README.md

RLix Examples

This directory contains examples for using RLix (rlix) for multi-pipeline GPU scheduling.

Available Examples

Multi-Pipeline Test

start_multi_pipeline_test.py - Demonstrates running multiple RL training pipelines concurrently under the RLix control plane.

Prerequisites

• Python 3.10+
• Ray installed and configured
• ROLL framework (for pipeline integrations)

Running the Examples

Multi-Pipeline Example

# Run a single pipeline
python examples/start_multi_pipeline_test.py --config_name full_finetune_pipeline1

# Run two full-finetune pipelines concurrently
python examples/start_multi_pipeline_test.py --config_name full_finetune_pipeline1,full_finetune_pipeline2

# Run two multi-LoRA pipelines concurrently
python examples/start_multi_pipeline_test.py --config_name multi_lora_pipeline1,multi_lora_pipeline2

Configuration

Example configurations are located in:

• examples/config/ - DeepSpeed and environment configs
• examples/rlix_test/ - RLix pipeline configs (full finetune and multi-LoRA)

Integration with ROLL

These examples require the ROLL fork, included as a git submodule:

# Initialize and install the ROLL submodule
git submodule update --init external/ROLL
pip install external/ROLL

Troubleshooting

• Ray connection errors: Ensure Ray is running (ray start --head)
• Import errors: Verify all dependencies are installed
• GPU allocation issues: Check Ray's GPU resources with ray status