rlinf-functional-testing

name	rlinf-functional-testing
description	RLinf: Validate RLinf training examples end-to-end from container tests to single-step training

RLinf Functional Testing

Purpose

RLinf-specific test configurations for the validation harness (L2-L6). These complement the generic testing framework in the shared functional-testing skill with concrete RLinf values.

L2: Container Validation

The container test (infrastructure/manifests/container-test.yaml) runs 10 tests verifying:

• Multi-venv availability (openvla, openvla-oft, openpi)
• EFA libraries (libfabric, aws-ofi-nccl)
• CUDA/GPU detection
• RLinf framework imports (import verl, from flash_attn import flash_attn_func)
• MuJoCo/ManiSkill rendering (MUJOCO_GL=osmesa)

L4: Model Download

The model download jobs (examples/<example>/manifests/model-download.yaml) download models to FSx:

Model	FSx Path	Size
RLinf/Openvla-oft-SFT-libero10-trajall	/fsx/models/openvla-oft-sft-libero10	~14 GB
RLinf/RLinf-Pi0-SFT-Spatial-Object-Goal	/fsx/models/rlinf-pi0-sft-spatial	~14 GB

L5: Per-Example Training Step Tests

Each example runs 1 PPO training step via HYDRA_OVERRIDES="runner.max_steps=1 runner.save_interval=-1 runner.val_check_interval=-1".

Example Configurations

Defined in tests/lib/cluster.sh L5_EXAMPLE_LIST:

Example Name	VENV_NAME	CONFIG_NAME	MODEL_PATH
maniskill-openvlaoft	openvla-oft	maniskill_ppo_openvlaoft_quickstart	/fsx/models/openvla-oft-sft-libero10
libero-pi0	openpi	libero_spatial_ppo_openpi_quickstart	/fsx/models/rlinf-pi0-sft-spatial

Running a Specific Example

./validate.sh --mode cluster --level 5 --skip-to 5 --example maniskill-openvlaoft

Expected Success Indicators

In the pod logs, look for:

• TRAINING_STEP_TEST_PASSED -- explicit success marker
• Global Step: 1/1 -- training step completed
• No Traceback, ERROR:, or FAILED in output

Common Failures

Symptom	Cause	Fix
OOM during FSDP init	GPU too small for 7B model	Use p-class instances (not g6)
ModuleNotFoundError: verl	Wrong venv activated	Check VENV_NAME matches example
input() hang (no output)	LIBERO import without LIBERO_DATASET_DIR	Set env var and pre-create config
FileNotFoundError: config.json	Model not downloaded	Run L4 first

L6: Multi-Node Training

Uses infrastructure/manifests/training-statefulset.yaml + training-service.yaml. Requires 2+ GPU nodes. Ray head starts on pod-0, workers connect via headless service training-svc.

The test patches HYDRA_OVERRIDES onto the StatefulSet for single-step validation.

Related Skills

• functional-testing -- generic testing framework and validation levels
• rlinf-pytorch-training-scripts -- Hydra configs and training entrypoints
• rlinf-dataset-preparation -- model download details

Validation Checklist

• L2 container test passes all 10 tests
• L4 all 3 models downloaded and verified on FSx
• L5 each example completes 1 training step without OOM or error
• L6 multi-node Ray cluster forms and completes 1 training step (if 2+ GPU nodes)