o
    wig                     @   s  d dl mZmZmZ d dl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mZ ddlmZ dd	lmZ dd
lmZmZ ddlmZmZ ddlmZ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$)e*Z+G dd dej,Z-G dd dej,Z.dej/de0dej/fddZ1	d8dej,dej/dej/dej/d eej/ d!e2d"e2fd#d$Z3d%d& Z4d9d'd(Z5G d)d* d*ej,Z6G d+d, d,ej,Z7G d-d. d.eZ8e"G d/d0 d0eZ9e"G d1d2 d2e9Z:G d3d4 d4ee!Z;e"G d5d6 d6e9eZ<g d7Z=dS ):    )CallableOptionalUnionN   )ACT2FN)CacheDynamicCache)GenerationMixin)create_causal_mask!create_sliding_window_causal_mask)FlashAttentionKwargs)GradientCheckpointingLayer)BaseModelOutputWithPastCausalLMOutputWithPast)ROPE_INIT_FUNCTIONSdynamic_rope_update)ALL_ATTENTION_FUNCTIONSPreTrainedModel)Unpack)
LossKwargsauto_docstringcan_return_tuplelogging)deprecate_kwarg   )Cohere2Configc                       s8   e Zd Zddef fddZe edd Z  Z	S )Cohere2RotaryEmbeddingNconfigc                    s   t    t|dr|jd ur|jd|jd| _nd| _|j| _|j| _|| _	t
| j | _| | j	|\}| _| jd|dd | j| _d S )Nrope_scaling	rope_typetypedefaultinv_freqF)
persistent)super__init__hasattrr   getr   max_position_embeddingsmax_seq_len_cachedoriginal_max_seq_lenr   r   rope_init_fnattention_scalingregister_bufferr"   original_inv_freq)selfr   devicer"   	__class__ i/home/ubuntu/sommelier/.venv/lib/python3.10/site-packages/transformers/models/cohere2/modeling_cohere2.pyr%   .   s   
zCohere2RotaryEmbedding.__init__c           
      C   s   | j d d d d f  |jd dd}|d d d d d f  }t|jjtr2|jjdkr2|jjnd}tj	|dd* | |  
dd}tj|ddd	}| | j }| | j }	W d    n1 shw   Y  |j|jd
|	j|jd
fS )Nr   r   mpscpuF)device_typeenabled   dimdtype)r"   floatexpandshape
isinstancer0   r    strtorchautocast	transposerepeat_interleavecosr,   sintor>   )
r/   xposition_idsinv_freq_expandedposition_ids_expandedr8   freqsembrH   rI   r3   r3   r4   forward?   s   (&zCohere2RotaryEmbedding.forwardN)
__name__
__module____qualname__r   r%   rD   no_gradr   rQ   __classcell__r3   r3   r1   r4   r   -   s
    r   c                       s&   e Zd Zd fdd	Zdd Z  ZS )	Cohere2LayerNormNh㈵>Fc                    s&   t    tt|| _|| _dS )zcThe hidden size can be a tuple or an int. The tuple is used for QKNorm to normalize across head_dimN)r$   r%   nn	ParameterrD   onesweightvariance_epsilon)r/   hidden_sizeepsbiasr1   r3   r4   r%   P   s   

zCohere2LayerNorm.__init__c                 C   sl   |j }|tj}|jddd}|| djddd}|| t|| j  }| jtj| }||S )Nr5   T)keepdimr:   )	r>   rJ   rD   float32meanpowrsqrtr^   r]   )r/   hidden_statesinput_dtyperd   variancer3   r3   r4   rQ   V   s   
zCohere2LayerNorm.forward)NrY   FrS   rT   rU   r%   rQ   rW   r3   r3   r1   r4   rX   O   s    rX   rg   n_repreturnc                 C   s^   | j \}}}}|dkr| S | dddddddddf |||||} | ||| ||S )z
    This is the equivalent of torch.repeat_interleave(x, dim=1, repeats=n_rep). The hidden states go from (batch,
    num_key_value_heads, seqlen, head_dim) to (batch, num_attention_heads, seqlen, head_dim)
    r   N)rA   r@   reshape)rg   rk   batchnum_key_value_headsslenhead_dimr3   r3   r4   	repeat_kv`   s
   0rr           modulequerykeyvalueattention_maskscalingdropoutc                 K   s   t || j}t || j}	t||dd| }
|d ur3|d d d d d d d |jd f }|
| }
tjj|
dtj	d
|j}
tjj|
|| jd}
t|
|	}|dd }||
fS )Nr:   r   r5   )r<   r>   )ptrainingr   )rr   num_key_value_groupsrD   matmulrF   rA   rZ   
functionalsoftmaxrc   rJ   r>   rz   r}   
contiguous)rt   ru   rv   rw   rx   ry   rz   kwargs
key_statesvalue_statesattn_weightscausal_maskattn_outputr3   r3   r4   eager_attention_forwardl   s   
&r   c                 C   sB   | dd d df }| ddd df }t j| |gddd}|S )N.r:   r   r5   r;   r{   )rD   stackflatten)rK   x1x2rot_xr3   r3   r4   rotate_half   s   r   c           	      C   sj   | j }|  } | }||}||}| | t| |  }|| t||  }|j|d|j|dfS )a  Applies 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`, *optional*):
            Deprecated and unused.
        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.
    r=   )r>   r?   	unsqueezer   rJ   )	qkrH   rI   rL   unsqueeze_dimr>   q_embedk_embedr3   r3   r4   apply_rotary_pos_emb   s   

