Update cpu affinity test; add TLLM NUMA AWARE WORKER AFFINITY as a optional env variable can be set as user

Author: SchumiDing

.github/workflows/e2e_ppo_grpo_trainer_trtllm_cpu_affinity_test.yml

@@ -87,6 +87,7 @@ jobs:
           fetch-depth: 0
       - name: Install the current repository
         run: |
+          pip3 install ray==2.41.0
           pip3 install -r requirements-test.txt
           pip3 install --no-deps -e .
       - name: Prepare GSM8K dataset

verl/workers/rollout/trtllm_rollout/trtllm_async_server.py

@@ -328,11 +328,16 @@ async def launch_servers(self):
             if not self.is_reward_model
             else f"trtllm_server_reward_{self.replica_rank}"
         )
-
-        runtime_env_vars = {
-            "RAY_EXPERIMENTAL_NOSET_CUDA_VISIBLE_DEVICES": "1",
-            # "TLLM_NUMA_AWARE_WORKER_AFFINITY": "0"
-        }
+        tllm_numa_aware_worker_affinity = os.getenv("TLLM_NUMA_AWARE_WORKER_AFFINITY")
+        if tllm_numa_aware_worker_affinity == "0":
+            runtime_env_vars = {
+                "RAY_EXPERIMENTAL_NOSET_CUDA_VISIBLE_DEVICES": "1",
+                "TLLM_NUMA_AWARE_WORKER_AFFINITY": "0"
+            }
+        else:
+            runtime_env_vars = {
+                "RAY_EXPERIMENTAL_NOSET_CUDA_VISIBLE_DEVICES": "1",
+            }
         server = TRTLLMHttpServer.options(
             scheduling_strategy=ray.util.scheduling_strategies.NodeAffinitySchedulingStrategy(
                 node_id=node_id,