ZjKSrSSKJr SSKJr SSKrSSKJr SSKJr SSK J r SS K J r SS K Jr SS KJrJr SS KJr SS KJrJrJr SSKJrJrJr SSKJr SSKJr \R@"\!5r"Sr#"SS\RH5r%Sr&S.Sjr'S/S\RHS\RPS\RPS\RPS\RPS-S\)S\)4Sjjr*"SS \RH5r+"S!S"\RH5r,"S#S$\RH5r-"S%S&\ 5r."S'S(\RH5r/\"S)S*\55r0S+r1\"S,S-\055r2S-S*/r3g)0zPyTorch Pixtral model.)Callable)OptionalN)nn)ACT2FN)GradientCheckpointingLayer)BaseModelOutput)dynamic_rope_update)ALL_ATTENTION_FUNCTIONSPreTrainedModel)Unpack)TransformersKwargsauto_docstringlogging)is_flash_attention_requestedmaybe_autocastmerge_with_config_defaults)capture_outputs)PixtralVisionConfigc/nUHnURSSupE[R"[R"U5[R"U5SS9n[R"USS9R SS5R SS5upxXq-U-n URU SS2S45 M [R"U5$)Nij)indexingdimr) shapetorchmeshgridarangestackreshapechunkappendcat) patch_embeds_list max_width positionspatchheightwidthmeshh_gridv_grididss }/root/GenerationalWealth/GenerationalWealth/venv/lib/python3.13/site-packages/transformers/models/pixtral/modeling_pixtral.pyposition_ids_in_meshgridr3%sI" BC( ~~ell62ELL4GRVWTr2::2qAGG2N 6)QT# # 99Y c ^\rSrSr%Sr\R \S'SS\4U4Sjjjr \ SS\S-S\ SS \ S-S \ S \44S jj5r\R "5\S 55rSrU=r$)PixtralRotaryEmbedding0a The key with pixtral embedding is just that you have a frequency for each pixel positions. If you have height x width pixels (or embedding pixels), then the frequency used for ROPE is given by indexing the pre_computed frequency on the width and height. What you output is of dimension (batch, height * width, dim) with dim the embed dim. This simply means that for each image hidden state, you are going to add a corresponding positional embedding, based on its index in the grid. inv_freqNconfigc>[TU]5 XlURRSUlUR nURS:wa/[ URRSURS35eU"URU5upVURSUSS9 URSUR5SS9 g) N rope_typedefaultz7 does not support non-default RoPE, but got `rope_type=`r8F) persistentoriginal_inv_freq) super__init__r9rope_parametersr;compute_default_rope_parameters ValueError __class____name__register_bufferclone)selfr9device layer_type rope_init_fnr8attention_scalingrEs r2rAPixtralRotaryEmbedding.__init__>s  44[A!%!E!E >>Y &>>**++bcgcqcqbrrst '34;;&G# ZeD 0(..2BuUr4rJz torch.deviceseq_lenreturnz torch.TensorcURSn[USS5=(d URUR-nSnURUR -n[ R"U5n[ R"U5nSU[ R"SUS5R5U- -- n [ R"XySSS25R5n [ R"XSSS25R5n [ R"U SS2SSS24RSUS5U SSS2SS24RUSS5/SS 9RSUS-5n [ R"X4SS 9n X4$) aH 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) rBgetattr hidden_sizenum_attention_heads image_size patch_sizer r"floatouterr'repeatr$) r9rJrObaserattention_factormax_patches_per_sidehwfreqsfreqs_hfreqs_wr8s r2rC6PixtralRotaryEmbedding.compute_default_rope_parametersNsf&%%l3fj$/c63E3EIcIc3c &00F4E4EE LL- . LL- .t QQ 7 = = ?# EFG++ass,224++aqt!t-335994 #**1.BAFa #**+?AF   '"cQh  99h1r:))r4cURUn[URR[5(a0URRS:waURROSn[ USS9 UnUR 5nUR5nSSS5 WRURS9WRURS94$!,(df  N@=f)NmpscpuF) device_typeenableddtype) r8 isinstancerJtypestrrcossintork)rIx position_idsrarhembrorps r2forwardPixtralRotaryEmbedding.forward{s l+'1!((--'E'E!((--[`J`ahhmmfk  U CC'')C'')CD vvAGGv$cff177f&;;; D Cs 5#C C )r9r;NN)NNN)rF __module__ __qualname____firstlineno____doc__r Tensor__annotations__rrA staticmethodrinttuplerYrCno_gradr ru__static_attributes__ __classcell__rEs@r2r6r60s llV2VV -1+/"**#d***(**t** ~u$ % ****X ]]_<