How to Run SolarWM Locally

01

1. Start with the official SolarWM repository

Clone the official repository with git clone https://github.com/Junchao-cs/SolarWM.git, enter the directory and use the backend documentation for the model family you intend to run.

Wan2.2, LTX-2.5 and MiniMax-H3 use separate runtime environments. Do not assume one dependency set is valid across all backbones.

02

2. Create the correct environment and probe it

After activating the documented backbone-specific environment, install the shared SolarWM source in editable mode with python -m pip install -e ..

Then run solarwm environment probe. The probe is important because a Python installation can succeed even when CUDA, backend packages or runtime dependencies are not usable together.

03

3. Choose a checkpoint that is actually released

At publication time, the Wan2.2 TI2V-5B family has the complete public progression: base, Stage 0.5, Stage 1 and Stage 2. The official model repository lists SolarWM-5B-sgf-stage2-81f as the released few-step causal SGF checkpoint.

The larger Wan2.2-14B, LTX-2.5 and MiniMax-H3 routes currently have bidirectional Stage 0.5 support while later causal stages remain marked as coming soon. Recheck the repository before assuming that status is unchanged.

04

4. Resolve configuration before allocating the model

SolarWM uses explicit configuration files and command-line overrides. Resolve the final configuration first so model paths, checkpoint paths, data roots and output locations can be checked before expensive model allocation.

Use the exact example configuration for the backbone and training stage you downloaded. Placeholder paths such as /path/to/... are examples, not official installation locations.

05

5. Run a short Stage 2 inference test first

For the released 5B Stage 2 route, the official documentation uses the SolarWM inference module under torchrun with the matching Stage 2 config, base-model directory, SGF checkpoint, data roots and output directory.

Before exploring long horizons, confirm that the environment loads, the checkpoint matches the config, camera-control input is accepted, frames are produced and output is written correctly. Increase sequence length only after that baseline succeeds.

06

6. Keep the data layout consistent

SolarWM data indexes store shard keys relative to a release root. For local storage, the documented pattern points both index_root and the local transport root at the SolarWM-Data release directory.

Inference users may not need the entire raw training corpus. The project separates annotations, pre-encoded artifacts and larger raw payloads, so obtain only the artifacts required by the recipe you are running.

07

7. Do not guess a universal GPU requirement

SolarWM does not publish one universal minimum VRAM figure that covers every model and mode. Memory needs vary with backbone, parameter count, precision, resolution, sequence length and whether you are training or running inference.

The 5B route is the most complete official local path, but that does not make any specific community VRAM estimate an official minimum. Treat community quantizations, repacks and offloading recipes as separate artifacts unless SolarWM explicitly publishes them.

08

8. Common setup failures

• Environment probe fails: confirm the correct backbone environment is active before installing the shared package.
• Checkpoint download fails: verify that Hugging Face access conditions were accepted and that the requested stage is released.
• Config resolves but inference fails: verify base model, checkpoint, data root and config belong to the same backbone/stage.
• A larger causal checkpoint is missing: Stage 1 and Stage 2 for several larger backbones are still marked as coming soon at publication time.