o
    eiX                     @   s*  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	m
Z
 ddlmZ dd	lmZmZm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mZ ddlmZmZ ddl m!Z! ddl"m#Z#m$Z$m%Z% ddl&m'Z'm(Z( ddl)m*Z* ddl+m,Z, edG dd dej-Z.G dd dej-Z/dd Z0edd=ddZ1d ej2d!e3d"ej2fd#d$Z4	%d>d&ej-d'ej2d(ej2d)ej2d*ej2dB d+e5d,e5d-e!e# fd.d/Z6ee1G d0d1 d1ej-Z7G d2d3 d3eZ8G d4d5 d5ej-Z9e$G d6d7 d7eZ:e$G d8d9 d9e:Z;e$G d:d; d;e:eZ<g d<Z=dS )?    )Callable)OptionalN)nn   )ACT2FN)CacheDynamicCache)GenerationMixin)use_kernel_forward_from_hubuse_kernel_func_from_hubuse_kernelized_func)create_causal_mask)FlashAttentionKwargs)GradientCheckpointingLayer)BaseModelOutputWithPastCausalLMOutputWithPast)ROPE_INIT_FUNCTIONSdynamic_rope_update)ALL_ATTENTION_FUNCTIONSPreTrainedModel)Unpack)TransformersKwargsauto_docstringcan_return_tuple)maybe_autocastmerge_with_config_defaults)capture_outputs   )BitNetConfigRMSNormc                       sF   e Zd Zddeddf fddZdejdejfdd	Zd
d Z  Z	S )BitNetRMSNormư>epsreturnNc                    s&   t    tt|| _|| _dS )z<
        BitNetRMSNorm is equivalent to T5LayerNorm
        N)super__init__r   	Parametertorchonesweightvariance_epsilon)selfhidden_sizer"   	__class__ h/home/ubuntu/transcripts/venv/lib/python3.10/site-packages/transformers/models/bitnet/modeling_bitnet.pyr%   -   s   

zBitNetRMSNorm.__init__hidden_statesc                 C   sJ   |j }|tj}|djddd}|t|| j  }| j|| S )N   T)keepdim)	dtypetor'   float32powmeanrsqrtr*   r)   )r+   r1   input_dtypevariancer/   r/   r0   forward5   s
   zBitNetRMSNorm.forwardc                 C   s   t | jj d| j S )Nz, eps=)tupler)   shaper*   )r+   r/   r/   r0   
extra_repr<   s   zBitNetRMSNorm.extra_repr)r!   )
__name__
__module____qualname__floatr%   r'   Tensorr=   r@   __classcell__r/   r/   r-   r0   r    +   s    r    c                       s*   e Zd Zdef fddZdd Z  ZS )	BitNetMLPconfigc                    s   t    || _|j| _|j| _tj| j| jdd| _tj| j| jdd| _tj| j| jdd| _	t
|j | _t|j|jd| _d S )NFbiasr"   )r$   r%   rH   r,   intermediate_sizer   Linear	gate_projup_proj	down_projr   
hidden_actact_fnr    rms_norm_epsffn_sub_normr+   rH   r-   r/   r0   r%   A   s   
zBitNetMLP.__init__c              	   C   s*   |  | | | || | }|S N)rP   rT   rR   rN   rO   )r+   xrP   r/   r/   r0   r=   L   s   &zBitNetMLP.forward)rA   rB   rC   r   r%   r=   rF   r/   r/   r-   r0   rG   @   s    rG   c                 C   sH   | dd| j d d f }| d| j d d df }tj| |fddS )z*Rotates half the hidden dims of the input..Nr3   r2   dim)r?   r'   cat)rW   x1x2r/   r/   r0   rotate_halfQ   s   r]   rotary_pos_embc                 C   sD   | |}| |}| | t| |  }|| t||  }||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.
        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]   )qkcossinunsqueeze_dimq_embedk_embedr/   r/   r0   apply_rotary_pos_embX   s
   

rg   r1   n_repr#   c                 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)r?   expandreshape)r1   rh   batchnum_key_value_headsslenhead_dimr/   r/   r0   	repeat_kvr   s
   0ro           modulequerykeyvalueattention_maskscalingdropoutkwargsc                 K   s   t || j}t || j}	t||dd| }
|d ur |
| }
tjj|
dtjd	|j
}
tjj|
|| jd}
t|
|	}|dd }||
fS )Nr2   r   r3   )rY   r5   )ptrainingr   )ro   num_key_value_groupsr'   matmul	transposer   
functionalsoftmaxr7   r6   r5   rw   rz   
contiguous)rq   rr   rs   rt   ru   rv   rw   rx   
key_statesvalue_statesattn_weightsattn_outputr/   r/   r0   eager_attention_forward~   s   
r   c                       s   e Zd ZdZdedef fddZ		ddejde	ejejf d	ejdB d
e
dB dejdB dee de	ejejdB f fddZ  ZS )BitNetAttentionz=Multi-headed attention from 'Attention Is All You Need' paperrH   	layer_idxc                    s   t    || _|| _t|d|j|j | _|j|j | _	| jd | _
|j| _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| _t|j|jd| _d S )Nrn   g      TrI   rK   )r$   r%   rH   r   getattrr,   num_attention_headsrn   rl   r{   rv   attention_dropout	is_causalr   rM   attention_biasq_projk_projv_projo_projr    rS   attn_sub_normr+   rH   r   r-   r/   r0   r%      s*   
zBitNetAttention.__init__Nr1   position_embeddingsru   past_key_valuescache_positionrx   r#   c                 K   s"  |j d d }g |d| jR }| ||dd}	| ||dd}
| ||dd}|\}}t|	|
||\}	}
|d urW|||d}||
|| j	|\}
}t
| jjt}|| |	|
||f| jskdn| j| jd|\}}|jg |dR   }| |}| |}||fS )Nr3   r   r2   )rc   rb   r   rp   )rw   rv   )r?   rn   r   viewr}   r   r   rg   updater   r   get_interfacerH   _attn_implementationr   rz   r   rv   rj   r   r   r   )r+   r1   r   ru   r   r   rx   input_shapehidden_shapequery_statesr   r   rb   rc   cache_kwargsattention_interfacer   r   r/   r/   r0   r=      s:   	


