
    qiQ                     j   d Z ddlZddlmZ ddlmZ ddlZddlmZ ddlm	Z	 ddl
mZ dd	lmZmZ dd
lmZ ddlmZ ddlmZmZmZmZ ddlmZ ddlmZ  ej8                  e      Ze G d de             Z G d dej@                        Z!	 d&dej@                  dejD                  dejD                  dejD                  dejD                  dz  de#de#fdZ$ G d dej@                        Z% G d dej@                        Z& G d  d!e      Z' G d" d#ej@                        Z( G d$ d%ej@                        Z)y)'zTPyTorch IdeficsVision model: a copy of CLIPVisionModel using a simpler config object    N)Callable)	dataclass)nn   )ACT2FN)GradientCheckpointingLayer)BaseModelOutputBaseModelOutputWithPooling)ALL_ATTENTION_FUNCTIONS)Unpack)ModelOutputTransformersKwargscan_return_tuplelogging)is_flash_attention_requested   )IdeficsVisionConfigc                       e Zd ZU dZdZej                  dz  ed<   dZej                  dz  ed<   dZ	e
ej                  df   dz  ed<   dZe
ej                  df   dz  ed<   y)IdeficsVisionModelOutputa  
    Base class for vision model's outputs that also contains image embeddings of the pooling of the last hidden states.

    Args:
        image_embeds (`torch.FloatTensor` of shape `(batch_size, output_dim)` *optional* returned when model is initialized with `with_projection=True`):
            The image embeddings obtained by applying the projection layer to the pooler_output.
        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.
        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.
    Nimage_embedslast_hidden_state.hidden_states
attentions)__name__
__module____qualname____doc__r   torchFloatTensor__annotations__r   r   tupler        T/opt/pipecat/venv/lib/python3.12/site-packages/transformers/models/idefics/vision.pyr   r   )   sr    * .2L%##d*126u((4/6:>M5**C/047>7;Je'',-4;r#   r   c                        e Zd Zdef fdZdej                  dededej                  fdZddej                  d	e
dej                  fd
Z xZS )IdeficsVisionEmbeddingsconfigc                    t         |           || _        |j                  | _        |j
                  | _        |j                  | _        t        j                  t        j                  | j                              | _        t        j                  |j                  | j                  | j                  | j                  d      | _        | j
                  | j                  z  dz  | _        | j                  dz   | _        t        j"                  | j                   | j                        | _        | j'                  dt        j(                  | j                         j+                  d      d       y )NF)in_channelsout_channelskernel_sizestridebias   r   position_ids)r   )
persistent)super__init__r'   hidden_size	embed_dim
image_size
patch_sizer   	Parameterr   randnclass_embeddingConv2dnum_channelspatch_embeddingnum_patchesnum_positions	Embeddingposition_embeddingregister_bufferarangeexpandselfr'   	__class__s     r$   r3   z IdeficsVisionEmbeddings.__init__H   s	   ++ ++ ++!||EKK,GH!yy++?? 
 !OOt>1D!--1"$,,t/A/A4>>"R^U\\$:L:L-M-T-TU\-]jopr#   
embeddingsheightwidthreturnc                    |j                   d   dz
  }| j                  | j                        }|j                   d   dz
  }||k(  r||k(  r|S |dddf   }|ddddf   }|j                   d   }	|| j                  j                  z  }
|| j                  j                  z  }|
dz   |dz   }}
t        j                  |      }|j                  dt        |      t        |      |	      }|j                  dddd      }|j                  t        j                  k(  }|r4t        j                  d       |j                  t        j                         }t"        j$                  j'                  ||
|z  ||z  fd	d
      }|r|j                  t        j                        }t        |
      |j                   d   k7  st        |      |j                   d   k7  rBt)        dt        |
      t        |      f d|j                   d   |j                   d   f d      |j                  dddd      j+                  dd|	      }t        j,                  |j/                  d      |fd      S )a#  
        This method allows to interpolate the pre-trained position encodings, to be able to use the model on higher
        resolution images.

