README
spark-projects
Personal playbooks for ML/robotics experiments on an NVIDIA DGX Spark (GB10, Grace-Blackwell, aarch64). Each subdirectory is a self-contained playbook: code, Dockerfiles, scripts, and notes. Large artifacts (checkpoints, datasets, ONNX/engines, venvs) are gitignored — these repos hold the recipe, not the data.
Environment across all playbooks: NVIDIA GB10, Ubuntu 24.04, aarch64, CUDA 13, driver 580.
Playbooks
pi05-spark-inference/ — π0.5 VLA inference on GB10
openpi π0.5 (Physical Intelligence) Vision-Language-Action inference, BF16 baseline → TensorRT FP8+NVFP4.
| backend | latency | rate | fidelity |
|---|---|---|---|
| PyTorch BF16 | ~203 ms | ~4.9 Hz | — |
| TensorRT FP8+NVFP4 | ~94.7 ms | ~10.6 Hz | cosine 0.997 vs PyTorch |
2.12× speedup, essentially lossless — matches the Jetson Thor reference (~94 ms).
Full pipeline: JAX→PyTorch convert → ONNX (ModelOpt) → trtexec engine → benchmark.
See pi05-spark-inference/notes/findings.md and RUNBOOK.md.
vla-onnx/ — LeRobot VLA → split ONNX → Orin Nano
One pipeline, three models: SmolVLA 450 M, X-VLA 0.9 B, EVO1 775 M. Fine-tune on GB10, cut into split ONNX graphs the 8 GB Orin can actually build engines for. Verify against PyTorch baseline.
Two environments on purpose — lerobot 0.5.1/torch 2.12 for the SmolVLA↔X-VLA comparison,
0.6.1/torch 2.11 for EVO1 and the X-VLA export. They are mutually exclusive pins.
See vla-onnx/README.md.
scene-reconstruction/ — video → Gaussian splat → Isaac Sim
(basically deprecated since Spirula Studio exists. I only use the usd_export from here!)
Smartphone video -> COLMAP -> 3DGRUT raw Gaussian splat -> SuperSplat cleanup/compression -> Isaac Sim NuRec USDZ on DGX Spark. See scene-reconstruction/README.md.
Where the deploy side lives
This repo is the Spark side: fine-tune, export, and validate the export. It stops at a parity-checked ONNX bundle.
| repo | owns |
|---|---|
| spark-projects (here) | fine-tuning, ONNX export, export parity, bundles |
| jetson-orin-nano-vla | does a base model fit on the 8 GB Orin, and what it costs — board prep, TRT build probes, the benchmark |
| kaivuriprokkis | the robot. Fine-tuned runtime, camera, control and generally a very messy dev repo |
TensorRT engines are built on-device and never copied; the ONNX bundle is the portable
artifact. Bundles ship with ship_bundle.sh <dir> orin <name> and land in ~/bundles/.
Notes
pi05-spark-inference/phase2/openpi_on_thor/contains scripts adapted from NVIDIA / Jetson AI Lab's "OpenPi π0.5 on Jetson Thor" tutorial, kept here for reproducibility.- Each playbook documents the exact aarch64/Blackwell dependency gotchas it hit — that's the main value, since stock install instructions rarely work cleanly on GB10.
Contributors
53 commits
Similar repos
Languages
Python
92.0%
Shell
7.4%
README
spark-projects
Personal playbooks for ML/robotics experiments on an NVIDIA DGX Spark (GB10, Grace-Blackwell, aarch64). Each subdirectory is a self-contained playbook: code, Dockerfiles, scripts, and notes. Large artifacts (checkpoints, datasets, ONNX/engines, venvs) are gitignored — these repos hold the recipe, not the data.
Environment across all playbooks: NVIDIA GB10, Ubuntu 24.04, aarch64, CUDA 13, driver 580.
Playbooks
pi05-spark-inference/ — π0.5 VLA inference on GB10
openpi π0.5 (Physical Intelligence) Vision-Language-Action inference, BF16 baseline → TensorRT FP8+NVFP4.
| backend | latency | rate | fidelity |
|---|---|---|---|
| PyTorch BF16 | ~203 ms | ~4.9 Hz | — |
| TensorRT FP8+NVFP4 | ~94.7 ms | ~10.6 Hz | cosine 0.997 vs PyTorch |
2.12× speedup, essentially lossless — matches the Jetson Thor reference (~94 ms).
Full pipeline: JAX→PyTorch convert → ONNX (ModelOpt) → trtexec engine → benchmark.
See pi05-spark-inference/notes/findings.md and RUNBOOK.md.
vla-onnx/ — LeRobot VLA → split ONNX → Orin Nano
One pipeline, three models: SmolVLA 450 M, X-VLA 0.9 B, EVO1 775 M. Fine-tune on GB10, cut into split ONNX graphs the 8 GB Orin can actually build engines for. Verify against PyTorch baseline.
Two environments on purpose — lerobot 0.5.1/torch 2.12 for the SmolVLA↔X-VLA comparison,
0.6.1/torch 2.11 for EVO1 and the X-VLA export. They are mutually exclusive pins.
See vla-onnx/README.md.
scene-reconstruction/ — video → Gaussian splat → Isaac Sim
(basically deprecated since Spirula Studio exists. I only use the usd_export from here!)
Smartphone video -> COLMAP -> 3DGRUT raw Gaussian splat -> SuperSplat cleanup/compression -> Isaac Sim NuRec USDZ on DGX Spark. See scene-reconstruction/README.md.
Where the deploy side lives
This repo is the Spark side: fine-tune, export, and validate the export. It stops at a parity-checked ONNX bundle.
| repo | owns |
|---|---|
| spark-projects (here) | fine-tuning, ONNX export, export parity, bundles |
| jetson-orin-nano-vla | does a base model fit on the 8 GB Orin, and what it costs — board prep, TRT build probes, the benchmark |
| kaivuriprokkis | the robot. Fine-tuned runtime, camera, control and generally a very messy dev repo |
TensorRT engines are built on-device and never copied; the ONNX bundle is the portable
artifact. Bundles ship with ship_bundle.sh <dir> orin <name> and land in ~/bundles/.
Notes
pi05-spark-inference/phase2/openpi_on_thor/contains scripts adapted from NVIDIA / Jetson AI Lab's "OpenPi π0.5 on Jetson Thor" tutorial, kept here for reproducibility.- Each playbook documents the exact aarch64/Blackwell dependency gotchas it hit — that's the main value, since stock install instructions rarely work cleanly on GB10.
Similar repos
Contributors
53 commits
Languages
Python
92.0%
Shell
7.4%