r   c                       s   e Zd ZdZddedee f fddZ		ddej	de
ej	ej	f d	eej	 d
ee deej dee de
ej	eej	 ee
ej	  f fddZ  ZS )Cohere2Attentionz=Multi-headed attention from 'Attention Is All You Need' paperNr   	layer_idxc                    s   t    || _|| _t|d|j|j | _|j|j | _	| jd | _
|j| _d| _|j| dkr4|jnd | _tj|j|j| j |jd| _tj|j|j| j |jd| _tj|j|j| j |jd| _tj|j| j |j|jd| _d S )Nrq   g      Tsliding_attentionra   )r$   r%   r   r   getattrr_   num_attention_headsrq   ro   r~   ry   attention_dropout	is_causallayer_typessliding_windowrZ   Linearattention_biasq_projk_projv_projo_projr/   r   r   r1   r3   r4   r%      s*   
zCohere2Attention.__init__rg   position_embeddingsrx   past_key_valuecache_positionr   rl   c                 K   s2  |j d d }g |d| jR }| ||dd}	| ||dd}
| ||dd}|\}}| jd urGt|	|
||\}	}
|d ur\|||d}|	|
|| j
|\}
}t}| jjdkrjt| jj }|| |	|
||f| jsvdn| j| j| jd|\}}|jg |dR   }| |}||fS )Nr5   r   r:   )rI   rH   r   eagerrs   )rz   ry   r   )rA   rq   r   viewrF   r   r   r   r   updater   r   r   _attn_implementationr   r}   r   ry   rm   r   r   )r/   rg   r   rx   r   r   r   input_shapehidden_shapequery_statesr   r   rH   rI   cache_kwargsattention_interfacer   r   r3   r3   r4   rQ      s<   	
	

zCohere2Attention.forwardrR   )NN)rS   rT   rU   __doc__r   r   intr%   rD   Tensortupler   
LongTensorr   r   rQ   rW   r3   r3   r1   r4   r      s(    r   c                       s$   e Zd Z fddZdd Z  ZS )
Cohere2MLPc                    sr   t    || _|j| _|j| _tj| j| jdd| _tj| j| jdd| _tj| j| jdd| _	t
|j | _d S NFr   )r$   r%   r   r_   intermediate_sizerZ   r   	gate_projup_proj	down_projr   
hidden_actact_fnr/   r   r1   r3   r4   r%      s   
zCohere2MLP.__init__c                 C   s$   |  | | || | }|S rR   )r   r   r   r   )r/   rK   r   r3   r3   r4   rQ      s    zCohere2MLP.forwardrj   r3   r3   r1   r4   r      s    
r   c                       s   e Zd Zdedef fddZeddd							dd
ejde	ejejf de
ej de
e de
e de
e de
ej dee de	eje
e	ejejf  f fddZ  ZS )Cohere2DecoderLayerr   r   c                    sL   t    |j| _t||d| _t|| _t|j|jd| _	|j
| | _d S )N)r   r   r_   r`   )r$   r%   r_   r   	self_attnr   mlprX   layer_norm_epsinput_layernormr   attention_typer   r1   r3   r4   r%     s   

