fix(omnidreams): support native windows demo

[gtong-nv]

Native Windows OmniDreams Interactive Demo Fixes

Summary

• Materialize --synthetic-scene before HUD/MJPEG scene-picker discovery so the
  Windows browser-stream path can start without pre-staged USDZ scenes.
• Register Torch and CUDA runtime DLL directories before importing the Ludus JIT
  extension on Windows, and keep the returned handles alive for the process
  lifetime.
• Add focused CPU tests for synthetic scene materialization and Windows DLL
  directory registration.

Why

The native Windows interactive-drive --stream-mjpeg --synthetic-scene path hit
two blockers before it could complete a lightweight smoke run:

1. HUD/MJPEG startup discovered available scenes before the synthetic scene was
   materialized, so startup failed when the cache had no staged scenes.
2. The Ludus CUDA extension compiled successfully, but importing the generated
   .pyd failed because Windows could not find dependent Torch/CUDA DLLs.

[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.

[gtong-nv]

/ok to test db23b12

[greptile-apps]

Greptile Summary

This PR fixes two blockers that prevented the native Windows interactive-drive --stream-mjpeg --synthetic-scene path from completing a smoke run: scene discovery now happens after the synthetic USDZ is materialized, and Torch/CUDA DLL directories are registered before the JIT extension is imported.

• demo.py: _materialize_synthetic_scene_for_picker is called before _discover_scene_options in both HUD and MJPEG paths, clearing the synthetic args so the downstream backend builder treats the result as a normal archive.
• _plugin.py: Windows DLL directory handles for Torch's lib/ and any discovered cudart64_*.dll directories are registered and kept alive at module scope; os.system(\"where cl.exe\") is replaced with subprocess.run([\"where.exe\", \"cl.exe\"]) for more reliable process isolation.
• omnidreams_singleview.py: Adds _add_windows_cuda_dll_directories, _cudnn_link_flags, _cuda_link_flags, and import-library-from-DLL generation to support the full cuDNN/CUBLAS/NVRTC link graph on Windows; nvidia-cudnn-cu13==9.23.2.1 is pinned as a Windows-only dependency.

Confidence Score: 4/5

Safe to merge with one issue worth resolving: a CUDA DLL directory is silently skipped because the package lookup uses a name that does not match any known PyPI namespace.

The scene-picker ordering fix and the Ludus DLL registration are both correct and well-tested. The omnidreams_singleview.py changes introduce _python_package_dir("nvidia.cu13"), which does not correspond to any standard NVIDIA PyPI package namespace (nvidia-cuda-runtime-cu13 installs as nvidia.cuda_runtime, not nvidia.cu13). The lookup silently returns None, so the CUDA binary directory is never added to the DLL search path. If the required DLLs are not bundled in PyTorch's lib/, extension import will fail with an opaque DLL-not-found error on Windows.

integrations/omnidreams/omnidreams/native/omnidreams_singleview.py — specifically the _add_windows_cuda_dll_directories function and its _python_package_dir("nvidia.cu13") call.

Important Files Changed

Filename	Overview
integrations/omnidreams/ludus-renderer/ludus_renderer/_ops/_plugin.py	Adds Windows DLL directory registration for Torch/CUDA before JIT extension import; replaces os.system with subprocess.run for cl.exe detection; minor formatting improvements.
integrations/omnidreams/omnidreams/interactive_drive/demo.py	Adds _materialize_synthetic_scene_for_picker to build the synthetic USDZ before HUD/MJPEG scene discovery, fixing Windows browser-stream startup ordering.
integrations/omnidreams/omnidreams/native/omnidreams_singleview.py	Adds Windows cuDNN/CUDA DLL directory registration and import library generation from DLL exports; _python_package_dir("nvidia.cu13") references a non-existent PyPI package namespace, silently skipping CUDA binary directory registration.
integrations/omnidreams/tests/interactive_drive/test_demo_scene_selection.py	Adds test coverage for _materialize_synthetic_scene_for_picker: verifies synthetic args are consumed, args.scene is set, and synthetic_scene is cleared to False.
integrations/omnidreams/tests/test_ludus_renderer_plugin.py	New test verifying Windows DLL directory handles are kept alive and deduplication works correctly across multiple calls.
integrations/omnidreams/pyproject.toml	Pins nvidia-cudnn-cu13==9.23.2.1 as a Windows-only dependency for the cuDNN DLL needed by the native extension.
integrations/omnidreams/omnidreams_singleview/patches/cutlass/sm120-tma-pool.patch	Whitespace-only change: trailing spaces replaced with empty lines in the patch context, no semantic change to the CUTLASS TMA descriptor pool patch.
uv.lock	Lock file updated to pin nvidia-cudnn-cu13==9.23.2.1 for win32 and 9.20.0.48 for Linux, with correct wheel hashes.

