Qwen3-Omni GRPO Rollout Generates Infinite Repetitive Characters

During GRPO fine-tuning of Qwen3-Omni-30B on H200 using Megatron-LM, the model gets stuck in an infinite loop during rollout generation, repeatedly outputting the same character and never stopping.

i try to add vllm_mm_processor_cache_gb 0, but still not work ......

megatron rlhf
--rlhf_type grpo
--model /path/to/Qwen3-Omni-30B-A3B-Instruct
--dataset /path/to/dataset.jsonl#8000
--train_type full
--tensor_model_parallel_size 4
--expert_model_parallel_size 4
--pipeline_model_parallel_size 1
--context_parallel_size 1
--global_batch_size 1024
--micro_batch_size 1
--num_generations 8
--steps_per_generation 1
--max_length 16384
--max_completion_length 2048
--max_pixels 1003520
--lr 1e-6
--min_lr 1e-7
--lr_warmup_fraction 0.05
--bf16 true
--beta 0.00
--epsilon 3e-4
--epsilon_high 4e-4
--loss_type grpo
--importance_sampling_level sequence
--dynamic_sample false
--overlong_filter true
--use_vllm true
--vllm_mode colocate
--vllm_tensor_parallel_size 8
--vllm_gpu_memory_utilization 0.4
--vllm_max_model_len 16384
--vllm_mm_processor_cache_gb 0
--reward_funcs random_reward
--reward_weights 1.0
--external_plugins /path/to/plugin.py
--freeze_vit true
--freeze_aligner true
--freeze_parameters talker code2wav
--attention_backend flash
--sequence_parallel true
--padding_free true
--offload_model true
--use_precision_aware_optimizer
--recompute_granularity selective
--sleep_level 2
--moe_permute_fusion true
--moe_grouped_gemm true
--moe_shared_expert_overlap true
--moe_aux_loss_coeff 1e-3
--vit_gradient_checkpointing false
--save_interval 50
--eval_interval 50
--log_interval 1
--log_completions true
--no_save_optim true
--no_save_rng true
--num_workers 8
--dataset_num_proc 8
--wandb_project megatron_swift
--finetune
--seed 42

Name: vllm
Version: 0.12.1.dev0+g4fd9d6a85.d20251217.cu130
Name: torch
Version: 2.9.1+cu130
Name: ms_swift
Version: 3.12.0.dev0

Start by reproducing the provided GRPO rollout command with the listed Qwen3-Omni, vLLM, PyTorch, and ms-swift versions, then inspect the rollout-generation logs and logged completions. Trace where repeated characters fail to terminate; done means rollout generation stops normally instead of producing an infinite repetitive sequence.