        Source:
        https://github.com/facebookresearch/dino/blob/de9ee3df6cf39fac952ab558447af1fa1365362a/vision_transformer.py#L174
        r   Nr   r0   g?r   r.   zUpcasting patch_pos_embed to fp32 for interpolation since `upsample_bicubic2d_out_frame` in nn.functional.interpolate is not implemented for 'torch.bfloat16' dtype. This will result in a slight overhead.bicubicF)scale_factormodealign_cornerszNumber of patches for images (z/) don't match the shape of position embedding ()dim)shaperA   r/   r'   r7   mathsqrtreshapeintpermutedtyper   bfloat16loggerwarning_oncetofloatr   
functionalinterpolate
ValueErrorviewcat	unsqueeze)rF   rH   rI   rJ   r>   	pos_embedr?   class_pos_embedpatch_pos_embedr5   num_h_patchesnum_w_patchessqrt_num_positionsfp32_upcastings                 r$   interpolate_pos_encodingz0IdeficsVisionEmbeddings.interpolate_pos_encoding_   sf    !&&q)A-++D,=,=>	!*Q.-'FeO#AqD/#AqrE*$$R(	$++"8"88!7!77 (5s':MC<O}!YY}5)11!S9K5LcRdNegpq)11!Q1=(..%..@h .00=O--33'*<<mN`>`a	 4 
 -00@O}!6!6r!::c->PTcTiTijlTm>m0]1CSEW1W0X Y00?0E0Eb0I?K`K`acKd0d/eefh  *11!Q1=BB1b)Tyy/33A6HaPPr#   pixel_valuesrn   c                 `   |j                   \  }}}}|sJ|| j                  k7  s|| j                  k7  r,t        d| d| d| j                   d| j                   d	      | j                  j                  j
                  }| j                  |j                  |            }|j                  d      j                  dd      }| j                  j                  |dd      }	t        j                  |	|gd	      }
|r|
| j                  |
||      z   }
|
S |
| j                  | j                        z   }
|
S )
NzInput image size (*z) doesn't match model (z8). You should try to set `interpolate_pos_encoding=True`)r[   r.   r   r0   rS   )rU   r6   rc   r=   weightr[   r_   flatten	transposer:   rD   r   re   rn   rA   r/   )rF   ro   rn   
batch_sizer<   rI   rJ   target_dtypepatch_embedsclass_embedsrH   s              r$   forwardzIdeficsVisionEmbeddings.forward   s8   2>2D2D/
L&%'(ET__,D (% 9)4??*;;su 
 ++2288++LOO,O,OP#++A.88A>++22:q"EYYl;C
 $#d&C&CJPVX]&^^J  $d&=&=d>O>O&PPJr#   F)r   r   r   r   r3   r   TensorrY   rn   r   boolry   __classcell__rG   s   @r$   r&   r&   G   sm    q2 q./Q5<< /Q /QUX /Q]b]i]i /QbE$5$5 QU bgbnbn r#   r&   modulequerykeyvalueattention_maskscalingdropoutc                    t        j                  ||j                  dd            |z  }|||z   }t        j                  j                  |dt         j                        j                  |j                        }t        j                  j                  ||| j                        }t        j                  ||      }	|	j                  dd      j                         }	|	|fS )Nr0   rQ   )rT   r[   )ptrainingr   r.   )r   matmulrt   r   ra   softmaxfloat32r_   r[   r   r   
contiguous)
r   r   r   r   r   r   r   kwargsattn_weightsattn_outputs
             r$   eager_attention_forwardr      s     <<s}}R'<=GL!#n4==((2U]](SVVW\WbWbcL==((6??([L,,|U3K''1-88:K$$r#   c                        e Zd ZdZdef fdZ	 	 	 ddej                  dej                  dz  dej                  dz  dedz  d	e	ej                  ej                  dz  f   f
d
Z
 xZS )IdeficsVisionAttentionz=Multi-headed attention from 'Attention Is All You Need' paperr'   c                    t         |           || _        |j                  | _        |j
                  | _        | j                  | j                  z  | _        | j                  | j                  z  | j                  k7  r&t        d| j                   d| j                   d      | j                  dz  | _	        |j                  | _        d| _        t        j                  | j                  | j                        | _        t        j                  | j                  | j                        | _        t        j                  | j                  | j                        | _        t        j                  | j                  | j                        | _        y )Nz;embed_dim must be divisible by num_heads (got `embed_dim`: z and `num_heads`: z).g      F)r2   r3   r'   r4   r5   num_attention_heads	num_headshead_dimrc   scaleattention_dropoutr   	is_causalr   Lineark_projv_projq_projout_projrE   s     r$   r3   zIdeficsVisionAttention.__init__   s   ++33$..8==4>>)T^^;MdnnM] ^NN#2'  ]]D(
//ii?ii?ii?		$..$..Ar#   Nr   r   causal_attention_maskoutput_attentionsrK   c           
         |j                   \  }}}| j                  |      }| j                  |      }	| j                  |      }
|j	                  ||| j
                  | j                        j                  dd      }|	j	                  ||| j
                  | j                        j                  dd      }	|
j	                  ||| j
                  | j                        j                  dd      }
t        | j                        s||||z   }n||}n	|du| _
        t        j                  | j                  j                  t              } || ||	|
|| j                  | j                  | j                   sdn| j"                        \  }}|j%                  |||      j'                         }| j)                  |      }|sd}||fS )z#Input shape: Batch x Time x Channelr   r.   N        )r   r   r   )rU   r   r   r   rd   r   r   rt   r   r'   r   r   get_interface_attn_implementationr   r   r   r   rX   r   r   )rF   r   r   r   r   ru   
seq_lengthr5   querieskeysvaluesattention_interfacer   r   s                 r$   ry   zIdeficsVisionAttention.forward   s    -:,?,?)
J	++m,{{=)]+,,z:t~~t}}U__`acdeyyZOYYZ[]^_ZT^^T]]S]]^_abc ,DKK8).C.O!/2G!G&2!62$>DN(?(M(MKK,,.E)
 %8nnJJ#}}C$,,	%
!\ "))*j)LWWYmmK0 LL((r#   )NNF)r   r   r   r   r   r3   r   r{   r|   r!   ry   r}   r~   s   @r$   r   r      s    GB2 B. /359)./)||/) t+/)  %||d2	/)
  $;/) 
u||U\\D00	1/)r#   r   c                   V     e Zd Z fdZdej
                  dej
                  fdZ xZS )IdeficsVisionMLPc                    t         |           || _        t        |j                     | _        t        j                  |j                  |j                        | _
        t        j                  |j                  |j                        | _        y N)r2   r3   r'   r   
hidden_actactivation_fnr   r   r4   intermediate_sizefc1fc2rE   s     r$   r3   zIdeficsVisionMLP.__init__  sd    #F$5$5699V//1I1IJ99V55v7I7IJr#   r   rK   c                 l    | j                  |      }| j                  |      }| j                  |      }|S r   )r   r   r   )rF   r   s     r$   ry   zIdeficsVisionMLP.forward  s4    /**=9/r#   )r   r   r   r3   r   r{   ry   r}   r~   s   @r$   r   r     s$    KU\\ ell r#   r   c                        e Zd Zdef fdZ	 d
dej                  dej                  dedz  dee	   de
ej                     f
d	Z xZS )IdeficsVisionEncoderLayerr'   c                 D   t         |           |j                  | _        t	        |      | _        t        j                  | j                  |j                        | _	        t        |      | _        t        j                  | j                  |j                        | _        y N)eps)r2   r3   r4   r5   r   	self_attnr   	LayerNormlayer_norm_epslayer_norm1r   mlplayer_norm2rE   s     r$   r3   z"IdeficsVisionEncoderLayer.__init__  sm    ++/7<<F<Q<QR#F+<<F<Q<QRr#   r   r   r   Nr   rK   c                     |}| j                  |      } | j                  d|||d|\  }}||z   }|}| j                  |      }| j                  |      }||z   }|f}|r||fz  }|S )aI  
        Args:
            hidden_states (`torch.FloatTensor`): input to the layer of shape `(batch, seq_len, embed_dim)`
            attention_mask (`torch.FloatTensor`): attention mask of size
                `(batch, 1, tgt_len, src_len)` where padding elements are indicated by very large negative values.
                `(config.encoder_attention_heads,)`.
            output_attentions (`bool`, *optional*):
                Whether or not to return the attentions tensors of all attention layers. See `attentions` under
                returned tensors for more detail.
        )r   r   r   r"   )r   r   r   r   )rF   r   r   r   r   residualr   outputss           r$   ry   z!IdeficsVisionEncoderLayer.forward%  s    " !((7&4dnn '
')/'
 	'
#| !=0 ((7/ =0 "&Gr#   rz   )r   r   r   r   r3   r   r{   r|   r   r   r!   r   ry   r}   r~   s   @r$   r   r     si    S2 S */	&||& &  $;	&
 +,& 
u  	!&r#   r   c                        e Zd ZdZdef fdZe	 	 	 	 ddej                  dz  de	dz  de	dz  de	dz  d	e
e   d
eez  fd       Z xZS )IdeficsVisionEncoderz
    Transformer encoder consisting of `config.num_hidden_layers` self attention layers. Each layer is a
    [`IdeficsVisionEncoderLayer`].