zBitNetAttention.forward)NN)rA   rB   rC   __doc__r   intr%   r'   rE   r>   r   
LongTensorr   r   r=   rF   r/   r/   r-   r0   r      s(    r   c                       s   e Zd Zdedef fddZ						ddejdejdB d	ejdB d
e	dB de
dB dejdB deejejf dB dee dejfddZ  ZS )BitNetDecoderLayerrH   r   c                    sR   t    |j| _t||d| _t|| _t|j|jd| _	t|j|jd| _
d S )N)rH   r   rK   )r$   r%   r,   r   	self_attnrG   mlpr    rS   input_layernormpost_attention_layernormr   r-   r/   r0   r%      s   

zBitNetDecoderLayer.__init__NFr1   ru   position_idsr   	use_cacher   r   rx   r#   c              
   K   s^   |}	|  |}| jd|||||||d|\}}
|	| }|}	| |}| |}|	| }|S )N)r1   ru   r   r   r   r   r   r/   )r   r   r   r   )r+   r1   ru   r   r   r   r   r   rx   residual_r/   r/   r0   r=      s&   




zBitNetDecoderLayer.forward)NNNFNN)rA   rB   rC   r   r   r%   r'   rE   r   r   boolr>   r   r   r=   rF   r/   r/   r-   r0   r      s6    	
r   c                       s~   e Zd ZU ejed< ddef fddZe			ddedB de	d de
dB d	ed
ef fddZe edd Z  ZS )BitNetRotaryEmbeddinginv_freqNrH   c                    s   t    |j| _|j| _|| _| jjd | _| j}| jdkr$t	| j }|| j|\}| _
| jd|dd | jd| dd d S )N	rope_typedefaultr   F)
persistentoriginal_inv_freq)r$   r%   max_position_embeddingsmax_seq_len_cachedoriginal_max_seq_lenrH   rope_parametersr   compute_default_rope_parametersr   attention_scalingregister_bufferclone)r+   rH   devicerope_init_fnr   r-   r/   r0   r%     s   


zBitNetRotaryEmbedding.__init__r   ztorch.deviceseq_lenr#   ztorch.Tensorc                 C   sZ   | j d }t| ddp| j| j }d}d|tjd|dtjdj|tjd|   }||fS )	a  
        Computes the inverse frequencies according to the original RoPE implementation
        Args:
            config ([`~transformers.PreTrainedConfig`]):
                The model configuration.
            device (`torch.device`):
                The device to use for initialization of the inverse frequencies.
            seq_len (`int`, *optional*):
                The current sequence length. Unused for this type of RoPE.
        Returns:
            Tuple of (`torch.Tensor`, `float`), containing the inverse frequencies for the RoPE embeddings and the
            post-processing scaling factor applied to the computed cos/sin (unused in this type of RoPE).
        
rope_thetarn   Ng      ?r   r2   r5   )r   r5   )	r   r   r,   r   r'   arangeint64r6   rD   )rH   r   r   baserY   attention_factorr   r/   r/   r0   r      s   
&z5BitNetRotaryEmbedding.compute_default_rope_parametersc           
      C   s   | j d d d d f  |jd dd|j}|d d d d d f  }t|jjtr6|jjdkr6|jjnd}t	|dd+ | |  
dd}tj||fdd	}| | j }| | j }	W d    n1 slw   Y  |j|jd
|	j|jd
fS )Nr   r3   r   mpscpuF)device_typeenabledr2   rX   r   )r   rD   ri   r?   r6   r   
isinstancetypestrr   r}   r'   rZ   rb   r   rc   r5   )
r+   rW   r   inv_freq_expandedposition_ids_expandedr   freqsembrb   rc   r/   r/   r0   r=   >  s   0&zBitNetRotaryEmbedding.forwardrV   )NNN)rA   rB   rC   r'   rE   __annotations__r   r%   staticmethodr   r   r>   rD   r   no_gradr   r=   rF   r/   r/   r-   r0   r     s&   
 

