`idpdZddlZddlmZddlmZmZddlZddlmZddl m Z ddl m Z dd l mZdd lmZdd lmZmZmZdd lmZmZd dlmZejeZerddlmZddlm Z m!Z!nd\Z Z!Zer ddl"m#Z#m$Z$nd\Z$Z#e%ee e!e#e$fZ&dej'de(fdZ)dZ*dZ+dZ,GddZ-Gddejj.Z/Gddej.Z0Gd d!ej.Z1Gd"d#eZ2eGd$d%eZ3eed&'Gd(d)eZ4eed*'Gd+d,eZ5eGd-d.e3Z6ed/'Gd0d1e3e Z7gd2Z8dS)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|jdkr ddddd|ddfnddd|ddf}tjj||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 ~/home/jaya/work/projects/VOICE-AGENT/VIET/agent-env/lib/python3.11/site-packages/transformers/models/mamba2/modeling_mamba2.pypad_tensor_by_sizer'Asj 47|7I3J3Ja3O3OAq!Q!Q//VWYZ\]_gijlmUnI 8  " "<ST " U UUc"t||}t|jdkr.||jdd||jdS||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.Ls&lH==L < !####L$6q$92z