CLI Reference¶
RL-Kernel exposes rlk-repro for the supported Qwen3-8B VIME consistency
workflow:
rlk-repro prepare --workspace /data/rlk-repro --download-data --convert-checkpoint
rlk-repro doctor --workspace /data/rlk-repro
rlk-repro plan --workspace /data/rlk-repro --mode consistency
rlk-repro run --workspace /data/rlk-repro --mode consistency --rollouts 8 --wait
rlk-repro validate --run-dir /data/rlk-repro/data/runs/convergence/<run-id> --seal
rlk-repro report --workspace /data/rlk-repro
Use COMMAND --help for path overrides and execution options. native runs
VIME's operators and consistency runs RL-Kernel's operators; neither mode
reuses rollout log-probabilities. See the
Qwen3-8B consistency guide for required
hardware, runtime setup, Ray configuration, and profile customization.
CUDA plan and run accept --tp-size, --cp-size,
--rollout-tp-size, and --rollout-cp-size. ROCm uses the separate
rl_engine.integrations.orchestrators.vime.experiments.rocm_attention.run_qwen3_8b module documented
in the guide because it owns ROCm-specific Ray, AITER/CK, RCCL, and HIP Graph
setup.
Developer commands¶
The repository also contains lower-level developer commands: