o
    wi~a                     @   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 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#m$Z$ ddl%m&Z& e$'e(Z)edG dd dej*Z+G dd dej*Z,dd Z-d:ddZ.dej/de0dej/fdd Z1	!d;d"ej*d#ej/d$ej/d%ej/d&eej/ d'e2d(e2fd)d*Z3G d+d, d,ej*Z4G d-d. d.eZ5G d/d0 d0ej*Z6e"G d1d2 d2eZ7e"G d3d4 d4e7Z8G d5d6 d6ee!Z9e"G d7d8 d8e7eZ:g d9Z;dS )<    )CallableOptionalUnionN)nn   )ACT2FN)CacheDynamicCache)GenerationMixin)use_kernel_forward_from_hub)create_causal_mask)FlashAttentionKwargs)GradientCheckpointingLayer)BaseModelOutputWithPastCausalLMOutputWithPast)ROPE_INIT_FUNCTIONSdynamic_rope_update)ALL_ATTENTION_FUNCTIONSPreTrainedModel)Unpack)
LossKwargsauto_docstringcan_return_tuplelogging   )BitNetConfigRMSNormc                       s.   e Zd Zd fdd	Zdd Zdd Z  ZS )	BitNetRMSNormư>c                    s&   t    tt|| _|| _dS )z<
        BitNetRMSNorm is equivalent to T5LayerNorm
        N)super__init__r   	Parametertorchonesweightvariance_epsilon)selfhidden_sizeeps	__class__ g/home/ubuntu/sommelier/.venv/lib/python3.10/site-packages/transformers/models/bitnet/modeling_bitnet.pyr    .   s   

zBitNetRMSNorm.__init__c                 C   sJ   |j }|tj}|djddd}|t|| j  }| j|| S )N   T)keepdim)	dtypetor"   float32powmeanrsqrtr%   r$   )r&   hidden_statesinput_dtypevariancer+   r+   r,   forward6   s
   zBitNetRMSNorm.forwardc                 C   s   t | jj d| j S )Nz, eps=)tupler$   shaper%   r&   r+   r+   r,   
extra_repr=   s   zBitNetRMSNorm.extra_repr)r   )__name__
__module____qualname__r    r9   r=   __classcell__r+   r+   r)   r,   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    rC   r'   intermediate_sizer   Linear	gate_projup_proj	down_projr   
hidden_actact_fnr   rms_norm_epsffn_sub_normr&   rC   r)   r+   r,   r    B   s   
zBitNetMLP.__init__c              	   C   s*   |  | | | || | }|S N)rK   rO   rM   rI   rJ   )r&   xrK   r+   r+   r,   r9   M   s   &zBitNetMLP.forward)r>   r?   r@   r   r    r9   rA   r+   r+   r)   r,   rB   A   s    rB   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..Nr.   r-   dim)r;   r"   cat)rR   x1x2r+   r+   r,   rotate_halfR   s   rX   c                 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.
        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.
    )	unsqueezerX   )qkcossinposition_idsunsqueeze_dimq_embedk_embedr+   r+   r,   apply_rotary_pos_embY   s
   

rb   r6   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)r;   expandreshape)r6   rc   batchnum_key_value_headsslenhead_dimr+   r+   r,   	repeat_kvt   s
   0rk           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   r.   )rT   r0   )ptrainingr   )rk   num_key_value_groupsr"   matmul	transposer;   r   
functionalsoftmaxr2   r1   r0   rs   rv   
contiguous)rm   rn   ro   rp   rq   rr   rs   kwargs
key_statesvalue_statesattn_weightscausal_maskattn_outputr+   r+   r,   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
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 )BitNetAttentionz=Multi-headed attention from 'Attention Is All You Need' paperrC   	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 )Nrj   g      TrD   rF   )r   r    rC   r   getattrr'   num_attention_headsrj   rh   rw   rr   attention_dropout	is_causalr   rH   attention_biasq_projk_projv_projo_projr   rN   attn_sub_normr&   rC   r   r)   r+   r,   r       s*   
zBitNetAttention.__init__Nr6   position_embeddingsrq   past_key_valuecache_positionr}   rd   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dkret| jj }|| |	|
||f| jsqdn| j| jd|\}}|jg |dR   }| |}| |}||fS )Nr.   r   r-   )r]   r\   r   eagerrl   )rs   rr   )r;   rj   r   viewry   r   r   rb   updater   r   rC   _attn_implementationr   rv   r   rr   rf   r|   r   r   )r&   r6   r   rq   r   r   r}   input_shapehidden_shapequery_statesr~   r   r\   r]   cache_kwargsattention_interfacer   r   r+   r+   r,   r9      s:   	


