o
    eiPW                     @   s(  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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$m%Z%m&Z& ddl'm(Z(m)Z)m*Z* ddl+m,Z, ddl-m.Z. G dd dejj/Z0G dd dej/Z1edd;ddZ2dej3de4dej3fdd Z5	!d<d"ej/d#ej3d$ej3d%ej3d&ej3dB d'e6d(e6d)e#e% fd*d+Z7d,d- Z8ee2G d.d/ d/ej/Z9G d0d1 d1ej/Z:G d2d3 d3eZ;e&G d4d5 d5e!Z<e&G d6d7 d7e<Z=e&G d8d9 d9e<eZ>g d:Z?dS )=    N)Callable)Optional   )initialization)ACT2FN)CacheDynamicCache)GenerationMixin)use_kernel_func_from_hubuse_kernelized_func)create_causal_mask)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   )NanoChatConfigc                       s<   e Zd Zddef fddZdd Zdd Zd	d
 Z  ZS )NanoChatRMSNormư>epsc                    s   t    || _d S N)super__init__r   )selfr   	__class__ l/home/ubuntu/transcripts/venv/lib/python3.10/site-packages/transformers/models/nanochat/modeling_nanochat.pyr"   .   s   

zNanoChatRMSNorm.__init__c                 C   s$   |t |djddd| j  S )N   T)keepdim)torchrsqrtpowmeanr   r#   xr&   r&   r'   _norm2   s   $zNanoChatRMSNorm._normc                 C   s   |  | |S r    )r1   floattype_asr/   r&   r&   r'   forward5   s   zNanoChatRMSNorm.forwardc                 C   s   d| j  S )Nzeps=r   )r#   r&   r&   r'   
extra_repr8   s   zNanoChatRMSNorm.extra_repr)r   )	__name__
__module____qualname__r2   r"   r1   r4   r6   __classcell__r&   r&   r$   r'   r   -   s
    r   c                       s~   e Zd ZU ejed< ddef fddZe			ddedB de	d de
dB d	ed
ef fddZe edd Z  ZS )NanoChatRotaryEmbeddinginv_freqNconfigc                    s   t    |j| _|j| _|| _| jjd | _| j}| jdkr$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_lenr=   rope_parametersr>   compute_default_rope_parametersr   attention_scalingregister_bufferclone)r#   r=   devicerope_init_fnr<   r$   r&   r'   r"   ?   s   


z NanoChatRotaryEmbedding.__init__rJ   ztorch.deviceseq_lenreturnz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_thetahead_dimNg      ?r   r(   dtype)rJ   rQ   )	rE   getattrhidden_sizenum_attention_headsr+   arangeint64tor2   )r=   rJ   rL   basedimattention_factorr<   r&   r&   r'   rF   O   s   
&z7NanoChatRotaryEmbedding.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   r)   r   mpscpuF)device_typeenabledr(   rY   rP   )r<   r2   expandshaperW   rJ   
isinstancetypestrr   	transposer+   catcosrG   sinrQ   )
r#   r0   position_idsinv_freq_expandedposition_ids_expandedr]   freqsembrg   rh   r&   r&   r'   r4   m   s   0&zNanoChatRotaryEmbedding.forwardr    )NNN)r7   r8   r9   r+   Tensor__annotations__r   r"   staticmethodr   inttupler2   rF   no_gradr   r4   r:   r&   r&   r$   r'   r;   <   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.
    )	unsqueezerotate_half)qkrg   rh   unsqueeze_dimq_embedk_embedr&   r&   r'   apply_rotary_pos_emb}   s
   

