OOM

感谢您开源了如此棒的工作，我在运行sft脚本的时候，在[`deepcopy`](https://github.com/ByteDance-Seed/Bagel/blob/c8d5e35f7d87d043b748caa87c45a946caf3265a/train/pretrain_unified_navit.py#L482C7-L482C32) 这行直接爆内存了，我有800G的内存，请问是什么原因造成的呢？

以下是我的运行脚本：
```
torchrun \
--nnodes=1 \
--node_rank=0 \
--nproc_per_node=8 \
train/pretrain_unified_navit.py \
--dataset_config_file ./data/configs/example.yaml \
--model_path /root/model/bagel \
--layer_module Qwen2MoTDecoderLayer \
--max_latent_size 64 \
--finetune_from_hf True \
--auto_resume False \
--resume-model-only False \
--finetune-from-ema True \
--log_every 1 \
--lr 2e-5 \
--num_worker 1 \
--expected_num_tokens 4096 \
--max_num_tokens 5120 \
--max_num_tokens_per_sample 4096 \
--cpu_offload True
```
感谢！

Read train/pretrain_unified_navit.py at the linked deepcopy line, then reproduce the failure with the supplied torchrun command and flags. Trace the memory use across the eight processes and document the cause and a verified way to complete the run without OOM.

Written by the indexing model from the issue text.