zBitNetAttention.forward)NN)r>   r?   r@   __doc__r   intr    r"   Tensorr:   r   r   
LongTensorr   r   r9   rA   r+   r+   r)   r,   r      s(    r   c                       s   e Zd Zdedef fddZ							ddej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ejf  dee deejeeejejf  f fddZ  ZS )BitNetDecoderLayerrC   r   c                    sR   t    |j| _t||d| _t|| _t|j|jd| _	t|j|jd| _
d S )N)rC   r   rF   )r   r    r'   r   	self_attnrB   mlpr   rN   input_layernormpost_attention_layernormr   r)   r+   r,   r       s   

zBitNetDecoderLayer.__init__NFr6   rq   r^   r   output_attentions	use_cacher   r   r}   rd   c	                 K   st   |}
|  |}| jd||||||||d|	\}}|
| }|}
| |}| |}|
| }|f}|r8||f7 }|S )N)r6   rq   r^   r   r   r   r   r   r+   )r   r   r   r   )r&   r6   rq   r^   r   r   r   r   r   r}   residualself_attn_weightsoutputsr+   r+   r,   r9      s.   
	



zBitNetDecoderLayer.forward)NNNFFNN)r>   r?   r@   r   r   r    r"   r   r   r   r   boolr:   r   r   FloatTensorr9   rA   r+   r+   r)   r,   r      s<    	
r   c                       s8   e Zd Zddef fddZe edd Z  Z	S )BitNetRotaryEmbeddingNrC   c                    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)r   r    hasattrr   getr   max_position_embeddingsmax_seq_len_cachedoriginal_max_seq_lenrC   r   rope_init_fnattention_scalingregister_bufferr   original_inv_freq)r&   rC   devicer   r)   r+   r,   r      s   
zBitNetRotaryEmbedding.__init__c           
      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	j
|dd+ | |  dd}t	j||fdd	}| | j }| | j }	W d    n1 smw   Y  |j|jd
|	j|jd
fS )Nr   r.   r   mpscpuF)device_typeenabledr-   rS   )r0   )r   floatre   r;   r1   r   
isinstancer   strr"   autocastry   rU   r\   r   r]   r0   )
r&   rR   r^   inv_freq_expandedposition_ids_expandedr   freqsembr\   r]   r+   r+   r,   r9   *  s   0&zBitNetRotaryEmbedding.forwardrQ   )
r>   r?   r@   r   r    r"   no_gradr   r9   rA   r+   r+   r)   r,   r     s
    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 )BitNetPreTrainedModel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 )Nrl   )r4   stdg      ?)rC   initializer_ranger   r   rH   r$   datanormal_rE   zero_	Embeddingpadding_idxr   fill_)r&   rm   r   r+   r+   r,   _init_weightsJ  s   


