GRPO多机LMDeploy训练warning，loss为0

[owl-10]

报错信息：
[2025-02-24 15:47:13] [TM][WARNING] [forward] Request failed for 1740383233880208906, code 6
[2025-02-24 15:47:14] [TM][WARNING] [RejectInvalidRequests] Skipping invalid infer request for id 1740383234397829963, code = 6
[2025-02-24 15:47:14] [TM][WARNING] [forward] Request failed for 1740383234397829963, code 6
脚本：
#!/bin/bash

export NNODES=4
export MASTER_PORT=${MASTER_PORT:-34229}
export GPUS=${GPUS:-8}
export NPROC_PER_NODE_NODE0=6
export NPROC_PER_NODE_OTHERS=6

export NCCL_IB_DISABLE=0
export NCCL_P2P_DISABLE=0
export NCCL_SHM_DISABLE=0
export TF_CPP_MIN_LOG_LEVEL=3
export LAUNCHER=pytorch

--- Determine NODE_RANK (Priority: MLP_ROLE_INDEX > NODE_RANK) ---

if [ -n "$MLP_ROLE_INDEX" ]; then
export NODE_RANK=$MLP_ROLE_INDEX
echo "Using MLP_ROLE_INDEX for NODE_RANK: $NODE_RANK"
elif [ -n "$NODE_RANK" ]; then
echo "Using NODE_RANK: $NODE_RANK"
else
echo "Error: NODE_RANK or MLP_ROLE_INDEX environment variable must be set (e.g., export NODE_RANK=0)."
exit 1
fi

--- Configure NPROC_PER_NODE based on NODE_RANK ---

if [ "$NODE_RANK" -eq 0 ]; then
export NPROC_PER_NODE=$NPROC_PER_NODE_NODE0
echo "NPROC_PER_NODE set to: $NPROC_PER_NODE (for Node 0)"
else
export NPROC_PER_NODE=$NPROC_PER_NODE_OTHERS
echo "NPROC_PER_NODE set to: $NPROC_PER_NODE (for Node Rank > 0)"
fi

export CUDA_VISIBLE_DEVICES="0,1,2,3,4,5,6,7"

echo "NNODES: $NNODES"
echo "NODE_RANK: $NODE_RANK"
echo "MASTER_ADDR: $MASTER_ADDR"
echo "MASTER_PORT: $MASTER_PORT"
echo "NPROC_PER_NODE: $NPROC_PER_NODE"
echo "CUDA_VISIBLE_DEVICES: $CUDA_VISIBLE_DEVICES"
echo "----------------------------------------"

--- Distributed Training Command using torchrun ---

torchrun
--nnodes=$NNODES
--node_rank=$NODE_RANK
--master_addr=$MLP_WORKER_0_HOST
--nproc_per_node=$NPROC_PER_NODE
--master_port=$MLP_WORKER_0_PORT
swift/cli/rlhf.py
--rlhf_type grpo
--reward_funcs traj_acc_think format
--model_type internvl2_5
--use_lmdeploy true
--lmdeploy_session_len 2048
--lmdeploy_cache_max_entry_count 0.8
--train_type full
--torch_dtype bfloat16
--max_completion_length 512
--num_train_epochs 3
--per_device_train_batch_size 2
--per_device_eval_batch_size 2
--learning_rate 1e-6
--gradient_accumulation_steps 16
--eval_steps 25
--save_steps 25
--save_total_limit 10
--logging_steps 1
--max_length 4096
--output_dir output_multi_node_think_zero
--warmup_ratio 0.05
--dataloader_num_workers 32
--dataset_num_proc 32
--num_generations 12
--log_completions True
--temperature 0.9
--top_p 0.9
--top_k 50
--async_generate true
--deepspeed zero1
--num_iterations 2
--num_infer_workers 2
--attn_impl flash_attn
--system 'examples/train/grpo/prompt_think.txt'