Enable the SLIME GRPO test case to run end-to-end on B300 / H200

3.test_cases/pytorch/slime/README.md

@@ -522,6 +522,8 @@ slime/                                    # 3.test_cases/pytorch/slime
 ├── env_vars.colocated.example           # Base config: colocated train+rollout, built-in reward
 ├── env_vars.disaggregated.example       # Overlay: reward model on a CPU pool + heavier GRPO
 ├── slime.Dockerfile                     # SLIME + SGLang + Megatron + EFA image
+├── patches/
+│   └── apply_slime_patches.py           # Self-neutralizing in-place fixes for the pinned upstream SLIME checkout (no-op once upstream merges them)
 ├── requirements.txt                     # Pinned Python RL dependencies
 ├── reward_service.Dockerfile            # CPU-only image for the remote reward service
 ├── reward_service/
@@ -532,8 +534,10 @@ slime/                                    # 3.test_cases/pytorch/slime
 │   ├── reward-service.yaml              # CPU reward service Deployment + Service
 │   └── data-prep-pod.yaml               # Utility pod for data preparation
 ├── recipe/
-│   ├── run_grpo_qwen3_4b.sh             # GRPO training launcher (Qwen3-4B, colocated)
-│   └── run_grpo_qwen3_30b_a3b.sh        # GRPO training launcher (Qwen3-30B-A3B MoE, disaggregated)
+│   ├── run_grpo_qwen3_4b.sh             # GRPO submit script (Qwen3-4B, colocated)
+│   ├── run_grpo_qwen3_30b_a3b.sh        # GRPO submit script (Qwen3-30B-A3B MoE, disaggregated)
+│   └── launcher/
+│       └── grpo_launch.sh               # Ray job entrypoint: sources the model script, expands MODEL_ARGS, execs train.py
 └── scripts/
     ├── convert_checkpoint.sh            # HF <-> Megatron conversion helper
     └── evaluate.sh                      # Evaluation launcher