rhk"dZddlZddlmZddlmZmZddlZddlZddlm Z ddl m Z ddl m Z dd lmZdd lmZdd lmZmZmZdd lmZmZd dlmZej6eZerddlmZddl m!Z!m"Z"nd\Z!Z"Zer ddl#m$Z$m%Z%nd\Z%Z$e&ee!e"e$e%fZ'dejPde)fdZ*dZ+dZ,dZ-GddZ.Gddejj^Z0Gdde j^Z1Gd d!e j^Z2Gd"d#eZ3eGd$d%eZ4eed&'Gd(d)eZ5eed*'Gd+d,eZ6eGd-d.e4Z7ed/'Gd0d1e4e Z8gd2Z9y)3zPyTorch MAMBA2 model.N) dataclass)OptionalUnion)nn)ACT2FN)GenerationMixin)GradientCheckpointingLayer)PreTrainedModel) ModelOutputauto_docstringlogging)is_causal_conv1d_availableis_mamba_2_ssm_available) Mamba2Config)selective_state_update)mamba_chunk_scan_combined mamba_split_conv1d_scan_combinedNNN)causal_conv1d_fncausal_conv1d_update)NN input_tensorpad_sizect|jdk(r ddddd|ddfnddd|ddf}tjjj ||ddS)z Padding x tensor with `pad_size` on the seq_len dim (dim=1) Assumes that we only have tensors of either size 4 or 3 rconstant)modevalue)lenshapetorchr functionalpad)rr pad_shapes }/var/www/html/ai-insurance-compliance-backend/venv/lib/python3.12/site-packages/transformers/models/mamba2/modeling_mamba2.pypad_tensor_by_sizer'Bsf 47|7I7I3Ja3OAq!Q!Q/VWYZ\]_gijlmUnI 88   " "<ST " UUct||}t|jdk(r.|j|jdd||jdS|j|jdd||jd|jdS)z Padding input_tensor with `pad_size` on the seq_len dim (dim=1) and simultaneously splitting it into chunk sequences. Assumes that we only have tensors of either size 4 or 3 rr)r'r r!reshape)rr chunk_sizes r&reshape_into_chunksr.Ms&lH=L <  !###L$6$6q$92z