o
    ei+U                     @   s  d dl mZ d dlmZ d dlZd dlmZ d dlm  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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, G dd dej-Z.G dd dej-Z/G dd dej-Z0dd Z1dej2de3dej2fdd Z4	!d9d"ej-d#ej2d$ej2d%ej2d&ej2dB d'e5d(e5d)e!e# fd*d+Z6d:d,d-Z7ee7G d.d/ d/ej-Z8G d0d1 d1eZ9e$G d2d3 d3eZ:e$G d4d5 d5e:Z;e$G d6d7 d7e:eZ<g d8Z=dS );    )Callable)OptionalN   )ACT2FN)CacheDynamicCache)GenerationMixin)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   )
OlmoConfigc                       s@   e Zd ZdZdeddf fddZdejdejfdd	Z  Z	S )
OlmoLayerNormz/LayerNorm but with no learnable weight or bias.hidden_sizereturnNc                    s   t    |f| _d S N)super__init__normalized_shape)selfr   	__class__ d/home/ubuntu/transcripts/venv/lib/python3.10/site-packages/transformers/models/olmo/modeling_olmo.pyr    4   s   
zOlmoLayerNorm.__init__hidden_statesc                 C   s,   |j }tj|jtjd| jd d dd|S )Ndtypegh㈵>)eps)r)   F
layer_normtotorchfloat32r!   )r"   r'   
orig_dtyper%   r%   r&   forward8   s    zOlmoLayerNorm.forward)
__name__
__module____qualname____doc__intr    r.   Tensorr1   __classcell__r%   r%   r#   r&   r   1   s    r   c                       s$   e Zd Z fddZdd Z  ZS )OlmoMLPc                    sr   t    || _|j| _|j| _tj| j| jdd| _tj| j| jdd| _tj| j| jdd| _	t
|j | _d S NFbias)r   r    configr   intermediate_sizennLinear	gate_projup_proj	down_projr   
hidden_actact_fnr"   r=   r#   r%   r&   r    @   s   
zOlmoMLP.__init__c                 C   s$   |  | | || | }|S r   )rC   rE   rA   rB   )r"   xrC   r%   r%   r&   r1   J   s    zOlmoMLP.forward)r2   r3   r4   r    r1   r8   r%   r%   r#   r&   r9   ?   s    
r9   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 )OlmoRotaryEmbeddinginv_freqNr=   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defaultrI   F)
persistentoriginal_inv_freq)r   r    max_position_embeddingsmax_seq_len_cachedoriginal_max_seq_lenr=   rope_parametersrJ   compute_default_rope_parametersr   attention_scalingregister_bufferclone)r"   r=   devicerope_init_fnrI   r#   r%   r&   r    R   s   


zOlmoRotaryEmbedding.__init__rV   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_thetahead_dimNg      ?r      r(   )rV   r)   )	rQ   getattrr   num_attention_headsr.   arangeint64r-   float)r=   rV   rX   basedimattention_factorrI   r%   r%   r&   rR   b   s   
&z3OlmoRotaryEmbedding.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    ||	fS 1 sow   Y  ||	fS )
Nr   r   mpscpuF)device_typeenabledr[   rb   )rI   r`   expandshaper-   rV   
isinstancetypestrr   	transposer.   catcosrS   sin)
r"   rG   position_idsinv_freq_expandedposition_ids_expandedrg   freqsembrq   rr   r%   r%   r&   r1      s   0&
zOlmoRotaryEmbedding.forwardr   )NNN)r2   r3   r4   r.   r7   __annotations__r   r    staticmethodr   r6   tupler`   rR   no_gradr   r1   r8   r%   r%   r#   r&   rH   O   s&   
 

rH   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..Nrd   r[   ri   )rk   r.   rp   )rG   x1x2r%   r%   r&   rotate_half   s   r~   r'   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)rk   rj   reshape)r'   r   batchnum_key_value_headsslenrZ   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   rd   )rb   r)   )ptrainingr   )r   num_key_value_groupsr.   matmulro   r?   
functionalsoftmaxr/   r-   r)   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   s^   | j |j }}||}||}| | 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.
    )r)   	unsqueezer~   r-   )	qkrq   rr   unsqueeze_dimq_typek_typeq_embedk_embedr%   r%   r&   apply_rotary_pos_emb   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jejdB f fddZ  ZS )OlmoAttentionz=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| _d S )NrZ   g      Tr;   )r   r    r=   r   r\   r   r]   rZ   r   r   r   attention_dropout	is_causalr?   r@   attention_biasq_projk_projv_projo_projr"   r=   r   r#   r%   r&   r       s(   
zOlmoAttention.__init__Nr'   position_embeddingsr   past_key_valuescache_positionr   c                 K   sx  |j d d }g |d| jR }| |}	| |}
| |}| jjd urJ|	j| jj | jjd |
j| jj | jjd |j| jj | jjd |	|	dd}	|
|	dd}
||	dd}|\}}t
|	|
||\}	}
|d ur|||d}||
|| j|\}
}t| jjt}|| |	|
||f| jsdn| j| jd|\}}|jg |dR   }| |}||fS )Nrd   )minmaxr   r[   )rr   rq   r   r   )r   r   )rk   rZ   r   r   r   r=   clip_qkvclamp_viewro   r   updater   r   get_interface_attn_implementationr   r   r   r   r   r   r   )r"   r'   r   r   r   r   r   input_shapehidden_shapequery_statesr   r   rq   rr   cache_kwargsattention_interfacer   r   r%   r%   r&   r1      sF   	




