[Packaging] Improve packaging logic with various miscellaneous changes

[ArielG-NV]

Resolves: #716

Changes:

• Taxi demo

  • Performance: Reach 30 FPS on an RTX PRO 6000 workstation by reusing main/BEV render buffers and reducing repeated work in ground snapping, traffic updates, and physics spatial indexing.
  • Fullscreen: Keep presentation resolution and UI settings synchronized with the fullscreen window size.
  • Offline assets: Move live-edit checkpoints and the FP8 VAE state into the FlashDreams cache so they are included in offline bundles.

• Offline packaging

  • Deployment: Package the executable, runtime libraries, and prepared caches as a complete folder that runs without Python, uv, or network access.
  • Working directory: Add --working-dir to include application assets and resolve relative paths consistently.
  • Cache location: Add --unpack-dir to control the writable runtime cache location, with Windows long-path support.
  • Dependencies: Include required ImGui fonts and Ninja; exclude Jedi to reduce packaged file counts.

• Fixes and diagnostics

  • Native extensions: Load prebuilt extensions directly in packaged applications and report missing binaries clearly instead of attempting target-side compilation.
  • Checkpoint reuse: Load cached Hugging Face checkpoints directly without download calls or unnecessary free-space checks.
  • Logging: Add Windows --noconsole support with persistent runtime logs
  • Documentation: Update offline-packager docs

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

[greptile-apps]

Confidence Score: 4/5

[Medium impact] The changes reach offline app packaging and the Crazy Robotaxi simulation workflow.

Fix the custom native build-directory path before merging so completed performance bundles can start reliably.

Findings

1.  Custom builds miss the bundle ▶

Summary

This PR changes offline bundles to ship as a complete raw folder. It adds working-directory assets, a configurable writable cache, console-free logging, and direct loading of prepared native extensions.

• Offline apps run from a complete folder with configurable working and cache paths.
• The taxi simulation reuses ground, traffic, physics, and rendering work between frames.
• Packaged OmniDreams apps load prepared native extensions directly.
• Robotaxi live-edit files and OmniDreams VAE state use the FlashDreams cache.
• Cached Hugging Face checkpoints load directly from their local files.
• Fullscreen size changes now update the saved taxi presentation settings.
• Several old repository, benchmark, attention, and packaging tests are removed.

Diagram

%%{init: {'theme': 'neutral'}}%%
flowchart LR
  A[Application command and working files] --> B[Preload and prepare caches]
  B --> C[Build executable and runtime folder]
  C --> D[Ship complete bundle]
  D --> E[Copy seed into writable user cache]
  D --> F[Load prepared native binaries]
  E --> G[Run application]
  F --> G

Loading

_{Reviews (6) · Last reviewed commit: "fix formatting" · Reviewed by Greptile}

[ArielG-NV]

/ok to test 823038c

[greptile-apps]

Comments Outside Diff

These findings could not be posted inline.

•  Packaging checks disappear tools/package-as-offline-exe/test_package_as_offline_exe.py:8 ▶

  Deleting test_package_as_offline_exe.py removes automated checks for supported modes, bundled GPU libraries, preload settings, and runtime/application argument separation. check_launcher_streams.py only checks missing output streams and does not replace those tests.

  This is non-blocking, but later packaging changes can break commands or omit required files without these checks catching them. Keep the existing tests and update them for the new bundle layout.

  Note: If this suggestion doesn't match your team's coding style, reply to this and let me know. I'll remember it for next time!

[aidanfnv]

LGTM

[ArielG-NV]

/ok to test f35a9ad

[ArielG-NV]

/ok to test 1cc0010

[ArielG-NV]

/ok to test d068528

[greptile-apps]

Custom builds miss the bundle

If the packaging host has OMNIDREAMS_SINGLEVIEW_NATIVE_BUILD_ROOT set, preload builds the extension outside the bundle. _cache_environment() keeps that variable, and _default_build_root() gives it priority over TORCH_EXTENSIONS_DIR.

The new frozen loader searches only TORCH_EXTENSIONS_DIR/torch_extensions. Packaging can therefore finish successfully, but the deployed performance app cannot start because its required binary was not shipped. Force preload to build into the bundle cache, or copy the selected binary into the directory this loader searches.

Knowledge Base Used: Interactive world-model integrations