Jetson (JetPack 6 / L4T R36)

The Orin rows in benchmarks.md were measured with this setup. JetPack’s Python environment has three traps that all present as a broken install. This page is the path around them.

Torch

The generic PyPI wheel does not run on Jetson. The community index only carries builds for its own library stack. Use NVIDIA’s JetPack wheel:

python3 -m venv .venv
.venv/bin/pip install "loudkit[audio]"
.venv/bin/pip install \
  https://developer.download.nvidia.com/compute/redist/jp/v61/pytorch/torch-2.5.0a0+872d972e41.nv24.08.17622132-cp310-cp310-linux_aarch64.whl
.venv/bin/pip install "numpy<2"   # the NVIDIA wheel is compiled against 1.x

Libraries the wheel needs but JetPack does not ship

libcusparseLt is not in L4T. Fetch NVIDIA’s aarch64 archive once:

mkdir -p ~/libs && cd ~/libs
curl -sL -o cslt.tar.xz https://developer.download.nvidia.com/compute/cusparselt/redist/libcusparse_lt/linux-sbsa/libcusparse_lt-linux-sbsa-0.6.3.2-archive.tar.xz
tar xf cslt.tar.xz && cp libcusparse_lt-*/lib/libcusparseLt.so* .

Every run needs CUPTI and the CUDA libraries on the loader path:

export LD_LIBRARY_PATH=$HOME/libs:/usr/local/cuda-12.6/lib64:/usr/local/cuda-12.6/extras/CUPTI/lib64:$LD_LIBRARY_PATH

What to expect (Orin Nano Super, 25 W)

--cuda-graphs is the difference between below and above real time: 0.73x to 1.83x streaming, and 2.06x for synthesize_long. First audio floors at ~2.3 s. ChunkConfig.first_chunk_max_tokens = 96 takes it from 3.1 s to 2.55 s. The remainder is genuine vocoder compute at this power budget, not overhead: a captured-graph vocoder was tried and measured neutral here. Pinning clocks (sudo jetson_clocks) is the remaining free lever.