zOlmoAttention.forward)NN)r2   r3   r4   r5   r   r6   r    r.   r7   rz   r   
LongTensorr1   r8   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 )OlmoDecoderLayerr=   r   c                    sF   t    |j| _t||d| _t|| _t|j| _t|j| _	d S )N)r=   r   )
r   r    r   r   	self_attnr9   mlpr   input_layernormpost_attention_layernormr   r#   r%   r&   r    &  s   

zOlmoDecoderLayer.__init__NFr'   r   rs   r   	use_cacher   r   r   r   c              
   K   s^   |}	|  |}| jd|||||||d|\}}
|	| }|}	| |}| |}|	| }|S )N)r'   r   rs   r   r   r   r   r%   )r   r   r   r   )r"   r'   r   rs   r   r   r   r   r   residual_r%   r%   r&   r1   /  s&   




zOlmoDecoderLayer.forward)NNNFNN)r2   r3   r4   r   r6   r    r.   r7   r   r   boolrz   r   r   r1   r8   r%   r%   r#   r&   r   %  s6    	
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 )OlmoPreTrainedModelr=   modelTr   r   )r'   
attentionsN)r2   r3   r4   r   rx   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%   r&   r   Q  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 )	OlmoModelr=   c                    s|   t     j| _ j| _t j j| j| _t	 fddt
 jD | _t j| _t d| _d| _|   d S )Nc                    s   g | ]}t  |qS r%   )r   ).0r   r=   r%   r&   
<listcomp>m  s    z&OlmoModel.__init__.<locals>.<listcomp>r   F)r   r    pad_token_idpadding_idx
vocab_sizer?   	Embeddingr   embed_tokens
ModuleListrangenum_hidden_layerslayersr   normrH   
rotary_embgradient_checkpointing	post_initrF   r#   r   r&   r    f  s   zOlmoModel.__init__N	input_idsr   rs   r   inputs_embedsr   r   r   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   )rV   )r=   r   r   r   r   rs   )rs   )r   r   rs   r   r   r   )last_hidden_stater   )
ValueErrorr   r   r=   get_seq_lengthr.   r^   rk   rV   r   r
   r   r   r   r   r   )r"   r   r   rs   r   r   r   r   r   past_seen_tokenscausal_maskr'   r   decoder_layerr%   r%   r&   r1   v  sP   

	
zOlmoModel.forward)NNNNNNN)r2   r3   r4   r   r    r   r   r   r.   r   r7   r   FloatTensorr   r   r   r   r1   r8   r%   r%   r#   r&   r   d  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 )OlmoForCausalLMz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@   r   r   r   rF   r#   r%   r&   r      s
   
zOlmoForCausalLM.__init__Nr   r   r   rs   r   r   labelsr   r   logits_to_keepr   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  
        Example:

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

        >>> model = OlmoForCausalLM.from_pretrained("meta-olmo/Olmo-2-7b-hf")
        >>> tokenizer = AutoTokenizer.from_pretrained("meta-olmo/Olmo-2-7b-hf")

        >>> 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."
        ```)r   r   rs   r   r   r   r   N)r   r   r   )lossr   r   r'   r   r%   )r   r   rl   r6   slicer   loss_functionr=   r   r   r   r'   r   )r"   r   r   rs   r   r   r   r   r   r   r   outputsr'   slice_indicesr   r   r%   r%   r&   r1     s0    zOlmoForCausalLM.forward)	NNNNNNNNr   )r2   r3   r4   _tied_weights_keys_tp_plan_pp_planr    r   r   r.   r   r7   r   r   r   r6   r   r   r   r1   r8   r%   r%   r#   r&   r     sN    		
r   )r   r   r   )r   )r   )>collections.abcr   typingr   r.   torch.nnr?   torch.nn.functionalr   r+   activationsr   cache_utilsr   r   
generationr   integrationsr	   masking_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_olmor   Moduler   r9   rH   r~   r7   r6   r   r`   r   r   r   r   r   r   r   __all__r%   r%   r%   r&   <module>   sj   @

O,PK