    Args:
        config: IdeficsVisionConfig
    r'   c                     t         |           || _        t        j                  t        |j                        D cg c]  }t        |       c}      | _        d| _	        y c c}w )NF)
r2   r3   r'   r   
ModuleListrangenum_hidden_layersr   layersgradient_checkpointing)rF   r'   _rG   s      r$   r3   zIdeficsVisionEncoder.__init__X  sQ    mmPUV\VnVnPo$p1%>v%F$pq&+# %qs   A#Nr   r   output_hidden_statesreturn_dictr   rK   c                 j   ||n| j                   j                  }||n| j                   j                  }||n| j                   j                  }|rdnd}|rdnd}|}	t	        | j
                        D ]*  \  }
}|r||	fz   } ||	|fd|i|}|d   }	|s"||d   fz   }, |r||	fz   }t        |	||      S )ad  
        Args:
            inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`):
                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.
            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)
            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.
        Nr"   r   r   r   )r   r   r   )r'   r   r   use_return_dict	enumerater   r	   )rF   inputs_embedsr   r   r   r   r   encoder_statesall_attentionsr   idxencoder_layerlayer_outputss                r$   ry   zIdeficsVisionEncoder.forward^  s   @ 2C1N-TXT_T_TqTq$8$D $++JjJj 	 &1%<k$++B]B]30d%"+DKK"8 	FC#!/=2B!B) #4 	M *!,M !/=3C2E!E	F  +}.>>N+>Vd
 	
