o
    eiW                     @   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mZ ddlmZ ddlmZmZmZ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$m%Z% ddl&m'Z'm(Z(m)Z) ddl*m+Z+m,Z, ddl-m.Z. ddl/m0Z0 dd Z1eddHddZ2dej3de4dej3fddZ5	 dId!ej6d"ej3d#ej3d$ej3d%ej3dB d&e7d'e7d(e%e' fd)d*Z8d+ej3d,e7d-e4dej3fd.d/Z9ee2G d0d1 d1ej6Z:G d2d3 d3ej6Z;ed4G d5d6 d6ej6Z<G d7d8 d8eZ=e(G d9d: d:e#Z>G d;d< d<ej6Z?e(G d=d> d>e>Z@e(G d?d@ d@e>eZAG dAdB dBee>ZBG dCdD dDee>ZCG dEdF dFee>ZDg dGZEdS )J    )Callable)OptionalN)nn   )ACT2FN)CacheDynamicCache)GenerationMixin)use_kernel_forward_from_hubuse_kernel_func_from_hubuse_kernelized_func)create_causal_mask!create_sliding_window_causal_mask)FlashAttentionKwargs)GenericForQuestionAnswering GenericForSequenceClassificationGenericForTokenClassification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   )Ministral3Configc                 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..N   dim)shapetorchcat)xx1x2 r-   p/home/ubuntu/transcripts/venv/lib/python3.10/site-packages/transformers/models/ministral3/modeling_ministral3.pyrotate_half#   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-   r.   apply_rotary_pos_emb*   s
   

r9   hidden_states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)r:   r;   batchnum_key_value_headsslenhead_dimr-   r-   r.   	repeat_kvD   s
   0rC           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$   r   r#   )r&   dtype)ptrainingr!   )rC   num_key_value_groupsr(   matmul	transposer   
functionalsoftmaxfloat32torM   rK   rO   
contiguous)rE   rF   rG   rH   rI   rJ   rK   rL   
key_statesvalue_statesattn_weightsattn_outputr-   r-   r.   eager_attention_forwardP   s   
r\   positions_idsbetamax_position_embeddingsc              	   C   s*   d|t dt | |    }|dS )Nr!   r#   )r(   logfloorr1   )r]   r^   r_   rJ   r-   r-   r.   _get_llama_4_attn_scalei   s    
rb   c                       s   e Zd ZdZdedef fddZ		ddejde	ejejf d	ejdB d
e
dB dejdB dee de	ejejdB f fddZ  ZS )Ministral3Attentionz=Multi-headed attention from 'Attention Is All You Need' paperconfig	layer_idxc                    s   t    || _|| _t|dd p|j|j | _|j|j | _	| jd | _
|j| _d| _tj|j|j| j dd| _tj|j|j| j dd| _tj|j|j| j dd| _tj|j| j |jdd| _d S )NrB   g      TFbias)super__init__rd   re   getattrhidden_sizenum_attention_headsrB   r@   rP   rJ   attention_dropout	is_causalr   Linearq_projk_projv_projo_projselfrd   re   	__class__r-   r.   ri   r   s   
 zMinistral3Attention.__init__Nr:   position_embeddingsrI   past_key_valuescache_positionrL   r<   c                 K   sP  |j d d }g |d| jR }| ||dd}	| ||dd}
| ||dd}|\}}t|	|
||\}	}
|	t|| j	j
d| j	j
d|	j }	|d urm|||d}||
|| j|\}
}t| j	jt}|| |	|
||f| jsdn| j| jt| j	dd d	|\}}|jg |dR   }| |}||fS )
Nr#   r!   r$   llama_4_scaling_beta original_max_position_embeddings)r5   r4   rz   rD   sliding_window)rK   rJ   r}   )r'   rB   rp   viewrR   rq   rr   r9   rb   rd   rope_parametersgetrV   rM   updatere   r   get_interface_attn_implementationr\   rO   rm   rJ   rj   r>   rW   rs   )ru   r:   rx   rI   ry   rz   rL   input_shapehidden_shapequery_statesrX   rY   r4   r5   cache_kwargsattention_interfacer[   rZ   r-   r-   r.   forward   sH   		

