up

Author: Superjomn

examples/grpo_trainer/run_qwen2-7b_math_trtllm.sh

@@ -14,6 +14,7 @@ export RAY_DEDUP_LOGS=0
 TP=${1:-4}
 PROJECT_NAME=${PROJECT_NAME:-"verl_grpo_example_gsm8k_math"}
 EXP_NAME=trtllm-qwen2-7b-tp${TP}-8gpus${EXP_NAME_SUFFIX:+"-"}${EXP_NAME_SUFFIX}
+NODES=${NODES:-1}
 
 if [ $TP -eq 4 ]; then
     MAX_BATCH_SIZE=1024
@@ -81,7 +82,7 @@ python3 -m verl.trainer.main_ppo \
     trainer.project_name="${PROJECT_NAME}" \
     trainer.experiment_name=${EXP_NAME} \
     trainer.n_gpus_per_node=8 \
-    trainer.nnodes=1 \
+    trainer.nnodes=${NODES} \
     trainer.save_freq=-1 \
     trainer.test_freq=5 \
     trainer.resume_mode=disable \

verl/workers/rollout/trtllm_rollout/trtllm_async_server.py

@@ -98,6 +98,8 @@ def get_server_address(self):
 
     async def launch_server(self):
         from tensorrt_llm import AsyncLLM
+        import tensorrt_llm.logger as trtllm_logger
+        trtllm_logger.set_level('info')
         from tensorrt_llm.llmapi import CudaGraphConfig, KvCacheConfig
         from tensorrt_llm.serve import OpenAIServer
 
@@ -134,6 +136,7 @@ async def launch_server(self):
             "sampler_type": "TRTLLMSampler",
             **engine_kwargs,
         }
+        print(f"llm_kwargs: {llm_kwargs}")
 
         self.llm = await AsyncLLM(**llm_kwargs)