packing and neat_packing

[AriesJin]

使用qwen2.5VL进行sft微调的时候，设置packing，但是发现迭代的次数并没有减少，是哪里设置不对嘛？这是全部的设置

model

“”“
model_name_or_path: Qwen/Qwen2.5-vl
image_max_pixels: 1000000
video_max_pixels: 8192
trust_remote_code: true

enable_liger_kernel: true
use_unsloth_gc: true
flash_attn: auto
packing: true

method

stage: sft
do_train: true
finetuning_type: full
freeze_vision_tower: true
freeze_multi_modal_projector: false
freeze_language_model: true

dataset

dataset: ****
template: qwen2_vl
cutoff_len: 65536
overwrite_cache: true
preprocessing_num_workers: 16
dataloader_num_workers: 4
tokenized_path: tokenized_dataset

output

output_dir: models/output
logging_steps: 10
save_steps: 200
save_total_limit: 5
plot_loss: true
overwrite_output_dir: true
save_only_model: false
report_to: none # choices: [none, wandb, tensorboard, swanlab, mlflow]
use_swanlab: true
swanlab_api_key: ****
swanlab_project: llama_factory
swanlab_run_name: ***

train

per_device_train_batch_size: 1
gradient_accumulation_steps: 2
torch_empty_cache_steps: 100
learning_rate: 1.0e-5
num_train_epochs: 3.0
lr_scheduler_type: cosine
warmup_ratio: 0.1
bf16: true
ddp_timeout: 180000000
gradient_checkpointing: true
resume_from_checkpoint: null

”“”