What

Adds the multi-turn SWE-bench agentic async-GRPO recipe for Qwen3-30B-A3B-Thinking (MoE, 30B total / ~3B active):

• examples/swe_bench/grpo_qwen3_30b_async_swe.yaml — the recipe config
• examples/swe_bench/run_grpo_repro_baseline_swe2.sh — vLLM baseline launcher (reproduces the ~8%-resolved reference run)
• examples/swe_bench/run_grpo_swe2_scale_gen.sh — generation-scaling sweep launcher with a BACKEND=vllm|sglang switch
• examples/swe_bench/REPRO_swe2.md — vLLM baseline reproduction guide
• examples/swe_bench/REPRO_swe2_sglang.md — SGLang reproduction guide

Why

Provides a reproducible reference for multi-turn SWE-bench RL (baseline ~8% resolved from step 1) and a working SGLang generation path at parity with vLLM — rollout completeness, throughput, and training-grade per-token logprob parity.

Status — draft, depends on #2447

The BACKEND=sglang path needs the enhanced SGLang backend (Megatron→SGLang MoE/PP weight-refit, router, fault-tolerance) from #2447, which is not yet merged. On current main's basic SGLang backend the SGLang path will not run the 30B-MoE recipe; the vLLM path is self-contained. Kept as a draft until #2447 lands. The companion gym-proxy token-splicing contiguity fix (required for multi-turn SGLang) is in NVIDIA-NeMo/Gym#1787.

Validation

• Port parity: SGLang multi-turn rollouts 8/8, contiguity failures 0, ~193 gen tok/s with full CUDA graph (≈ vLLM).
• Logprob parity (teacher-forced, 27,493 tokens): median |Δ| 1.38e-3; cross-engine median ≈ the within-engine bf16/MoE noise floor (1.24e-3) — i.e. vLLM differs from SGLang no more than SGLang differs from itself. Details in REPRO_swe2_sglang.md.

Reproduced end-to-end from a clean clone.