Completion Length Static (Wrong length logged to WANDB)

[Syazvinski]

Completion length is stuck at 203 for me; it seems to be the length of the prompt?

Training with command:

torchrun \ --nproc_per_node=4 \ --nnodes=1 \ --node_rank=0 \ --master_addr=127.0.0.1 \ --master_port=12345 \ src/open_r1/grpo.py \ --output_dir output_train \ --model_name_or_path Qwen/Qwen2.5-VL-3B-Instruct \ --dataset_name mydataset \ --deepspeed local_scripts/zero3.json \ --max_completion_length 512 \ --per_device_train_batch_size 1 \ --gradient_accumulation_steps 2 \ --logging_steps 1 \ --bf16 \ --report_to wandb \ --gradient_checkpointing false \ --attn_implementation flash_attention_2 \ --num_train_epochs 5 \ --run_name Qwen2.5-VL-3B-GRPO-RUN1 \ --save_steps 100 \ --save_only_model true \ --num_generations 4 \ --max_pixels 518400

Yet the completion length stays the same throughout the whole trianing process. Setting debug to true and looking at the logs, I can see that the completion length DOES vary, it just seems that the logger isnt properly getting the average length of generations.