zMinistral3Attention.forward)NN)__name__
__module____qualname____doc__r"   intri   r(   Tensortupler   
LongTensorr   r   r   __classcell__r-   r-   rv   r.   rc   n   s(    rc   c                       s$   e Zd Z fddZdd Z  ZS )Ministral3MLPc                    sr   t    || _|j| _|j| _tj| j| jdd| _tj| j| jdd| _tj| j| jdd| _	t
|j | _d S NFrf   )rh   ri   rd   rk   intermediate_sizer   ro   	gate_projup_proj	down_projr   
hidden_actact_fnru   rd   rv   r-   r.   ri      s   
zMinistral3MLP.__init__c                 C   s$   |  | | || | }|S N)r   r   r   r   )ru   r*   r   r-   r-   r.   r      s    zMinistral3MLP.forward)r   r   r   ri   r   r   r-   r-   rv   r.   r      s    
r   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 )Ministral3RMSNormư>epsr<   Nc                    s&   t    tt|| _|| _dS )z@
        Ministral3RMSNorm is equivalent to T5LayerNorm
        N)rh   ri   r   	Parameterr(   onesweightvariance_epsilon)ru   rk   r   rv   r-   r.   ri      s   

zMinistral3RMSNorm.__init__r:   c                 C   sJ   |j }|tj}|djddd}|t|| j  }| j|| S )Nr$   r#   T)keepdim)	rM   rV   r(   rU   powmeanrsqrtr   r   )ru   r:   input_dtypevariancer-   r-   r.   r      s
   zMinistral3RMSNorm.forwardc                 C   s   t | jj d| j S )Nz, eps=)r   r   r'   r   )ru   r-   r-   r.   
extra_repr   s   zMinistral3RMSNorm.extra_repr)r   )
r   r   r   floatri   r(   r   r   r   r   r-   r-   rv   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 )Ministral3DecoderLayerrd   re   c                    sR   t    |j| _t||d| _t|| _t|j|jd| _	t|j|jd| _
d S )N)rd   re   r   )rh   ri   rk   rc   	self_attnr   mlpr   rms_norm_epsinput_layernormpost_attention_layernormrt   rv   r-   r.   ri      s   

zMinistral3DecoderLayer.__init__NFr:   rI   position_idsry   	use_cacherz   rx   rL   r<   c              
   K   s^   |}	|  |}| jd|||||||d|\}}
|	| }|}	| |}| |}|	| }|S )N)r:   rI   r   ry   r   rz   rx   r-   )r   r   r   r   )ru   r:   rI   r   ry   r   rz   rx   rL   residual_r-   r-   r.   r      s&   




zMinistral3DecoderLayer.forward)NNNFNN)r   r   r   r"   r   ri   r(   r   r   r   boolr   r   r   r   r   r-   r-   rv   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 )Ministral3PreTrainedModelrd   modelTr   ry   )r:   
attentionsN)r   r   r   r"   __annotations__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   rc   _can_record_outputsr-   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 )Ministral3RotaryEmbeddinginv_freqNrd   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)rh   ri   r_   max_seq_len_cachedoriginal_max_seq_lenrd   r   r   compute_default_rope_parametersr   attention_scalingregister_bufferclone)ru   rd   devicerope_init_fnr   rv   r-   r.   ri     s   


