Add Slime multi-node RL training example

[xyuzh]

Summary

• Adds a complete example for multi-node RL training using Slime (Megatron-LM + SGLang disaggregated rollout)
• Runs GRPO training of Qwen3-1.7B on 2 workers x 4x A10G (8 GPUs total, TP=2 x PP=2 across nodes)
• Includes Dockerfile with A10G (SM86) compatibility patches for sgl_kernel and Triton, PyTorch 2.7.1 pinning for pre-built wheel compatibility, and runtime patches for all nodes

Files

File	Description
job.yaml	Anyscale job config (m5.2xlarge head + 2x g5.12xlarge workers)
Dockerfile.anyscale	Docker image with Slime, Megatron-LM, SGLang, flash-attn, TE
anyscale-smoke-2node-a10g.sh	Anyscale entrypoint (downloads model/data, converts weights, trains)
patch_all_nodes.py	Runtime patches for sgl_kernel on A10G via Ray
run-qwen3-4B-smoke-2node-a10g.sh	Bare-metal variant for Qwen3-4B
README.md	Documentation with cluster layout, quick start, and OOM tips

Test plan

• Submit job via anyscale job submit -f job.yaml on a fresh Anyscale workspace
• Verify Docker image builds successfully (~15 min)
• Verify weight conversion completes on GPU worker
• Verify multi-node NCCL init succeeds for PP across workers
• Verify training loss decreases over rollout-train cycles

🤖 Generated with Claude Code