fix(grpo): TRL 1.2 stock path + LoRA double-wrap fix + smoke memory caps

scripts/job_train.sh

@@ -65,13 +65,26 @@ print("SFT adapter:", run_sft(cfg, m, t))
 PY
 
 # GRPO 50 steps (Path 3 — 0.5B sanity check; switch to qwen_1_5b + max_steps=400 for Path 1)
+# Memory overrides for single-A10G + vLLM colocate at 0.5B:
+#  - num_generations=4 (was 8)            → halves rollout VRAM
+#  - max_prompt_length=2048 (was 4096)    → halves KV cache per prompt
+#  - max_completion_length=512 (was 2048) → quarters generation memory
+#  - vllm_gpu_memory_utilization=0.35     → leaves ~16GB headroom for trainer
 python <<'PY'
 from hydra import initialize_config_dir, compose
 from train.model_load import load_model_and_tokenizer
 from train.grpo import run_grpo
 from rewards.compose import make_reward_fn
 with initialize_config_dir(config_dir="/w/configs", version_base="1.3"):
-    cfg = compose(config_name="config", overrides=["model=qwen_0_5b_smoke","train=grpo","train.max_steps=50"])
+    cfg = compose(config_name="config", overrides=[
+        "model=qwen_0_5b_smoke",
+        "train=grpo",
+        "train.max_steps=50",
+        "train.num_generations=4",
+        "train.max_prompt_length=2048",
+        "train.max_completion_length=512",
+        "train.vllm_gpu_memory_utilization=0.35",
+    ])
 m, t = load_model_and_tokenizer(cfg)
 print("GRPO merged:", run_grpo(cfg, m, t, make_reward_fn(cfg.reward), "http://localhost:8001"))
 PY