Interactive serving

Run a demo in a browser or a local window with flashdreams-run-v2. Use a registered name from --help or the model pages.

Browser

uv run --no-sync flashdreams-run-v2 APPLICATION_SLUG \
    --mode webrtc --host 127.0.0.1 --port 0

Open the URL printed by the server. Port 0 picks an available port. To accept connections from another machine, use --host 0.0.0.0 and a fixed port, then connect to the server's address.

Local window

uv run --no-sync flashdreams-run-v2 APPLICATION_SLUG --mode native-window

Demo arguments

Runtime options go before --; demo options go after it:

uv run --no-sync flashdreams-run-v2 cam2v-lingbot \
    --mode webrtc --host 127.0.0.1 --port 8089 -- \
    --example-data --total-blocks 21

List runtime options with flashdreams-run-v2 --help. For this demo's options, use flashdreams-run-v2 cam2v-lingbot -- --help.

Input and display

The runtime sends user input to the app's loops and displays their output. The app handles controls and model inputs. A custom UI loop can draw controls over the generated frames; otherwise, the runtime displays the frames directly.

Model and UI loops run on separate threads. Use invoke_async to send changes between them. Create a demo shows an ImGui text field sending a prompt to the model loop.

Multiple GPUs

For a model adapter that supports distributed execution, launch all ranks:

uv run torchrun --nproc_per_node=2 --no-python flashdreams-run-v2 \
    cam2v-lingbot --mode webrtc -- --example-data --total-blocks 21

Rank zero handles the window and output. The session's parallel context synchronizes steps and inputs; the model integration handles parallel inference and gathers the output for rank zero.

See the CLI reference for runtime options, or Integrate model adapter for a demo to connect another model.