o
    Zh	x                     @   s0  d dl mZ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&m'Z' ddl(m)Z) e& rd dl*m+Z+ ddl,m-Z- e'.e/Z0edG dd dej1Z2dej3de4dej3fddZ5	d;dej1dej3dej3d ej3d!eej3 d"e6d#e6fd$d%Z7d<d&d'Z8d(d) Z9G d*d+ d+ej1Z:G d,d- d-ej1Z;G d.d/ d/eZ<G d0d1 d1ej1Z=e$G d2d3 d3eZ>e$G d4d5 d5e>Z?G d6d7 d7ee#Z@e$G d8d9 d9e>eZAg d:ZBdS )=    )CallableOptionalTupleUnionN   )ACT2FN)CacheDynamicCache)GenerationMixin)use_kernel_forward_from_hub)AttentionMaskConverter)FlashAttentionKwargs)GradientCheckpointingLayer)BaseModelOutputWithPastCausalLMOutputWithPast)ROPE_INIT_FUNCTIONSdynamic_rope_update)ALL_ATTENTION_FUNCTIONSPreTrainedModel)Unpack)
LossKwargsauto_docstringcan_return_tupleis_torch_flex_attn_availablelogging   )Olmo2Config)	BlockMask)make_flex_block_causal_maskZRMSNormc                       s.   e Zd Zd fdd	Zdd Zdd Z  ZS )	Olmo2RMSNormư>c                    s&   t    tt|| _|| _dS )z;
        Olmo2RMSNorm is equivalent to T5LayerNorm
        N)super__init__nn	ParametertorchZonesweightvariance_epsilon)selfhidden_sizeeps	__class__ W/var/www/auris/lib/python3.10/site-packages/transformers/models/olmo2/modeling_olmo2.pyr"   &   s   

zOlmo2RMSNorm.__init__c                 C   sJ   |j }|tj}|djddd}|t|| j  }| j| |S )N   T)Zkeepdim)	dtypetor%   float32powmeanZrsqrtr'   r&   )r(   hidden_statesZinput_dtypeZvariancer-   r-   r.   forward.   s
   zOlmo2RMSNorm.forwardc                 C   s   t | jj d| j S )Nz, eps=)tupler&   shaper'   r(   r-   r-   r.   
extra_repr5   s   zOlmo2RMSNorm.extra_repr)r    )__name__
__module____qualname__r"   r7   r;   __classcell__r-   r-   r+   r.   r   $   s    r   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)r9   expandreshape)r6   r@   batchnum_key_value_headsslenhead_dimr-   r-   r.   	repeat_kv9   s
   0rH           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   r0   )dimr1   )ptrainingr   )rH   num_key_value_groupsr%   matmul	transposer9   r#   Z
functionalZsoftmaxr3   r2   r1   rP   rT   
contiguous)rJ   rK   rL   rM   rN   rO   rP   kwargs
key_statesvalue_statesattn_weightscausal_maskattn_outputr-   r-   r.   eager_attention_forwardE   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.
        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.
    )r1   	unsqueezerotate_halfr2   )
qkcossinposition_idsZunsqueeze_dimZq_typeZk_typeZq_embedZk_embedr-   r-   r.   apply_rotary_pos_emb_   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..Nr0   r/   rR   )r9   r%   cat)xx1Zx2r-   r-   r.   ra   {   s   ra   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j	eej	 ee
ej	  f fddZ  ZS )Olmo2Attentionz=Multi-headed attention from 'Attention Is All You Need' paperNconfig	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 |j| _t|j| j |j| _d S )NrG   g      Tbias)r!   r"   rm   rn   getattrr)   Znum_attention_headsrG   rE   rU   rO   attention_dropoutZ	is_causalr#   LinearZattention_biasq_projk_projv_projo_projr   rms_norm_epsq_normk_normr(   rm   rn   r+   r-   r.   r"      s,   
zOlmo2Attention.__init__r6   position_embeddingsrN   past_key_valuecache_positionrA   c                 K   s`  |j d d }g |d| jR }| | |}	| | |}
| |}|	|dd}	|
|dd}
||dd}|\}}t	|	|
||\}	}
|d urc|||d}|
|
|| j|\}
}t}| jjdkr| jjdkr}|ddr}td	 nt| jj }|| |	|
||f| jsd
n| j| jd|\}}|jg |dR   }| |}||fS )Nr0   r   r/   )re   rd   r~   eagersdpaoutput_attentionsFz`torch.nn.functional.scaled_dot_product_attention` does not support `output_attentions=True`. Falling back to eager attention. This warning can be removed using the argument `attn_implementation="eager"` when loading the model.rI   )rP   rO   )r9   rG   ry   rt   rz   ru   rv   viewrW   rg   updatern   r_   rm   _attn_implementationgetloggerwarning_oncer   rT   rr   rO   rC   rX   rw   )r(   r6   r|   rN   r}   r~   rY   Zinput_shapeZhidden_shapeZquery_statesrZ   r[   rd   re   Zcache_kwargsZattention_interfacer^   r\   r-   r-   r.   r7      sF   	


