T^(h~ dZddlZddlmZddlmZmZmZmZm Z ddl Z ddl Z ddl m Z ddl mZddlmZdd lmZdd lmZmZmZmZdd lmZdd lmZmZmZmZmZm Z d dl!m"Z"m#Z#m$Z$ejJe&Z'e"Z(dHde jRde jTdee+fdZ, dIde jZde jTde j\de+fdZ/dIdZ0dZ1dZ2dZ3dZ4eGddeZ5eGddeZ6Gd d!e jnZ8Gd"d#e jnZ9Gd$d%e jnZ:Gd&d'e jnZ;Gd(d)e jnZ<Gd*d+e jnZ=Gd,d-e jnZ>Gd.d/e jnZ?Gd0d1e jnZ@Gd2d3e jnZAGd4d5e jnZBGd6d7eZCGd8d9eCZDGd:d;eCZEedeCeZFGd?d@e jnZGedAe1GdBdCeCZHedDe1GdEdFeCeZIgdGZJy)JzPyTorch KOSMOS-2 model.N) dataclass)AnyListOptionalTupleUnion)nn)CrossEntropyLoss)ACT2FN)GenerationMixin)BaseModelOutput)BaseModelOutputWithPastAndCrossAttentionsBaseModelOutputWithPooling!CausalLMOutputWithCrossAttentions)PreTrainedModel) ModelOutputadd_start_docstrings%add_start_docstrings_to_model_forwardloggingreplace_return_docstrings torch_int) Kosmos2ConfigKosmos2TextConfigKosmos2VisionConfigmaskdtypetgt_lenc2|j\}}||n|}|ddddddfj|d||j|}d|z }|j|jtj t j |jS)z_ Expands attention_mask from `[bsz, seq_len]` to `[bsz, 1, tgt_seq_len, src_seq_len]`. Nr?)sizeexpandto masked_filltorchboolfinfomin)rrrbszsrc_len expanded_mask inverted_masks j/var/www/skyplay_api_hub/venv/lib/python3.12/site-packages/transformers/models/kosmos2/modeling_kosmos2.py _expand_maskr/3s99;LC ,g'GD$)*11#q'7KNNuUM-'M  $ $]%5%5ejj%A5;;uCUCYCY ZZinput_ids_shapedevicepast_key_values_lengthc|\}}tj||ftj|j|}tj|j d|}|j ||dzj|j ddkd|j|}|dkDr0tjtj|||||gd}|ddddddfj|d|||zS)zB Make causal mask used for bi-directional self-attention. )r2rrrr2dimN) r&fullr(r)aranger" masked_fill_viewr$catzerosr#)r1rr2r3r*rr mask_conds r._make_causal_maskr@As #LC ::w(%++e*<*@*@ PD TYYr]6:Ii9q="6"6tyy}a"HH!L 775>D!yy%++g/EU[abdhioqr dAq ! ( (a'DZ:Z [[r0c|j|j}tj|dj ||z|z}|j |zS)a Replace non-padding symbols with their position numbers. Position numbers begin at padding_idx+1. Padding symbols are ignored. This is modified from fairseq's `utils.make_positions`. Args: x: torch.Tensor x: Returns: torch.Tensor rr7)neintr&cumsumtype_aslong) input_ids padding_idxr3rincremental_indicess r."create_position_ids_from_input_idsrJSsW << $ ( ( *D <<!4<