r#   )NNNN)r   r   r   r   r   r3   r   r   r{   r|   r   r   r!   r	   ry   r}   r~   s   @r$   r   r   O  s    ,2 ,  /3)-,0#'=
 t+=
  $;	=

 #Tk=
 D[=
 +,=
 
	 =
 =
r#   r   c                        e Zd Zdef fdZ	 	 	 	 	 ddej                  dz  dedz  dedz  dedz  dedz  d	ee	z  fd
Z
 xZS )IdeficsVisionTransformerr'   c                     t         |           || _        |j                  }t	        |      | _        t        j                  ||j                        | _	        t        |      | _        t        j                  ||j                        | _        y r   )r2   r3   r'   r4   r&   rH   r   r   r   pre_layrnormr   encoderpost_layernorm)rF   r'   r5   rG   s      r$   r3   z!IdeficsVisionTransformer.__init__  sj    &&	1&9LL8M8MN+F3 ll9&:O:OPr#   Nro   r   r   rn   r   rK   c                    ||n| j                   j                  }||n| j                   j                  }||n| j                   j                  }|t	        d      | j                  ||      }| j                  |      }| j                  ||||      }|d   }|dddddf   }	| j                  |	      }	|s
||	f|dd z   S t        ||	|j                  |j                        S )z
        Returns:

        Nz You have to specify pixel_values)rn   )r   r   r   r   r   r   )r   pooler_outputr   r   )r'   r   r   r   rc   rH   r   r   r   r
   r   r   )
rF   ro   r   r   rn   r   r   encoder_outputsr   pooled_outputs
             r$   ry   z IdeficsVisionTransformer.forward  s    2C1N-TXT_T_TqTq$8$D $++JjJj 	 &1%<k$++B]B]?@@Ogh))-8,,'/!5#	 ' 
 ,A.)!Q'2++M:%}58KKK)/')77&11	
 	
r#   )NNNFN)r   r   r   r   r3   r   r   r|   r!   r
   ry   r}   r~   s   @r$   r   r     s    Q2 Q 26)-,005#'+
''$.+
  $;+
 #Tk	+

 #'++
 D[+
 
+	++
r#   r   )r   )*r   rV   collections.abcr   dataclassesr   r   r   activationsr   modeling_layersr   modeling_outputsr	   r
   modeling_utilsr   processing_utilsr   utilsr   r   r   r   utils.genericr   configuration_ideficsr   
get_loggerr   r]   r   Moduler&   r{   r`   r   r   r   r   r   r   r"   r#   r$   <module>r      s6   [  $ !   ! 9 K 5 &  : 6 
		H	% <{ < <:`bii `V %II%<<% 
% <<	%
 LL4'% % %.F)RYY F)Tryy  / : /fM
299 M
b7
ryy 7
r#   