zOlmo2Attention.forwardN)NN)r<   r=   r>   __doc__r   r   intr"   r%   Tensorr   r   
LongTensorr7   r?   r-   r-   r+   r.   rl      s$    rl   c                       s$   e Zd Z fddZdd Z  ZS )Olmo2MLPc                    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 NFro   )r!   r"   rm   r)   Zintermediate_sizer#   rs   	gate_projup_proj	down_projr   Z
hidden_actact_fnr(   rm   r+   r-   r.   r"      s   
zOlmo2MLP.__init__c                 C   s$   |  | | || | }|S r   )r   r   r   r   )r(   rj   r   r-   r-   r.   r7      s    zOlmo2MLP.forward)r<   r=   r>   r"   r7   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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jeeejejf  f fddZ  ZS )Olmo2DecoderLayerrm   rn   c                    sR   t    |j| _t||d| _t|| _t|j|jd| _	t|j|jd| _
d S )N)rm   rn   r*   )r!   r"   r)   rl   	self_attnr   mlpr   rx   post_attention_layernormpost_feedforward_layernormr{   r+   r-   r.   r"      s   

zOlmo2DecoderLayer.__init__NFr6   rN   rf   r}   r   	use_cacher~   r|   rA   c	                 K   st   |}
| j d||||||||d|	\}}| |}|
| }|}
| |}| |}|
| }|f}|r8||f7 }|S )N)r6   rN   rf   r}   r   r   r~   r|   r-   )r   r   r   r   )r(   r6   rN   rf   r}   r   r   r~   r|   rY   ZresidualZself_attn_weightsoutputsr-   r-   r.   r7      s.   	




zOlmo2DecoderLayer.forward)NNNFFNN)r<   r=   r>   r   r   r"   r%   r   r   r   r   boolr   FloatTensorr7   r?   r-   r-   r+   r.   r      s8    	r   c                       s8   e Zd Zddef fddZe edd Z  Z	S )Olmo2RotaryEmbeddingNrm   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   r   r   Zmax_position_embeddingsZmax_seq_len_cachedZoriginal_max_seq_lenrm   r   Zrope_init_fnattention_scalingZregister_bufferr   Zoriginal_inv_freq)r(   rm   devicer   r+   r-   r.   r"     s   
zOlmo2RotaryEmbedding.__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 }	||	fW  d    S 1 sqw   Y  d S )
Nr   r0   r   ZmpscpuF)device_typeenabledr/   rh   )r   floatrB   r9   r2   r   
isinstancer   strr%   ZautocastrW   ri   rd   r   re   )
r(   rj   rf   Zinv_freq_expandedZposition_ids_expandedr   ZfreqsZembrd   re   r-   r-   r.   r7   *  s   0&$zOlmo2RotaryEmbedding.forwardr   )
r<   r=   r>   r   r"   r%   Zno_gradr   r7   r?   r-   r-   r+   r.   r     s
    r   c                   @   sH   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d ZdS )Olmo2PreTrainedModel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 )NrI   )r5   stdg      ?)rm   Zinitializer_ranger   r#   rs   r&   dataZnormal_rp   Zzero_	Embeddingpadding_idxr   Zfill_)r(   rJ   r   r-   r-   r.   _init_weightsH  s   