r   c                   @   sH   e Zd ZU eed< dZdZdgZdgZdZ	dZ
dZdZdZeedZdS )BitNetPreTrainedModelrH   modelTr   r   )r1   
attentionsN)rA   rB   rC   r   r   base_model_prefixsupports_gradient_checkpointing_no_split_modules_skip_keys_device_placement_supports_flash_attn_supports_sdpa_supports_flex_attn_can_compile_fullgraph_supports_attention_backendr   r   _can_record_outputsr/   r/   r/   r0   r   N  s   
 
r   c                       s   e Zd Zdef fddZeee							ddej	dB dej
dB dej	dB dedB d	ejdB d
ej	dB dedB dee defddZ  ZS )BitNetModelrH   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 r/   )r   ).0r   rH   r/   r0   
<listcomp>j  s    z(BitNetModel.__init__.<locals>.<listcomp>rK   r   F)r$   r%   pad_token_idpadding_idx
vocab_sizer   	Embeddingr,   embed_tokens
ModuleListrangenum_hidden_layerslayersr    rS   normr   
rotary_embgradient_checkpointing	post_initrU   r-   r   r0   r%   c  s   zBitNetModel.__init__N	input_idsru   r   r   inputs_embedsr   r   rx   r#   c              
   K   s   |d u |d uA rt d|d u r| |}|r!|d u r!t| jd}|d u r<|d ur-| nd}	tj|jd |jd|	 }|d u rE|	d}t
| j|||||d}
|}| j||d}| jd | jj D ]}||f|
|||||d|}qb| |}t||d	S )
Nz:You must specify exactly one of input_ids or inputs_embedsr   r   r   )r   )rH   r   ru   r   r   r   )r   )ru   r   r   r   r   r   )last_hidden_stater   )
ValueErrorr   r   rH   get_seq_lengthr'   r   r?   r   r_   r   r   r   r   r   r   )r+   r   ru   r   r   r   r   r   rx   past_seen_tokenscausal_maskr1   r   decoder_layerr/   r/   r0   r=   s  sP   

	
zBitNetModel.forward)NNNNNNN)rA   rB   rC   r   r%   r   r   r   r'   r   rE   r   FloatTensorr   r   r   r   r=   rF   r/   r/   r-   r0   r   a  s>    	
r   c                       s   e Zd ZddiZdZdZ fddZee									dde	j
dB de	jdB d	e	j
dB d
edB de	jdB de	j
dB dedB de	j
dB dee	jB dee defddZ  ZS )BitNetForCausalLMzlm_head.weightzmodel.embed_tokens.weightNc                    s@   t  | t|| _|j| _tj|j|jdd| _| 	  d S )NFrI   )
r$   r%   r   r   r   r   rM   r,   lm_headr   rU   r-   r/   r0   r%     s
   
zBitNetForCausalLM.__init__r   r   ru   r   r   r   labelsr   r   logits_to_keeprx   r#   c
              
   K   s   | j d|||||||d|
}|j}t|	trt|	 dn|	}| |dd|ddf }d}|durB| 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, transformers.,
            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, transformers., config.vocab_size]`.

        Example:

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

        >>> model = BitNetForCausalLM.from_pretrained("microsoft/bitnet-b1.58-2B-4T")
        >>> tokenizer = AutoTokenizer.from_pretrained("microsoft/bitnet-b1.58-2B-4T")

        >>> prompt = f'<|begin_of_text|>User: Hey, are you conscious? Can you talk to me?<|eot_id|>Assistant: '
        >>> inputs = tokenizer(prompt, return_tensors="pt")

        >>> # Generate
        >>> generate_ids = model.generate(inputs.input_ids, max_length=100)
        >>> tokenizer.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
        "User: Hey, are you conscious? Can you talk to me?Assistant: No, I'm not conscious. I'm an artificial intelligence designed to assist with information and tasks. How can I help you today?"
        ```)r   ru   r   r   r   r   r   N)logitsr   r   )lossr   r   r1   r   r/   )r   r   r   r   slicer   loss_functionrH   r   r   r   r1   r   )r+   r   ru   r   r   r   r   r   r   r   rx   outputsr1   slice_indicesr   r   r/   r/   r0   r=     s0   %zBitNetForCausalLM.forward)	NNNNNNNNr   )rA   rB   rC   _tied_weights_keys_tp_plan_pp_planr%   r   r   r'   r   rE   r   r   r   r   r   r   r   r=   rF   r/   r/   r-   r0   r     sN    		
r   )r   r   r   )r   )rp   )>collections.abcr   typingr   r'   r   activationsr   cache_utilsr   r   
generationr	   integrationsr
   r   r   masking_utilsr   modeling_flash_attention_utilsr   modeling_layersr   modeling_outputsr   r   modeling_rope_utilsr   r   modeling_utilsr   r   processing_utilsr   utilsr   r   r   utils.genericr   r   utils.output_capturingr   configuration_bitnetr   Moduler    rG   r]   rg   rE   r   ro   rD   r   r   r   r   r   r   r   __all__r/   r/   r/   r0   <module>   sn   
H-APP