z#BitNetPreTrainedModel._init_weightsN)r>   r?   r@   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   r+   r+   r+   r,   r   :  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 )BitNetModelrC   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   rC   r+   r,   
<listcomp>a  s    z(BitNetModel.__init__.<locals>.<listcomp>rF   r   F)r   r    pad_token_idr   
vocab_sizer   r   r'   embed_tokens
ModuleListrangenum_hidden_layerslayersr   rN   normr   
rotary_embgradient_checkpointing	post_initrP   r)   r   r,   r    Z  s   zBitNetModel.__init__c                 C      | j S rQ   r   r<   r+   r+   r,   get_input_embeddingsj     z BitNetModel.get_input_embeddingsc                 C   
   || _ d S rQ   r   r&   rp   r+   r+   r,   set_input_embeddingsm     
z BitNetModel.set_input_embeddingsN	input_idsrq   r^   r   inputs_embedsr   r   output_hidden_statesr   flash_attn_kwargsrd   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}t	|t
d tfsFtd|d u rO| |}|rX|d u rXt }|	d u rt|d urd| nd}tj|||jd  |jd}	|d u r}|	d}t| j |||	||d}|}| ||}|rd	nd }|rd	nd }| jd | j j D ]&}|r||f7 }||f||||||	|d
|
}|d }|r||d f7 }q| |}|r||f7 }t||r|nd ||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`.FzBThe `past_key_values` should be either a `Cache` object or `None`.r   r   )r   )rC   input_embedsrq   r   r   r^   r+   )rq   r^   r   r   r   r   r   )last_hidden_stater   r6   
attentions)rC   r   r   r   
ValueErrorr   rv   loggerwarning_oncer   r   r   r   r	   get_seq_lengthr"   aranger;   r   rY   r   r   r   r   r   r   )r&   r   rq   r^   r   r   r   r   r   r   r   past_seen_tokensr   r6   r   all_hidden_statesall_self_attnsdecoder_layerlayer_outputsr+   r+   r,   r9   p  s   

	
	


zBitNetModel.forward)	NNNNNNNNN)r>   r?   r@   r   r    r   r   r   r   r   r"   r   r   r   r   r   r   r   r   r9   rA   r+   r+   r)   r,   r   X  sL    	
r   c                   @   s   e Zd ZdS )KwargsForCausalLMN)r>   r?   r@   r+   r+   r+   r,   r    s    r  c                       s   e Zd ZdgZdZd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 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 )"BitNetForCausalLMzlm_head.weightNc                    s@   t  | t|| _|j| _tj|j|jdd| _| 	  d S )NFrD   )
r   r    r   r   r   r   rH   r'   lm_headr   rP   r)   r+   r,   r      s
   
zBitNetForCausalLM.__init__c                 C   s   | j jS rQ   r   r   r<   r+   r+   r,   r     s   z&BitNetForCausalLM.get_input_embeddingsc                 C   s   || j _d S rQ   r  r   r+   r+   r,   r     s   z&BitNetForCausalLM.set_input_embeddingsc                 C   r   rQ   r  r<   r+   r+   r,   get_output_embeddings  r   z'BitNetForCausalLM.get_output_embeddingsc                 C   r   rQ   r  )r&   new_embeddingsr+   r+   r,   set_output_embeddings  r   z'BitNetForCausalLM.set_output_embeddingsc                 C   r   rQ   r   )r&   decoderr+   r+   r,   set_decoder  r   zBitNetForCausalLM.set_decoderc                 C   r   rQ   r  r<   r+   r+   r,   get_decoder  r   zBitNetForCausalLM.get_decoderr   r   rq   r^   r   r   labelsr   r   r   r   logits_to_keepr}   rd   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 }d}|durX| 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?"
        ```N)	r   rq   r^   r   r   r   r   r   r   )logitsr  r   )lossr  r   r6   r  r+   )rC   r   r   r   r   r   r   slicer  loss_functionr   r   r   r6   r  )r&   r   rq   r^   r   r   r  r   r   r   r   r  r}   r   r6   slice_indicesr  r  r+   r+   r,   r9     s:   '
zBitNetForCausalLM.forward)NNNNNNNNNNr   )r>   r?   r@   _tied_weights_keys_tp_plan_pp_planr    r   r   r  r  r  r  r   r   r   r"   r   r   r   r   r   r   r   r   r  r   r9   rA   r+   r+   r)   r,   r    sf    		
r  )r  r   r   )Nr   )rl   )<typingr   r   r   r"   r   activationsr   cache_utilsr   r	   
generationr
   integrationsr   masking_utilsr   modeling_flash_attention_utilsr   modeling_layersr   modeling_outputsr   r   modeling_rope_utilsr   r   modeling_utilsr   r   processing_utilsr   utilsr   r   r   r   configuration_bitnetr   
get_loggerr>   r  Moduler   rB   rX   rb   r   r   rk   r   r   r   r   r   r   r   r  r  __all__r+   r+   r+   r,   <module>   sd   


I5"}l