The Size of ema.safetensors.

Nobody has claimed this yet.

The size of the ema.safetensors saved during the training phase is nearly twice as large as the original official ema file.
official ema.safetensors: 27.21GB
my training ema.safetensors: 52.79GB
Causing out-of-memory (OOM) during reasoning.

Here is my scripts of training:
``` bash
torchrun \
--nnodes=1 \
--node_rank=0 \
--nproc_per_node=8 \
--master_addr=127.0.0.1 \
--master_port=12345 \
train/pretrain_unified_navit.py \
--dataset_config_file ./data/configs/t2i.yaml \
--layer_module Qwen2MoTDecoderLayer \
--max_latent_size 64 \
--model_path $model_path \
--use_flex True \
--resume-from $model_path \
--finetune_from_hf True \
--auto_resume True \
--resume-model-only True \
--finetune-from-ema True \
--wandb_offline True \
--results_dir $output_path \
--checkpoint_dir $ckpt_path \
--log_every 10 \
--save_every 1000 \
--lr 2e-5 \
--total_steps 10000 \
--visual_und False\
--expected_num_tokens 10240 \
--max_num_tokens 11520 \
--max_num_tokens_per_sample 10240 \
--num_workers
```

Start in train/pretrain_unified_navit.py and trace how the training command creates or saves ema.safetensors, especially with --finetune-from-ema and checkpoint options. Compare the saved checkpoint contents and tensor sizes with the official 27.21GB file, then verify that the resulting checkpoint does not cause the reported out-of-memory failure during reasoning.

Written by the indexing model from the issue text.