Add Miles Qwen3-8B GRPO training example on H100

[xyuzh]

Summary

• Adds a single-node RL training example for Qwen3-8B with GRPO on 8x H100-80GB using Anyscale
• Uses Miles with Megatron backend (TP=2, DP=2) for training and 3 disaggregated SGLang engines for rollout
• Includes Dockerfile, Anyscale job config, and entrypoint script that handles model download, HF→Megatron weight conversion, and async GRPO training

Files

File	Description
miles_qwen3_8b_h100/Dockerfile.anyscale	Docker image with Miles, Megatron-LM, SGLang, flash-attn, TE
miles_qwen3_8b_h100/job.yaml	Anyscale job config (m5.2xlarge head + 1x p5.48xlarge worker)
miles_qwen3_8b_h100/entrypoint.sh	Downloads model/data, converts weights, runs async GRPO training
miles_qwen3_8b_h100/README.md	Setup instructions, cluster layout, and tuning guide

Cluster Layout

Head node (m5.2xlarge):  driver only, no GPUs
Worker 0 (8x H100-80GB):
  GPU 0-3: Training (TP=2, DP=2)
  GPU 4-7: Rollout (3 SGLang engines + 1 driver)

Test plan

• Build the Docker image via Anyscale
• Submit the job with anyscale job submit -f job.yaml
• Verify weight conversion completes successfully
• Confirm training loss decreases and reward increases over rollouts