zCohere2DecoderLayer.__init__last_cache_positionz4.53.0)versionNFrg   r   rx   r   output_attentions	use_cacher   r   rl   c              
   K   s`   |}	|  |}| jd|||||||d|\}
}| |}|	|
 | }|f}|r.||f7 }|S )ax  
        Args:
            hidden_states (`torch.FloatTensor`): input to the layer of shape `(batch, seq_len, embed_dim)`
            position_embeddings (`tuple[torch.FloatTensor, torch.FloatTensor]`):
                Tuple containing the cosine and sine positional embeddings of shape `(batch_size, seq_len, head_dim)`,
                with `head_dim` being the embedding dimension of each attention head.
            attention_mask (`torch.FloatTensor`, *optional*):
                attention mask of size `(batch_size, sequence_length)` if flash attention is used or `(batch_size, 1,
                query_sequence_length, key_sequence_length)` if default attention is used.
            past_key_value (`Tuple(torch.FloatTensor)`, *optional*): cached past key and value projection states
            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`).
            cache_position (`torch.LongTensor` of shape `(sequence_length)`, *optional*):
                Indices depicting the position of the input sequence tokens in the sequence
        )rg   r   rx   r   r   r   r   Nr3   )r   r   r   )r/   rg   r   rx   r   r   r   r   r   residualhidden_states_attentionself_attn_weightshidden_states_mlpoutputsr3   r3   r4   rQ     s&   



zCohere2DecoderLayer.forward)NNFFN)rS   rT   rU   r   r   r%   r   rD   r   r   r   r   boolr   r   r   FloatTensorrQ   rW   r3   r3   r1   r4   r     s6    
	
r   c                   @   sL   e Zd ZeZdZdZdgZdgZdZ	dZ
dZdZdZdZdZdZdd ZdS )Cohere2PreTrainedModelmodelTr   past_key_valuesc                 C   s   | j j}t|tjr"|jjjd|d |jd ur |jj	  d S d S t|tj
rC|jjjd|d |jd urA|jj|j 	  d S d S t|trQ|jjd d S d S )Nrs   )rd   stdg      ?)r   initializer_rangerB   rZ   r   r]   datanormal_ra   zero_	Embeddingpadding_idxrX   fill_)r/   rt   r   r3   r3   r4   _init_weightsZ  s   


z$Cohere2PreTrainedModel._init_weightsN)rS   rT   rU   r   config_classbase_model_prefixsupports_gradient_checkpointing_no_split_modules_skip_keys_device_placement_supports_flash_attn_3_supports_flash_attn_2_supports_sdpa_supports_flex_attn_supports_cache_class_supports_quantized_cache_supports_static_cache_supports_attention_backendr   r3   r3   r3   r4   r   J  s    r   c                       s   e Zd Zdef fddZdd Zdd Zee									dd	e	e
j d
e	e
j de	e
j de	e de	e
j de	e de	e de	e de	e
j dee defddZ  ZS )Cohere2Modelr   c                    s   t     j| _ j| _t j j| j| _t	 fddt
 jD | _t j jd| _t d| _d| _|   d S )Nc                    s   g | ]}t  |qS r3   )r   ).0r   r   r3   r4   
<listcomp>q  s    z)Cohere2Model.__init__.<locals>.<listcomp>r   r   F)r$   r%   pad_token_idr   
vocab_sizerZ   r   r_   embed_tokens
ModuleListrangenum_hidden_layerslayersrX   r   normr   
rotary_embgradient_checkpointing	post_initr   r1   r   r4   r%   j  s   zCohere2Model.__init__c                 C      | j S rR   r   r/   r3   r3   r4   get_input_embeddingsz     z!Cohere2Model.get_input_embeddingsc                 C   
   || _ d S rR   r   r/   rw   r3   r3   r4   set_input_embeddings}     
z!Cohere2Model.set_input_embeddingsN	input_idsrx   rL   r   inputs_embedsr   r   output_hidden_statesr   flash_attn_kwargsrl   c
              
   K   s  |d ur|n| j j}|d ur|n| j j}|d ur|n| j j}|d u |d uA r*td| jr9| jr9|r9td d}|d u rB| 	|}|rN|d u rN| jsNt
 }|	d u rj|d urZ| nd}tj|||jd  |jd}	|d u rs|	d}t| }ts| j |||	||d}td	i |td	i |d}|}| ||}|rd	nd }|rd	nd }| jD ](}|r||f7 }||f|||j ||||	d
|
}|d }|r||d f7 }q| |}|r||f7 }t||||dS )Nz:You must specify exactly one of input_ids or inputs_embedszX`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`.Fr   r   )r0   )r   input_embedsrx   r   r   rL   )full_attentionr   r3   )r   rx   r   r   r   r   )last_hidden_stater   rg   
attentions)r   r   r  r   
ValueErrorr   r}   loggerwarning_oncer   r   get_seq_lengthrD   arangerA   r0   r   rB   dictr
   r   r   r   r   r   r   )r/   r  rx   rL   r   r  r   r   r  r   r  past_seen_tokenscausal_mask_mappingmask_kwargsrg   r   all_hidden_statesall_self_attnsdecoder_layerlayer_outputsr3   r3   r4   rQ     s   






