fix vllm memory leak

[hjh0119]

PR type

• Bug Fix
• New Feature
• Document Updates
• More Models or Datasets Support

PR information

fix vllm 0.7.3 memory leak when n > 1

#3508
vllm-project/vllm#14326

Write the detail information belongs to this PR.

Experiment results

Paste your experiment result here(if needed).

[hjh0119]

Test: GRPO training with Tensor Parallelism (TP). (When TP is not used, request_config.n is equal to 1, so this memory leak does not occur.)

MAX_PIXELS=602112 \
CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 \
NPROC_PER_NODE=8 \
swift rlhf \
    --rlhf_type grpo \
    --model 'Qwen/Qwen2.5-VL-7B-Instruct' \
    --external_plugins examples/train/grpo/plugin/plugin.py \
    --reward_funcs external_r1v_acc format \
    --use_vllm true \
    --train_type lora \
    --torch_dtype bfloat16 \
    --dataset 'lmms-lab/multimodal-open-r1-8k-verified#1000' \
    --max_completion_length 2048 \
    --num_train_epochs 3 \
    --per_device_train_batch_size 1 \
    --per_device_eval_batch_size 1 \
    --learning_rate 1e-6 \
    --eval_steps 100 \
    --save_steps 200 \
    --save_total_limit 2 \
    --logging_steps 5 \
    --warmup_ratio 0.05 \
    --dataloader_num_workers 4 \
    --dataset_num_proc 4 \
    --num_generations 4 \
    --temperature 1.0 \
    --top_p 0.9 \
    --top_k 50 \
    --async_generate false \
    --system 'examples/train/grpo/prompt.txt' \
    --deepspeed zero3 \
    --log_completions true \
    --num_iterations 1 \
    --num_infer_workers 8 \
    --gradient_accumulation_steps 2 \
    --tensor_parallel_size 4 \
    --sleep_level 1 \
    --vllm_gpu_memory_utilization 0.7    

Add a print log in grpo_trainer.py and include the log output at step 10.

        distributed_idx = self.round_robin(len(all_inputs), get_node_setting()[1] * self.args.num_infer_workers)
        if self.infer_rank >= 0:
            engine = self.engine.engine
            if hasattr(engine, 'llm_engine'):
                engine = engine.llm_engine
            print(f"DEBUG: len of self.engine.engine.seq_id_to_seq_group: {len(engine.seq_id_to_seq_group)}")

before fix:

DEBUG: len of self.engine.engine.seq_id_to_seq_group: 40

after fix:

DEBUG: len of self.engine.engine.seq_id_to_seq_group: 1

[hjh0119]

Added a strict requirement for the version to avoid compatibility issues with lower versions.

[tastelikefeet]

nice work