Run

python3 scripts/run.py config/<model>/<arm>.toml

That is the whole interface. <model> is a directory under config/, <arm> is one configuration of it — usually baseline plus one or more optimized arms.

python3 scripts/run.py config/wan22_ti2v_5b/baseline.toml
python3 scripts/run.py config/wan22_ti2v_5b/wan5b_kernel_easycache_pisa.toml

Each run writes a bundle under runs/<timestamp>-<id>/: launch.sh, manifest.resolved.toml, metadata.json, and outputs/ with the video and benchmark.json.

Before you run

python3 scripts/run.py <config> --print

Resolves the config and checks its paths without running anything — the runtime directory, the entry script, the interpreter — and fails naming whichever is missing.

Overriding a value

python3 scripts/run.py <config> --set KEY=VALUE [--set KEY=VALUE ...]

Changes a key for one run without editing the file. For example, a short smoke run instead of the full 50 steps:

python3 scripts/run.py config/wan22_ti2v_5b/baseline.toml \
    --set WAN22_STEPS=4 --set WAN22_WARMUP_PASSES=0

The config format

id            = "<model>_<arm>"
kind          = "baseline"           # or "optimized"
model_profile = "<model>"            # inherits models/<model>.toml

[runtime]
root = "models/<model>/<variant>"    # the code this config launches

[env]
KEY = "value"                        # exported as an environment variable

The model profile carries the workload and whatever every arm of that model shares; the config carries what makes this arm different. --print shows the merged result.

id            = "wan22_ti2v_5b_baseline"
kind          = "baseline"
model_profile = "wan22_ti2v_5b"

[runtime]
root = "models/wan22_ti2v_5b/baseline"

[env]
WAN22_STEPS = "50"