LinkerError: CUDA_ERROR_UNSUPPORTED_PTX_VERSION when training LSTM Transducer BPE with NeMo ASR

[Leuss99]

I am trying to train an LSTM Transducer BPE model using NeMo ASR on an NVIDIA GeForce RTX 3060 Ti GPU. When I start training, I get a CUDA linker error related to the PTX version:
LinkerError: [222] Call to cuLinkAddData results in CUDA_ERROR_UNSUPPORTED_PTX_VERSION
ptxas application ptx input, line 9; fatal: Unsupported .version 8.8; current version is '8.5'

Environment:

NeMo: 2.4.0
PyTorch: 2.3.1+cu121
CUDA (PyTorch built with): 12.1
NVIDIA Driver: 555.42.06 (CUDA Version: 12.5)
GPU: NVIDIA GeForce RTX 3060 Ti, 8GB VRAM

What I tried:

Downgrading PyTorch from 2.8.0+cu128 to 2.3.1+cu121
Checked CUDA/driver compatibility (NVIDIA-SMI shows CUDA 12.5, PyTorch built with CUDA 12.1)
Installed minimal NeMo ASR packages
Confirmed CUDA is available (torch.cuda.is_available() → True

How I can fix it?