get_decoderz%IdeficsForVisionText2Text.get_decoders zzr1c|j}|j}t|jddr`|j|_|j dkDr@|j |j k(sJ|jj|j_t|drJt|dr=|j|_ t|drt|dr|j |_yyyyy) z Overwrite `transformers.modeling_utils.PreTrainedModel.tie_weights` to handle the case of IdeficsDecoupledLinear and IdeficsDecoupledEmbedding. tie_word_embeddingsTrrrgrrnN) rrgetattrrrornrrprrrgr)rqoutput_embeddingsinput_embeddingss r2 tie_weightsz%IdeficsForVisionText2Text.tie_weightss !668446 4;; 5t <'7'>'>  $99A=(@@DTDnDnnnn9I9^9^9e9e!//6 $n 5'BRTd:e-=-L-L  *(*CD "=J=Mrr%rr9r:r;r<labelsrrrrrrrec| | n|jj} | | n|jj} ||n|jj}|j ||||||||| | | | |||}|d}|j |}d}| | j |j} |s|dd|jddz dfj |j}|dddddf|dk7j}| dddf|dk7j}n1|dddddfj}| dddfj}t}||jd|jd|jd}|s|f|ddz}||f|zS|St|||j|j|j |j"S)a  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]`. Returns: Example: ```python >>> from transformers import AutoProcessor, IdeficsForVisionText2Text >>> model = IdeficsForVisionText2Text.from_pretrained("HuggingFaceM4/idefics-9b") >>> processor = AutoProcessor.from_pretrained("HuggingFaceM4/idefics-9b") >>> dogs_image_url_1 = "https://huggingface.co/datasets/hf-internal-testing/fixtures_nlvr2/raw/main/image1.jpeg" >>> dogs_image_url_2 = "https://huggingface.co/datasets/hf-internal-testing/fixtures_nlvr2/raw/main/image2.jpeg" >>> prompts = [ ... [ ... "User:", ... dogs_image_url_1, ... "Describe this image.\nAssistant: An image of two dogs.\n", ... "User:", ... dogs_image_url_2, ... "Describe this image.\nAssistant:", ... ] ... ] >>> inputs = processor(prompts, return_tensors="pt") >>> generate_ids = model.generate(**inputs, max_new_tokens=6) >>> processor.batch_decode(generate_ids, skip_special_tokens=True) ```N)rGr>rr%rr9r:r;r<rrrrrrrr.r8)r5r6r%r&r'r()rrrrr\rrCrDr@rr rArr4r%r&r'r()rqrGr>rr%rr9r:r;r<rrrrrrrr2r&r6r5shift_attention_mask shift_logits shift_labelsloss_fctrs r2r}z!IdeficsForVisionText2Text.forwards#n2C1N-TXT_T_TqTq$8$D $++JjJj &1%rrrrr<rGrr<)rrcpoprlprepare_inputs_for_generationr@)rqrGr>rrr%rr9r(r<rrr images_kwargs model_inputsrrss r2rz7IdeficsForVisionText2Text.prepare_inputs_for_generations   *{{((8K 45.s$gjZ%<%m $D>m 'tnm #+4.m d^!m "!!1!12#m $ u33 4%m hCm d !+b$) 38n!  c3h 4r1r)rrZrK)rFNN)r)Or, dataclassesrtypingrrrrrr r-torch.nn.functionalr rrutorch.utils.checkpointtorch.nnr activationsr cache_utilsrrr generationrmodeling_attn_mask_utilsrmodeling_outputsrmodeling_utilsrr pytorch_utilsrutilsrrrrrconfiguration_ideficsr perceiverrvisionr!torch.nn.attention.flex_attentionr integrations.flex_attentionr! get_loggerr)rrr#r4rMrarQrcrPrModulerappendrrrrrr#r4LLAMA_START_DOCSTRINGrKrrZr__all__r0r1r2rs(!:: %!;;)>+?110, !;J   H %! )C[)C )CX &CK&C &CV *#Z+- k  k \> RYY> DJRYYJ.N+$ uxx$ N(:PP$l9ryyl9`L"))L^LbiiL^"Y?_? ?0?DY])] ]@w 6wt Rr1