Sequence Diagram

%%{init: {'theme': 'neutral'}}%%
sequenceDiagram
    participant CLI as interactive-drive CLI
    participant Demo as demo.py
    participant Synth as build_synthetic_scene_to_temp
    participant Disc as _discover_scene_options
    participant Plugin as _get_plugin()
    participant DLL as os.add_dll_directory

    CLI->>Demo: _run_streaming / _run_slangpy_hud(args)
    Demo->>Demo: _resolve_demo_paths(args)
    Demo->>Synth: _materialize_synthetic_scene_for_picker(args)
    Note over Demo,Synth: NEW: synthetic USDZ built before discovery
    Synth-->>Demo: scene_path (temp USDZ)
    Demo->>Demo: "args.scene = scene_path, args.synthetic_scene = False"
    Demo->>Disc: _discover_scene_options(args.scene_dir, args.scene)
    Disc-->>Demo: scene_options (includes synthetic scene)
    Demo->>Plugin: _get_plugin() [JIT compile/load]
    Plugin->>DLL: _ensure_windows_dll_directories()
    Note over Plugin,DLL: NEW: torch/lib + cudart64 dirs registered
    DLL-->>Plugin: handles kept in _dll_directory_handles[]
    Plugin-->>Demo: loaded .pyd extension

Loading

%%{init: {'theme': 'base', 'themeVariables': {"darkMode": true, "background": "#0d1117", "primaryColor": "#21262d", "primaryTextColor": "#e6edf3", "primaryBorderColor": "#8b949e", "lineColor": "#8b949e", "textColor": "#e6edf3", "edgeLabelBackground": "#161b22", "actorBkg": "#21262d", "actorBorder": "#8b949e", "actorTextColor": "#e6edf3", "actorLineColor": "#8b949e", "signalColor": "#8b949e", "signalTextColor": "#e6edf3", "noteBkgColor": "#373320", "noteBorderColor": "#d4a72c", "noteTextColor": "#f0e6c0", "labelBoxBkgColor": "#21262d", "labelBoxBorderColor": "#8b949e", "labelTextColor": "#e6edf3", "loopTextColor": "#e6edf3", "activationBkgColor": "#30363d", "activationBorderColor": "#8b949e"}}}%%
sequenceDiagram
    participant CLI as interactive-drive CLI
    participant Demo as demo.py
    participant Synth as build_synthetic_scene_to_temp
    participant Disc as _discover_scene_options
    participant Plugin as _get_plugin()
    participant DLL as os.add_dll_directory

    CLI->>Demo: _run_streaming / _run_slangpy_hud(args)
    Demo->>Demo: _resolve_demo_paths(args)
    Demo->>Synth: _materialize_synthetic_scene_for_picker(args)
    Note over Demo,Synth: NEW: synthetic USDZ built before discovery
    Synth-->>Demo: scene_path (temp USDZ)
    Demo->>Demo: "args.scene = scene_path, args.synthetic_scene = False"
    Demo->>Disc: _discover_scene_options(args.scene_dir, args.scene)
    Disc-->>Demo: scene_options (includes synthetic scene)
    Demo->>Plugin: _get_plugin() [JIT compile/load]
    Plugin->>DLL: _ensure_windows_dll_directories()
    Note over Plugin,DLL: NEW: torch/lib + cudart64 dirs registered
    DLL-->>Plugin: handles kept in _dll_directory_handles[]
    Plugin-->>Demo: loaded .pyd extension

Loading

_{Reviews (4): Last reviewed commit: "test(omnidreams): run ludus dll director..." | Re-trigger Greptile}

[gtong-nv]

/ok to test ad2ce6e

[jarcherNV]

LGTM

[gtong-nv]

/ok to test 6160246

[gtong-nv]

/ok to test 67c881e

[greptile-apps]

Want your agent to iterate on Greptile's feedback? Try greploops.