qi)dZddlZddlmZddlZddlmZddlmZddl m Z ddl m Z dd l mZdd lmZdd lmZdd lmZmZmZmZdd lmZddlmZej8eZdej>de fdZ!dZ"dZ#dZ$GddZ%Gddej jLZ'GddejLZ(GddejLZ)GddeZ*eGd d!eZ+eed"#Gd$d%eZ,eed&#Gd'd(eZ-eGd)d*e+Z.ed+#Gd,d-e+e Z/gd.Z0y)/zPyTorch MAMBA2 model.N) dataclass)nn)initialization)ACT2FN)GenerationMixin)lazy_load_kernel)GradientCheckpointingLayer)PreTrainedModel) ModelOutputauto_docstringis_torchdynamo_compilinglogging)resolve_internal_import) Mamba2Config 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 \/opt/pipecat/venv/lib/python3.12/site-packages/transformers/models/mamba2/modeling_mamba2.pypad_tensor_by_sizer!'sf 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!rrreshape)rr chunk_sizes r reshape_into_chunksr(2s&lH=L <  !###L$6$6q$92z