CLI#

FlashDreams exposes a unified command line entry point: flashdreams-run.

Core commands#

List all available runner slugs:

uv run flashdreams-run --help

Inspect one runner’s full options:

uv run flashdreams-run lingbot-world-fast --help

Run a single-GPU inference (run is the default mode):

uv run flashdreams-run lingbot-world-fast --total-blocks 7

Launch the LingBot v2 Cam2V application:

uv run --no-sync flashdreams-run-v2 cam2v-lingbot \
    --mode webrtc --host 0.0.0.0 --port 8089 -- --example-data

The common command shape is flashdreams-run <runner> [mode]. A runner only advertises modes it implements; unsupported pairs fail before CUDA initialization. Shared modes are run, mp4, null, webrtc, and local-window.

Run a multi-GPU inference:

uv run torchrun --nproc_per_node=4 --no-python flashdreams-run \
    lingbot-world-fast --total-blocks 7

Resolve config only (no model instantiation):

uv run flashdreams-run lingbot-world-fast --no-instantiate

Post-processing presets#

Post-processing presets run on decoded RGB frames from a video runner. Select one with --postprocess.preset:

uv run flashdreams-run lingbot-world-fast \
    --postprocess.preset rtx-super-resolution

The rtx-super-resolution preset wraps NVIDIA VFX Python bindings for RTX Video Super Resolution. Install the optional dependency with uv pip install 'flashdreams[rtx-postprocess]' and run on a supported RTX GPU before selecting this preset.

Native v2 applications receive their own arguments after --. Interactive Drive exposes the equivalent setting with the hyphenated --postprocess-preset option:

uv run flashdreams-run-v2 interactive-drive --mode webrtc -- \
    --postprocess-preset rtx-super-resolution

The preset starts enabled and can be toggled between generated chunks with the Post-processing checkbox in the Interactive Drive HUD. Use flashdreams-run-v2 interactive-drive -- --help to list the presets registered in the current environment.

See also#