z"Olmo2PreTrainedModel._init_weightsN)r<   r=   r>   r   Zconfig_classZbase_model_prefixZsupports_gradient_checkpointingZ_no_split_modulesZ_skip_keys_device_placementZ_supports_flash_attn_2Z_supports_sdpaZ_supports_flex_attnZ_supports_cache_classZ_supports_quantized_cacheZ_supports_static_cacheZ_supports_attention_backendr   r-   r-   r-   r.   r   9  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	d"d
ee
jdf de
jde
jdedef
ddZed
e
jdedede
jde
jdefdd Z  ZS )#
Olmo2Modelrm   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   ).0rn   rm   r-   r.   
<listcomp>_  s    z'Olmo2Model.__init__.<locals>.<listcomp>r   r   F)r!   r"   Zpad_token_idr   
vocab_sizer#   r   r)   embed_tokensZ
ModuleListrangenum_hidden_layerslayersr   rx   normr   
rotary_embgradient_checkpointing	post_initr   r+   r   r.   r"   X  s   zOlmo2Model.__init__c                 C      | j S r   r   r:   r-   r-   r.   get_input_embeddingsh     zOlmo2Model.get_input_embeddingsc                 C   
   || _ d S r   r   r(   rM   r-   r-   r.   set_input_embeddingsk     
zOlmo2Model.set_input_embeddingsN	input_idsrN   rf   r   inputs_embedsr   r   output_hidden_statesr~   flash_attn_kwargsrA   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}| |||	||}|}| ||}|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   r-   )rN   rf   r}   r   r   r~   r|   )last_hidden_stater   r6   
attentions)rm   r   r   r   
ValueErrorr   rT   r   r   r   r   r   r   r	   get_seq_lengthr%   aranger9   r   r`   _update_causal_maskr   r   r   r   r   )r(   r   rN   rf   r   r   r   r   r   r~   r   past_seen_tokensr]   r6   r|   Zall_hidden_statesZall_self_attnsZdecoder_layerZlayer_outputsr-   r-   r.   r7   n  sx   



	


zOlmo2Model.forwardFr   input_tensorc                 C   s:  | j jdkr|d ur|dk r|S d S | j jdkr&t|tjr$t|}|S |d ur.| nd}|d ur7|jnd}| j jdkrO|sO|sOt	j
|||| jdrOd S |j}|jd }	|r^| }
nt|tjri|jd	 n||	 d }
| j||	|
|||jd d
}| j jdkr|d ur|jjdv r|st|j}t	||}|S )NZflash_attention_2rI   Zflex_attentionr   Fr   )r   Zpast_key_values_lengthZis_trainingr   r0   )sequence_lengthtarget_lengthr1   r~   
batch_size)cudaZxpuZnpu)rm   r   anyr   r%   r   r   r   Zis_compileabler   Z_ignore_causal_mask_sdparT   r1   r9   Zget_max_cache_shape5_prepare_4d_causal_attention_mask_with_cache_positionr   r   finfominZ_unmask_unattended)r(   rN   r   r~   r   r   r   Zusing_compilable_cacher1   r   r   r]   	min_dtyper-   r-   r.   r     sT   




zOlmo2Model._update_causal_maskr   r   r1   r   c                 K   sD  | dur|   dkr| }|S t|j}tj||f|||jd}|dkr+tj|dd}|tj||jd|ddk9 }|ddddddf 	|ddd}| dur|
 }| jd }	|ddddddd|	f | ddddddf |j }
|
dk}
|ddddddd|	f |
||ddddddd|	f< |S )	aM  
        Creates a causal 4D mask of shape `(batch_size, 1, query_length, key_value_length)` from a 2D mask of shape
        `(batch_size, key_value_length)`, or if the input `attention_mask` is already 4D, do nothing.

