Jetson Nano 4 GB bring-up
This target is for the original Jetson Nano 4 GB development module (P3448-0000) on NVIDIA's reference carrier. It uses microSD storage. The 2 GB, production eMMC, and Orin Nano modules are different targets.
Local build
Run from the repository root:
./kas-container build kas/build-configs/jetson-nano-kirkstone.yaml
For a dependency-resolution check without compilation:
./kas-container shell kas/build-configs/jetson-nano-kirkstone.yaml \
-c 'bitbake -n pantavisor-starter'
The configuration pins OE4T's kirkstone-l4t-r32.7.x layer, which provides
L4T R32.7.6, Linux 4.9, and CUDA 10.2. NVIDIA's binary userspace requires
glibc. The CUDA recipes supply a separate GCC 8 compiler; the entire system
does not need to use GCC 8. LXC 3 and cgroup v1 are selected for this kernel.
Do not replace this layer with the regular kirkstone or scarthgap
meta-tegra branches: those target newer NVIDIA BSPs.
Artifacts and boot
Outputs are under build/tmp-kirkstone/deploy/images/jetson-nano-devkit/:
pantavisor-starter-jetson-nano-devkit.sdcard.gz: compressed whole-card image for a microSD card (16 GB or larger).pantavisor-starter-jetson-nano-devkit.tegraflash.tar.gz: NVIDIA recovery flashing tools, QSPI firmware, partition layout, and the APP filesystem.pantavisor-bsp-jetson-nano-devkit.pvrexport.tgz: Pantavisor BSP export.pv-jetson-cuda.pvrexport.tgz: CUDA diagnostic container.
The tegraflash archive supplies doflash.sh for recovery flashing and
dosdcard.sh for recreating a raw .sdcard image. A rootfs .ext4 file is
not a whole-card image. Before running a flashing script, inspect its
arguments and NVIDIA's recovery procedure for the attached board. Initial
QSPI firmware may need to be flashed even when booting from microSD.
The APP filesystem is labeled root and holds both /boot and Pantavisor
storage. Its boot script reads the revision from /boot/uboot.txt, tracks
try-boot state in /pv.env, and boots the revision's kernel and initramfs.
It preserves the device tree patched by NVIDIA's CBoot firmware, including
memory reservations needed by the GPU. Updating that firmware/device tree
through Pantavisor is not implemented by this target.
CUDA acceptance test
The factory revision includes pv-jetson-cuda. Its startup runs
deviceQuery and vectorAdd before signaling READY. The latter launches a
CUDA kernel and validates the result; enumeration alone is insufficient.
The container declares the required nvidia driver, mapped to nvgpu, and
its mdev rules expose nvhost* and nvmap device nodes.
On a running device, enter the container and repeat the tests:
pventer -c pv-jetson-cuda
/usr/bin/cuda-samples/deviceQuery
/usr/bin/cuda-samples/vectorAdd
Both must pass. A successful compilation or a successful CPU test does not establish CUDA support on hardware. This first board target contains CUDA diagnostics; llama.cpp, a chat model, and a chat UI are subsequent app builds.
API build integration
The KAS target is intended to be built and tested locally before being
registered in the cloud build catalog. It produces a native NVIDIA
tegraflash artifact, not a Raspberry Pi image-base archive. API image
assembly must understand that format before it can compose factory images
from arbitrary application builds. Do not use the Raspberry Pi image-base
assembler for this target.