r|   hidden_statesn_reprM   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)ra   r`   reshape)r}   r~   batchnum_key_value_headsslenrO   r&   r&   r'   	repeat_kv   s
   0r           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 )Nr(   r   r)   )rY   rQ   )ptrainingr   )r   num_key_value_groupsr+   matmulre   nn
functionalsoftmaxfloat32rW   rQ   r   r   
contiguous)r   r   r   r   r   r   r   r   
key_statesvalue_statesattn_weightsattn_outputr&   r&   r'   eager_attention_forward   s   
r   c                 C   sH   | dd| j d d f }| d| j d d df }tj|| fddS )zJRotates half the hidden dims of the input with flipped signs for NanoChat..Nr)   r(   r_   )ra   r+   rf   )r0   x1x2r&   r&   r'   rv      s   rv   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B 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 )NanoChatAttentionz=Multi-headed attention from 'Attention Is All You Need' paperr=   	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d| _t|jd| _d S )NrO   g      Tbiasr5   )r!   r"   r=   r   rR   rS   rT   rO   r   r   r   attention_dropout	is_causalr   Linearattention_biasq_projk_projv_projo_projr   rms_norm_epsq_normk_normr#   r=   r   r$   r&   r'   r"      s,   
zNanoChatAttention.__init__Nr}   position_embeddingsr   past_key_valuescache_positionr   rM   c                 K   s,  |j d d }g |d| jR }| ||dd}	| ||dd}
| ||dd}|\}}t|	|
||\}	}
| |	}	| 	|
}
|d ura|||d}|
|
|| j|\}
}t| jjt}|| |	|
||f| jsudn| j| jd|\}}|jg |dR   }| |}||fS )Nr)   r   r(   )rh   rg   r   r   )r   r   )ra   rO   r   viewre   r   r   r|   r   r   updater   r   get_interfacer=   _attn_implementationr   r   r   r   r   r   r   )r#   r}   r   r   r   r   r   input_shapehidden_shapequery_statesr   r   rg   rh   cache_kwargsattention_interfacer   r   r&   r&   r'   r4      s<   	



zNanoChatAttention.forward)NNNN)r7   r8   r9   __doc__r   rq   r"   r+   rn   rr   r   
LongTensorr   r   r4   r:   r&   r&   r$   r'   r      s,    r   c                       s2   e Zd Z fddZdejdejfddZ  ZS )NanoChatMLPc                    sL   t    || _t|j | _tj|j|j	dd| _
tj|j	|jdd| _d S NFr   )r!   r"   r=   r   
hidden_actactivation_fnr   r   rS   intermediate_sizefc1fc2r#   r=   r$   r&   r'   r"     s
   
zNanoChatMLP.__init__r}   rM   c                 C   s"   |  |}| |}| |}|S r    )r   r   r   )r#   r}   r&   r&   r'   r4     s   


zNanoChatMLP.forward)r7   r8   r9   r"   r+   rn   r4   r:   r&   r&   r$   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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 )NanoChatDecoderLayerr=   r   c                    sJ   t    |j| _t||d| _t|| _t|jd| _	t|jd| _
d S )N)r=   r   r5   )r!   r"   rS   r   	self_attnr   mlpr   r   input_layernormpost_attention_layernormr   r$   r&   r'   r"   !  s   

zNanoChatDecoderLayer.__init__NFr}   r   ri   r   	use_cacher   r   r   rM   c              
   K   s^   |}	|  |}| jd|||||||d|\}}
|	| }|}	| |}| |}|	| }|S )N)r}   r   ri   r   r   r   r   r&   )r   r   r   r   )r#   r}   r   ri   r   r   r   r   r   residual_r&   r&   r'   r4   ,  s&   




zNanoChatDecoderLayer.forward)NNNFNN)r7   r8   r9   r   rq   r"   r+   rn   r   r   boolrr   r   r   r4   r:   r&   r&   r$   r'   r      s6    	
r   c                       sd   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ejdd	f fd
dZ  ZS )NanoChatPreTrainedModelr=   modelTr   r   )r}   
attentionsr   rM   Nc                    sH   t  | t|tr"tj|jjd| jj	t
d| jj  d d S d S )Nr   r(   )r.   std)r!   _init_weightsrb   r   initnormal_r   weightr=   initializer_rangemathsqrtnum_hidden_layers)r#   r   r$   r&   r'   r   `  s   

