This issue organizes the remaining tasks to bring SGLang feature support to parity with vLLM.

Features to support:

• Support use_conversation_multi_turn=true. [Dependencies] Upgrade to torch 2.7  #73 observed issues with SGLang version 0.4.8.post1 returning an empty response when enabled.
• Support SGLang colocated with training with TP==1 (pending in [Generator] Support non-remote (e.g. colocated) SGLang engine #68)
• Support SGLang colocated with training with TP>1
• Need to fix remote server sleep for SGLang, which currently uses vllm API
  

  SkyRL/skyrl-train/skyrl_train/inference_engines/remote_inference_engine.py

  Lines 144 to 147 in c88a678

  	async def sleep(self, *args: Any, **kwargs: Any):
  	async with aiohttp.ClientSession() as session:
  	resp = await session.post(f"{self.url}/sleep", json={"level": kwargs.get("level", 1)})
  	return await resp.json()

• ...