Add rl_perf_repro functional test to CI and make script consistent with verl e2e

Signed-off-by: Shuyi Xiong <219646547+shuyixiong@users.noreply.github.com>

Author: shuyixiong

tests/integration/defs/ray_orchestrator/RL/run_rl_perf_reproduce.py

@@ -41,7 +41,7 @@ def __init__(self, async_llm_kwargs: dict, sampling_kwargs: dict):
 
     async def init_llm(self):
         """Initialize the AsyncLLM instance with configured parameters."""
-        self.llm = AsyncLLM(
+        self.llm = await AsyncLLM(
             model=self.async_llm_kwargs["model"],
             backend="pytorch",
             orchestrator_type=self.async_llm_kwargs["orchestrator_type"],
@@ -61,7 +61,6 @@ async def init_llm(self):
             batch_wait_timeout_iters=32,
             batch_wait_max_tokens_ratio=0.5,
         )
-        await self.llm.setup_async()
         self.sampling_params = SamplingParams(
             temperature=self.sampling_kwargs["temperature"],
             top_p=self.sampling_kwargs["top_p"],

tests/integration/test_lists/test-db/l0_dgx_h100.yml

@@ -301,3 +301,4 @@ l0_dgx_h100:
   tests:
     - unittest/_torch/ray_orchestrator/multi_gpu -m "gpu4"
     - unittest/llmapi/test_async_llm.py -m "gpu4"
+    - ray_orchestrator/RL/test_rl_perf_reproduce.py