z%NanoChatPreTrainedModel._init_weights)r7   r8   r9   r   ro   base_model_prefixsupports_gradient_checkpointing_no_split_modules_skip_keys_device_placement_supports_flash_attn_supports_sdpa_supports_flex_attn_can_compile_fullgraph_supports_attention_backendr   r   _can_record_outputsr   Moduler   r:   r&   r&   r$   r'   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 )NanoChatModelr=   c                    s~   t     j| _ j| _t j j| j| _t	 fddt
 jD | _t jd| _t d| _d| _|   d S )Nc                    s   g | ]}t  |qS r&   )r   ).0r   r=   r&   r'   
<listcomp>s  s    z*NanoChatModel.__init__.<locals>.<listcomp>r5   r   F)r!   r"   pad_token_idpadding_idx
vocab_sizer   	EmbeddingrS   embed_tokens
ModuleListranger   layersr   r   normr;   
rotary_embgradient_checkpointing	post_initr   r$   r   r'   r"   l  s   zNanoChatModel.__init__N	input_idsr   ri   r   inputs_embedsr   r   r   rM   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|}qg| |}t||d	S )
Nz:You must specify exactly one of input_ids or inputs_embedsr   r   r   )rJ   )r=   r   r   r   r   ri   )ri   )r   r   ri   r   r   )last_hidden_stater   )
ValueErrorr   r   r=   get_seq_lengthr+   rU   ra   rJ   ru   r   r   r   r   r   r   )r#   r   r   ri   r   r   r   r   r   past_seen_tokenscausal_maskr}   r   decoder_layerr&   r&   r'   r4   }  sP   

	


zNanoChatModel.forward)NNNNNNN)r7   r8   r9   r   r"   r   r   r   r+   r   rn   r   FloatTensorr   r   r   r   r4   r:   r&   r&   r$   r'   r   j  s>    	
r   c                       s   e Zd ZddiZddiZddgdgfiZ fddZee																	
dde	j
d	B de	jd	B de	j
d	B ded	B de	jd	B de	j
d	B ded	B de	j
d	B dee	jB dee defddZ  ZS )NanoChatForCausalLMzlm_head.weightzmodel.embed_tokens.weightlm_headcolwise_gather_outputr}   logitsc                    s@   t  | t|| _|j| _tj|j|jdd| _| 	  d S r   )
r!   r"   r   r   r   r   r   rS   r   r   r   r$   r&   r'   r"     s
   
zNanoChatForCausalLM.__init__Nr   r   r   ri   r   r   labelsr   r   logits_to_keepr   rM   c
              
   K   s   | j d|||||||d|
}|j}t|	trt|	 dn|	}| |dd|ddf }| jjdurE|| jj }t	|}|| jj }d}|durW| j
||| jfi |
}t|||j|j|jdS )ak  
        Example:

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

        >>> model = AutoModelForCausalLM.from_pretrained("karpathy/nanochat-d32")

        >>> tokenizer = AutoTokenizer.from_pretrained("karpathy/nanochat-d32")

        >>> conversation = [
                {"role": "user", "content": "What is the capital of France?"},
            ]

        >>> inputs = tokenizer.apply_chat_template(
                conversation, add_generation_prompt=True, tokenize=True, return_dict=True, return_tensors="pt"
            ).to(device)

        >>> with torch.no_grad():
        >>>     outputs = model.generate(**inputs, max_new_tokens=64, do_sample=False)

        >>> generated_tokens = outputs[0, inputs["input_ids"].shape[1] :]
        >>> output = tokenizer.decode(generated_tokens, skip_special_tokens=True)
        ```)r   r   ri   r   r   r   r   N)lossr   r   r}   r   r&   )r   r   rb   rq   slicer   r=   final_logit_softcappingr+   tanhloss_functionr   r   r   r}   r   )r#   r   r   ri   r   r   r   r   r   r  r   outputsr}   slice_indicesr   r  r&   r&   r'   r4     s8   (
zNanoChatForCausalLM.forward)	NNNNNNNNr   )r7   r8   r9   _tied_weights_keys_tp_plan_pp_planr"   r   r   r+   r   rn   r   r   r   rq   r   r   r   r4   r:   r&   r&   r$   r'   r     sN    		
r   )r   r   r   )r   )r   )@r   collections.abcr   typingr   r+   torch.nnr    r   r   activationsr   cache_utilsr   r   
generationr	   integrationsr
   r   masking_utilsr   modeling_layersr   modeling_outputsr   r   modeling_rope_utilsr   r   modeling_utilsr   r   processing_utilsr   utilsr   r   utils.genericr   r   r   utils.output_capturingr   configuration_nanochatr   r   r   r;   r|   rn   rq   r   r2   r   rv   r   r   r   r   r   r   __all__r&   r&   r&   r'   <module>   sn   A
M.QW