        Args:
            attention_mask (`torch.Tensor`):
                A 2D attention mask of shape `(batch_size, key_value_length)` or a 4D attention mask of shape
                `(batch_size, 1, query_length, key_value_length)`.
            sequence_length (`int`):
                The sequence length being processed.
            target_length (`int`):
                The target length: when generating with static cache, the mask should be as long as the static cache,
                to account for the 0 padding, the part of the cache that is not filled yet.
            dtype (`torch.dtype`):
                The dtype to use for the 4D attention mask.
            cache_position (`torch.Tensor`):
                Indices depicting the position of the input sequence tokens in the sequence.
            batch_size (`torch.Tensor`):
                Batch size.
        N   )Z
fill_valuer1   r   r   )Zdiagonalr   r0   r   )rR   r%   r   r   fullr   Ztriur   rC   rB   cloner9   r2   Zmasked_fill)rN   r   r   r1   r~   r   rY   r]   r   Zmask_lengthZpadding_maskr-   r-   r.   r     s,    $
6  z@Olmo2Model._prepare_4d_causal_attention_mask_with_cache_position)	NNNNNNNNN)F)r<   r=   r>   r   r"   r   r   r   r   r   r%   r   r   r   r   r   r   r   r   r7   r   r   staticmethodr   r1   r   r?   r-   r-   r+   r.   r   V  s    	
d
Dr   c                   @   s   e Zd ZdS )KwargsForCausalLMN)r<   r=   r>   r-   r-   r-   r.   r   J  s    r   c                       s
  e Zd ZdgZddiZddgdgfiZ fddZdd	 Zd
d Zdd Z	dd Z
dd Zdd Zee											d%deej deej deej dee deej deej dee dee dee deej d eeejf d!ee d"efd#d$Z  ZS )&Olmo2ForCausalLMzlm_head.weightlm_headZcolwise_repr6   logitsc                    s@   t  | t|| _|j| _tj|j|jdd| _| 	  d S r   )
r!   r"   r   r   r   r#   rs   r)   r   r   r   r+   r-   r.   r"   S  s
   
zOlmo2ForCausalLM.__init__c                 C   s   | j jS r   r   r   r:   r-   r-   r.   r   \  s   z%Olmo2ForCausalLM.get_input_embeddingsc                 C   s   || j _d S r   r   r   r-   r-   r.   r   _  s   z%Olmo2ForCausalLM.set_input_embeddingsc                 C   r   r   r   r:   r-   r-   r.   get_output_embeddingsb  r   z&Olmo2ForCausalLM.get_output_embeddingsc                 C   r   r   r   )r(   Znew_embeddingsr-   r-   r.   set_output_embeddingse  r   z&Olmo2ForCausalLM.set_output_embeddingsc                 C   r   r   r   )r(   decoderr-   r-   r.   set_decoderh  r   zOlmo2ForCausalLM.set_decoderc                 C   r   r   r   r:   r-   r-   r.   get_decoderk  r   zOlmo2ForCausalLM.get_decoderNr   r   rN   rf   r   r   labelsr   r   r   r~   logits_to_keeprY   rA   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 )at  
        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, Olmo2ForCausalLM

        >>> model = Olmo2ForCausalLM.from_pretrained("meta-olmo2/Olmo2-2-7b-hf")
        >>> tokenizer = AutoTokenizer.from_pretrained("meta-olmo2/Olmo2-2-7b-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."
        ```N)	r   rN   rf   r   r   r   r   r   r~   )r   r   r   )lossr   r   r6   r   r-   )rm   r   r   r   r   r   r   slicer   Zloss_functionr   r   r   r6   r   )r(   r   rN   rf   r   r   r   r   r   r   r~   r   rY   r   r6   Zslice_indicesr   r   r-   r-   r.   r7   n  s:   '
zOlmo2ForCausalLM.forward)NNNNNNNNNNr   )r<   r=   r>   Z_tied_weights_keysZ_tp_planZ_pp_planr"   r   r   r   r   r   r   r   r   r   r%   r   r   r   r   r   r   r   r   r   r   r7   r?   r-   r-   r+   r.   r   M  sf    		
r   )r   r   r   )rI   )Nr   )Ctypingr   r   r   r   r%   Ztorch.nnr#   Zactivationsr   Zcache_utilsr   r	   Z
generationr
   Zintegrationsr   Zmodeling_attn_mask_utilsr   Zmodeling_flash_attention_utilsr   Zmodeling_layersr   Zmodeling_outputsr   r   Zmodeling_rope_utilsr   r   Zmodeling_utilsr   r   Zprocessing_utilsr   utilsr   r   r   r   r   Zconfiguration_olmo2r   Z!torch.nn.attention.flex_attentionr   Zintegrations.flex_attentionr   Z
get_loggerr<   r   Moduler   r   r   rH   r   r_   rg   ra   rl   r   r   r   r   r   r   r   __all__r-   r-   r-   r.   <module>   sl   


R4! tl