zCohere2Model.forward)	NNNNNNNNN)rS   rT   rU   r   r%   r   r  r   r   r   rD   r   r   r   r   r   r   r   r   rQ   rW   r3   r3   r1   r4   r   h  sL    	
r   c                   @   s   e Zd ZdS )KwargsForCausalLMN)rS   rT   rU   r3   r3   r3   r4   r    s    r  c                       s  e Zd ZdgZddiZddgdgfiZ fddZdd	 Zd
d Zdd Z	dd Z
dd Zdd Zee											d%deej deej deej deeeeej f  deej deej dee dee dee deej d eeejf d!ee d"efd#d$Z  ZS )&Cohere2ForCausalLMzlm_head.weightlm_headcolwise_reprg   logitsc                    sP   t  | t|| _|j| _tj|j|jdd| _|j	| _	|j
| _
|   d S r   )r$   r%   r   r   r   rZ   r   r_   r  logit_scaletie_word_embeddingsr   r   r1   r3   r4   r%     s   
zCohere2ForCausalLM.__init__c                 C   s   | j jS rR   r   r   r   r3   r3   r4   r     s   z'Cohere2ForCausalLM.get_input_embeddingsc                 C   s   || j _d S rR   r!  r  r3   r3   r4   r    s   z'Cohere2ForCausalLM.set_input_embeddingsc                 C   r   rR   r  r   r3   r3   r4   get_output_embeddings  r   z(Cohere2ForCausalLM.get_output_embeddingsc                 C   r  rR   r"  )r/   new_embeddingsr3   r3   r4   set_output_embeddings  r  z(Cohere2ForCausalLM.set_output_embeddingsc                 C   r  rR   r   )r/   decoderr3   r3   r4   set_decoder  r  zCohere2ForCausalLM.set_decoderc                 C   r   rR   r&  r   r3   r3   r4   get_decoder  r   zCohere2ForCausalLM.get_decoderNr   r  rx   rL   r   r  labelsr   r   r  r   logits_to_keepr   rl   c                 K   s   |dur|n| j j}|	dur|	n| j j}	| jd||||||||	|
d	|}|j}t|tr4t| dn|}| |dd|ddf }|| j	 }d}|dur]| j
d||| j jd|}t|||j|j|jd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]`.

        Example:

        ```python
        >> from transformers import AutoTokenizer, Cohere2ForCausalLM

        >> model = Cohere2ForCausalLM.from_pretrained("Cohere2ForAI/c4ai-command-r-v01")
        >> tokenizer = AutoTokenizer.from_pretrained("Cohere2ForAI/c4ai-command-r-v01")

        >> prompt = "Hey, are you conscious? Can you talk to me?"
        >> inputs = tokenizer(prompt, return_tensors="pt")

        >> # Generate
        >> generate_ids = model.generate(inputs.input_ids, max_length=30)
        >> tokenizer.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
        "Hey, are you conscious? Can you talk to me?\nI'm not conscious, but I can talk to you."
        ```N)	r  rx   rL   r   r  r   r   r  r   )r  r*  r   )lossr  r   rg   r  r3   )r   r   r  r   r  rB   r   slicer  r  loss_functionr   r   r   rg   r  )r/   r  rx   rL   r   r  r*  r   r   r  r   r+  r   r   rg   slice_indicesr  r,  r3   r3   r4   rQ     s<   '

zCohere2ForCausalLM.forward)NNNNNNNNNNr   )rS   rT   rU   _tied_weights_keys_tp_plan_pp_planr%   r   r  r#  r%  r(  r)  r   r   r   rD   r   r   r   r   listr   r   r   r   r  r   rQ   rW   r3   r3   r1   r4   r    sf    	
r  )r  r   r   )rs   )Nr   )>typingr   r   r   rD   torch.nnrZ   activationsr   cache_utilsr   r   
generationr	   masking_utilsr
   r   modeling_flash_attention_utilsr   modeling_layersr   modeling_outputsr   r   modeling_rope_utilsr   r   modeling_utilsr   r   processing_utilsr   utilsr   r   r   r   utils.deprecationr   configuration_cohere2r   
get_loggerrS   r  Moduler   rX   r   r   rr   r?   r   r   r   r   r   r   r   r   r  r  __all__r3   r3   r3   r4   <module>   sb   
"

HFo