OpenRA-Bench / VENDOR.md
yxc20098's picture
Vendor the necessary openra_rl_training + openra_env modules
9063f2b
|
Raw
History Blame Contribute Delete
1.19 kB

A newer version of the Gradio SDK is available: 6.19.0

Upgrade

Vendored dependencies

OpenRA-Bench vendors a small, faithful subset of two sibling repos so the benchmark runs without external source checkouts, and so the evaluation stack is frozen for reproducibility.

openra_rl_training/ โ€” from OpenRA-RL-Training

  • scenario.py โ€” ScenarioDefinition, VALID_ACTOR_TYPES
  • training/reward_funcs.py โ€” DEFAULT_REWARD_WEIGHTS (composite scorer)
  • training/rust_env_pool.py โ€” RustEnvPool (wraps the engine)
  • training/minimap_renderer.py โ€” render_minimap (terrain minimap)

openra_env/ โ€” from OpenRA-RL

  • game_data.py โ€” RA_UNITS / RA_BUILDINGS (consumed by scenario.py)

These are verbatim copies โ€” do not hand-edit. To update: re-copy from the source repos and re-run the full suite (pytest tests/).

NOT vendored

  • openra_train โ€” the Rust engine, a compiled extension. Build it from the OpenRA-Rust repo: maturin develop --release. It cannot be vendored as source.
  • openra_env.client / openra_env.models โ€” the legacy gRPC client. The bench never imports them at runtime (the only reference is example code inside a docstring), so they are intentionally left out.