qiUdZddlmZddlmZddlZddlmZddlmZddl m Z dd l m Z dd l mZdd lmZdd lmZmZdd lmZddlmZmZmZddlmZmZddlmZej>e Z!dZ"GddejFZ$dZ%d.dZ& d/dejFdejNdejNdejNdejNdzde(de(fdZ)Gdd ejFZ*Gd!d"ejFZ+Gd#d$ejFZ,Gd%d&e Z-Gd'd(ejFZ.eGd)d*eZ/d+Z0eGd,d-e/Z1d-d*gZ2y)0zPyTorch Pixtral model.)Callable)OptionalN)nn)ACT2FN)FlashAttentionKwargs)GradientCheckpointingLayer)BaseModelOutput)dynamic_rope_update)ALL_ATTENTION_FUNCTIONSPreTrainedModel)Unpack)auto_docstringcan_return_tuplelogging)is_flash_attention_requestedmaybe_autocast)PixtralVisionConfigcg}|D]}|jdd\}}tjtj|tj|d}tj|dj ddj dd\}}||z|z} |j| dddftj|S)Nij)indexingdimr) shapetorchmeshgridarangestackreshapechunkappendcat) patch_embeds_list max_width positionspatchheightwidthmeshh_gridv_grididss ^/opt/pipecat/venv/lib/python3.12/site-packages/transformers/models/pixtral/modeling_pixtral.pyposition_ids_in_meshgridr2%sI"$ BC( ~~ell62ELL4GRVWTr2::2qAGG2Ny 6)QT# $ 99Y c eZdZUdZej ed<d deffd Ze ddedzde dde dzd e d e ffd Zejed ZxZS)PixtralRotaryEmbeddinga The key with pixtral embedding is just that you have a frequency for each pixel positions. If you have height x width pixels (or embedding pixels), then the frequency used for ROPE is given by indexing the pre_computed frequency on the width and height. What you output is of dimension (batch, height * width, dim) with dim the embed dim. This simply means that for each image hidden state, you are going to add a corresponding positional embedding, based on its index in the grid. inv_freqNconfigct|||_|jjd|_|j }|jdk7r/t |jjd|jd||j|\}}|jd|d|jd|jdy) N rope_typedefaultz7 does not support non-default RoPE, but got `rope_type=`r6F) persistentoriginal_inv_freq) super__init__r7rope_parametersr9compute_default_rope_parameters ValueError __class____name__register_bufferclone)selfr7device layer_type rope_init_fnr6attention_scalingrCs r1r?zPixtralRotaryEmbedding.__init__>s  44[A!%!E!E >>Y &>>**++bcgcqcqbrrst '34;;&G## ZeD 0(..2BuUr3rHz torch.deviceseq_lenreturnz torch.Tensorc|jd}t|ddxs|j|jz}d}|j|j z}t j|}t j|}d|t jd|dj|z zz } t j|| dddj} t j|| dddj} t j| dddddfjd|d| dddddfj|ddgd jd|dz} t j| | fd } | |fS) 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_thetahead_dimNg?rrrrr) r@getattr hidden_sizenum_attention_heads image_size patch_sizerr!floatouterr&repeatr#) r7rHrLbaserattention_factormax_patches_per_sidehwfreqsfreqs_hfreqs_wr6s r1rAz6PixtralRotaryEmbedding.compute_default_rope_parametersNso&%%l3fj$/c63E3EIcIc3c &00F4E4EE LL- . LL- .t QQ 7 = = ?# EFG++ass,224++aqt!t-335994 #**1.BAFa #**+?AF   '"cQh  99h1r:)))r3c|j|}t|jjtr/|jjdk7r|jjnd}t |d5|}|j }|j}dddj|jj|jfS#1swYAxYw)NmpscpuF) device_typeenableddtype) r6 isinstancerHtypestrrcossintorg)rGx position_idsr^rdembrkrls r1forwardzPixtralRotaryEmbedding.forward{s l+'1!((--'E!((--[`J`ahhmmfk  U C C'')C'')C  vvAGGv$cff177f&;;;   s 2#CCNN)NNN)rD __module__ __qualname____doc__rTensor__annotations__rr? staticmethodrinttuplerVrAno_gradr rq __classcell__rCs@r1r5r50s llV2V -1+/"**#d***(**t** ~u$ % ****XU]]_<