DreamZero SGLang Evaluation#

This guide runs DreamZero LIBERO evaluation through the RLinf SGLang embodied backend. Use this path for inference-only evaluation with an SGLang-native DreamZero checkpoint layout.

Compared with the original DreamZero eval path in DreamZero Supervised Fine-Tuning and Franka Real-World Deployment, this backend runs the large DreamZero network in a separately launched sglang serve process instead of loading it in the rollout worker. The worker still needs the DreamZero transform stack installed, though: rlinf.models.embodiment.dreamzero.sglang_adapter imports rlinf.data.datasets.dreamzero.data_transforms, which pulls in the groot package installed by install_dreamzero_deps. The eval driver launches the SGLang server group and pushes each server URL to the rollout workers; each worker is a thin client that posts batched observations to the URL assigned to its rank over the VLA action API, then denormalizes the returned action chunks before stepping LIBERO.

Install the Test Environment#

Set up RLinf with the embodied, LIBERO, and DreamZero SGLang dependencies:

cd /path/to/RLinf
bash requirements/install.sh embodied --env libero --model dreamzero \
  --venv /path/to/dreamzero_test

DreamZero support is still under review in SGLang, so use SGLang code that includes sgl-project/sglang#30679 and install it with the diffusion extra:

source /path/to/dreamzero_test/bin/activate
cd /path/to/sglang_dreamzero
pip install -e "python[diffusion]"

Prepare the Checkpoint#

Download the LIBERO SFT Diffusers checkpoint from RLinf/RLinf-DreamZero-WAN2.2-5B-LIBERO-SFT-Diffusers. Point rollout.model.model_path at the downloaded checkpoint directory.

hf download RLinf/RLinf-DreamZero-WAN2.2-5B-LIBERO-SFT-Diffusers \
  --local-dir /path/to/RLinf-DreamZero-WAN2.2-5B-LIBERO-SFT-Diffusers

The checkpoint should contain experiment_cfg/metadata.json. If metadata is not available in the checkpoint, generate it from the LIBERO dataset and set rollout.model.metadata_json_path explicitly:

python toolkits/lerobot/generate_dreamzero_metadata.py \
  --preset libero_sim \
  --dataset-root /path/to/libero \
  --output-metadata /path/to/metadata.json

Run LIBERO-Spatial#

The default SGLang eval config is evaluations/libero/libero_spatial_dreamzero_eval_sglang.yaml.

cd /path/to/RLinf
bash evaluations/run_eval.sh libero libero_spatial_dreamzero_eval_sglang \
  rollout.model.model_path=/path/to/RLinf-DreamZero-WAN2.2-5B-LIBERO-SFT-Diffusers

For a custom metadata file, add:

rollout.model.metadata_json_path=/path/to/metadata.json

Important Config Fields#

Field	Purpose
`rollout.rollout_backend: sglang`	Selects the SGLang rollout backend.
`rollout.sglang.serving_mode: embodied`	Selects `SGLangEmbodiedWorker`; the SGLang serve group is launched by the driver script (`launch_sglang_router_and_server`), and the worker connects to the driver-assigned server URL for inference.
`rollout.sglang.server.pipeline_config.default_num_inference_steps`	Controls the DreamZero denoising steps used by the server.
`rollout.sglang.server.pipeline_config.cfg_scale`	Classifier-free guidance scale for action inference.
`rollout.sglang.server.cfg_parallel_degree`	Splits positive and negative CFG branches across ranks when set to `2`.
`rollout.sglang.server.tp_size`	Tensor-parallel size for the DreamZero DiT.
`rollout.sglang.server.sp_degree`	Sequence-parallel size for the DreamZero DiT attention sequence.
`rollout.model.model_path`	DreamZero Diffusers checkpoint directory loaded by SGLang.
`rollout.model.metadata_json_path`	Normalization statistics used before and after action inference.
`rollout.model.num_action_chunks`	Number of actions returned per model request; `env.eval.max_steps_per_rollout_epoch` must be divisible by this value.

Parallel Overrides#

The supported DreamZero SGLang evaluation entry is libero_spatial_dreamzero_eval_sglang. For local experiments, adjust parallelism by overriding fields on this config instead of switching to a different YAML:

bash evaluations/run_eval.sh libero libero_spatial_dreamzero_eval_sglang \
  rollout.sglang.server.num_gpus=2 \
  rollout.sglang.server.cfg_parallel_degree=2 \
  rollout.model.model_path=/path/to/RLinf-DreamZero-WAN2.2-5B-LIBERO-SFT-Diffusers

Validation#

Evaluation logs are written under logs/<timestamp>-<config>/. Check eval_embodiment.log for the SGLang server command, endpoint readiness, per-episode results, and the final eval/success_once metric.

The LIBERO-Spatial SGLang config uses auto_reset: True and ordered reset states to cover the full suite with fewer parallel environments. See Evaluation Configuration for the LIBERO trajectory accounting rules.

Troubleshooting#

If SGLang cannot find model components, confirm that rollout.model.model_path points to the downloaded Diffusers checkpoint directory.
If metadata loading fails, set rollout.model.metadata_json_path to an existing metadata.json generated for libero_sim.
If local HTTP requests unexpectedly use a proxy, set NO_PROXY=127.0.0.1,localhost before launching evaluation.