z"Ministral3RotaryEmbedding.__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_thetarB   Ng      ?r   r$   rM   )r   rM   )	r   rj   rk   rl   r(   arangeint64rV   r   )rd   r   r   baser&   attention_factorr   r-   r-   r.   r   (  s   
&z9Ministral3RotaryEmbedding.compute_default_rope_parametersc           
      C   s   | j d d d d f  |jd dd|j}|d d d d d f  }t|jjtr6|jjdkr6|jjnd}t	|dd+ | |  
dd}tj||fdd	}| | j }| | j }	W d    n1 slw   Y  |j|jd
|	j|jd
fS )Nr   r#   r!   mpscpuF)device_typeenabledr$   r%   r   )r   r   r=   r'   rV   r   
isinstancetypestrr   rR   r(   r)   r4   r   r5   rM   )
ru   r*   r   inv_freq_expandedposition_ids_expandedr   freqsembr4   r5   r-   r-   r.   r   F  s   0&z!Ministral3RotaryEmbedding.forwardr   )NNN)r   r   r   r(   r   r   r"   ri   staticmethodr   r   r   r   r   no_gradr   r   r   r-   r-   rv   r.   r     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dB dej	dB dee defddZ  ZS )Ministral3Modelrd   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   ).0re   rd   r-   r.   
<listcomp>_  s    z,Ministral3Model.__init__.<locals>.<listcomp>r   r   F)rh   ri   pad_token_idpadding_idx
vocab_sizer   	Embeddingrk   embed_tokens
ModuleListrangenum_hidden_layerslayersr   r   normr   
rotary_embgradient_checkpointing	post_initr   rv   r   r.   ri   X  s   zMinistral3Model.__init__N	input_idsrI   r   ry   inputs_embedsr   rz   rL   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 rF|	d}| jj
d u rNtnt}
|
| j|||||d}|}| j||d}| jd | jj D ]}||f||||||d|}qm| |}t||r|d	S d d	S )
Nz:You must specify exactly one of input_ids or inputs_embedsr   r   r!   )r   )rd   r   rI   rz   ry   r   )r   )rI   r   ry   r   rz   rx   )last_hidden_statery   )
ValueErrorr   r   rd   get_seq_lengthr(   r   r'   r   r1   r}   r   r   r   r   r   r   r   )ru   r   rI   r   ry   r   r   rz   rL   past_seen_tokensmask_functioncausal_maskr:   rx   decoder_layerr-   r-   r.   r   h  sX   

	

zMinistral3Model.forward)NNNNNNN)r   r   r   r"   ri   r   r    r   r(   r   r   r   FloatTensorr   r   r   r   r   r   r-   r-   rv   r.   r   V  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 )Ministral3ForCausalLMz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   )
rh   ri   r   r   r   r   ro   rk   r  r   r   rv   r-   r.   ri     s
   
zMinistral3ForCausalLM.__init__Nr   r   rI   r   ry   r   labelsr   rz   logits_to_keeprL   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, Ministral3ForCausalLM

        >>> model = Ministral3ForCausalLM.from_pretrained("meta-ministral3/Ministral3-2-7b-hf")
        >>> tokenizer = AutoTokenizer.from_pretrained("meta-ministral3/Ministral3-2-7b-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   rI   r   ry   r   r   rz   N)r  r  r   )lossr  ry   r:   r   r-   )r   r   r   r   slicer  loss_functionrd   r   r   ry   r:   r   )ru   r   rI   r   ry   r   r  r   rz   r	  rL   outputsr:   slice_indicesr  r
  r-   r-   r.   r     s0    zMinistral3ForCausalLM.forward)	NNNNNNNNr   )r   r   r   _tied_weights_keys_tp_plan_pp_planri   r   r   r(   r   r   r   r  r   r   r   r   r   r   r   r-   r-   rv   r.   r    sN    		
r  c                   @      e Zd ZdS ) Ministral3ForTokenClassificationNr   r   r   r-   r-   r-   r.   r        r  c                   @   r  )#Ministral3ForSequenceClassificationNr  r-   r-   r-   r.   r    r  r  c                   @   r  )Ministral3ForQuestionAnsweringNr  r-   r-   r-   r.   r    r  r  )r  r  r   r   r  r  )r!   )rD   )Fcollections.abcr   typingr   r(   r   activationsr   cache_utilsr   r   
generationr	   integrationsr
   r   r   masking_utilsr   r   modeling_flash_attention_utilsr   modeling_layersr   r   r   r   modeling_outputsr   r   modeling_rope_utilsr   r   modeling_utilsr   r   processing_utilsr   utilsr   r   r   utils.genericr   r   utils.output_capturingr    configuration_ministral3r"   r/   r9   r   r   rC   Moduler   r\   rb   rc   r   r   r   r   r   r   r  r  r  r  __all__r-   r-   r-   r.   <module>   sv   
C+APK