Investigate JetPack 7.2 / CUDA 13 on AGX Orin 64GB (potential sm_87 wheel unlock)

[Vortiago]

Summary

JetPack 7.2 (announced 2026-06-01 at GTC Taipei, detailed ~06-12) ships CUDA 13 to the entire Jetson Orin family — Ubuntu 24.04, kernel 6.8. Orin was previously stuck on CUDA 12 throughout the JetPack 6.x line. This is potentially significant for this pipeline, which is pinned to JetPack 6.2.2 / L4T r36.5 and makes several deliberate concessions because of sm_87 (CUDA 12) wheel scarcity:

• HQQ instead of bitsandbytes for 4-bit (no sm_87 bitsandbytes wheel upstream)
• eager attention (flash-attn / SDPA unreliable on sm_87)

A CUDA 13 toolchain could change wheel availability and unwind part of docs/versions.md's rationale.

This is INVESTIGATE, not UPGRADE

Open questions to answer before any migration:

• Is AGX Orin 64GB explicitly supported by JetPack 7.2? (Launch messaging emphasized the 32GB module + "Super Mode" — 64GB support unconfirmed.)
• Does CUDA 13 on sm_87 actually improve wheel availability — specifically bitsandbytes and flash-attn? If not, the HQQ + eager-attention choices stand regardless.
• Compatibility of the current pin matrix (transformers, trl, peft, HQQ, torch) against a CUDA 13 / Ubuntu 24.04 base.
• Migration cost vs. benefit: is there a concrete training/inference win, or is this churn?

If the answers don't show a real win for the 64GB module, document the finding in docs/versions.md and close — keeping the current stack is a valid outcome.

Complexity: moderate (investigation/spike).

References

• JetPack 7.2 overview: https://www.seeedstudio.com/blog/2026/06/12/nvidia-jetpack-7-2-release-agentic-ai-comes-to-jetson-agx-orin/

_{Filed from a weekly dependency scan (week of 2026-06-22).}