Clear divide of loss before and after commit 9bec3c98a22c91b1c28fda757db51eb780291641

[HideLord]

Reminder

• I have read the README and searched the existing issues.

Reproduction

Command:

deepspeed --num_gpus 2 --master_port=9901 src/train_bash.py     --deepspeed ds_config.json     --flash_attn     --stage sft     --do_train True     --model_name_or_path yam-peleg/Experiment26-7B     --finetuning_type lora     --template mistral     --dataset_dir data     --dataset double_take_dataset     --data_seed 42     --seed 42     --cutoff_len 1500     --learning_rate 0.00003     --num_train_epochs 1.0     --max_samples 100000     --per_device_train_batch_size 4     --per_device_eval_batch_size 4     --gradient_accumulation_steps 1     --lr_scheduler_type polynomial     --max_grad_norm 1.0     --logging_steps 5     --bf16 True     --lora_rank 108     --lora_alpha 216     --lora_target all     --val_size 0.05     --evaluation_strategy steps   --save_steps 0.1    --eval_steps 0.1   --load_best_model_at_end True     --plot_loss True     --run_name DoubleTake_v16.9     --output_dir saves/Mistral/lora/DoubleTake_v16.9

Deepspeed config:

{
  "train_batch_size": "auto",
  "train_micro_batch_size_per_gpu": "auto",
  "gradient_accumulation_steps": "auto",
  "gradient_clipping": "auto",
  "zero_allow_untested_optimizer": true,
  "fp16": {
    "enabled": "auto",
    "loss_scale": 0,
    "initial_scale_power": 16,
    "loss_scale_window": 1000,
    "hysteresis": 2,
    "min_loss_scale": 1
  },
  "zero_optimization": {
    "stage": 2,
    "allgather_partitions": true,
    "allgather_bucket_size": 5e8,
    "reduce_scatter": true,
    "reduce_bucket_size": 5e8,
    "overlap_comm": false,
    "contiguous_gradients": true
  }
}

The ones with the higher loss are the commits after (and including) 9bec3c98a22c91b1c28fda757db51eb780291641. The same can be seen with the train loss as well:

All the tests were performed with the same dataset and seed and arguments.

Expected behavior

No regression in the train/eval loss.

System Info

• transformers version: 4.39.1
• Platform: Linux-5.15.0-97-generic-x86_64-with-glibc2.35
• Python version: 3.10.12
• Huggingface_hub version: 0.22.0
• Safetensors version: 0.4.2
• Accelerate version: 0.28.0
• Accelerate config: not found
• PyTorch version (GPU?): 2.2.1+cu121 (True)
• Tensorflow version (GPU?): not installed (NA)
• Flax version (CPU?/GPU?/TPU?): not installed (NA)
• Jax version: not installed
• JaxLib version: not installed
• Using GPU in script?: yes
• Using distributed or parallel set-up in script?: yes

Others

Please tell me if you need more info, and I will help if possible!

[hiyouga]

It should not cause a loss regression, could you try fine-tuning models with different seeds?

[HideLord]

It should not cause a loss regression, could you try fine-tuning models with different seeds?

I started another run, but it will take some time. Will update this comment once a few hundred steps have ran.

[HideLord]

Here are another two runs with a new seed 123. The data shuffling has been turned off so that the runs are comparable:

deepspeed --num_gpus 2 --master_port=9901 src/train_bash.py     --deepspeed ds_config.json     --flash_attn     --stage sft     --do_train True     --model_name_or_path yam-peleg/Experiment26-7B     --finetuning_type lora     --template mistral     --dataset_dir data     --dataset double_take_dataset     --data_seed 123     --seed 123     --cutoff_len 1500     --learning_rate 0.00003     --num_train_epochs 1.0     --max_samples 100000     --per_device_train_batch_size 4     --per_device_eval_batch_size 4     --gradient_accumulation_steps 1     --lr_scheduler_type polynomial     --max_grad_norm 1.0     --logging_steps 5     --bf16 True     --lora_rank 108     --lora_alpha 216     --lora_target all     --val_size 0.05     --evaluation_strategy steps   --save_steps 0.1    --eval_steps 0.1   --load_best_model_at_end True     --plot_loss True     --run_name DoubleTake_v16.9     --output_dir saves/Mistral/lora/DoubleTake_v16.9

Black = With commit 9bec3c98a22c91b1c28fda757db51eb780291641
Red = With commit 7b8f5029018f0481f7da83cc5ee4408d95c9beb2

Eval Loss:

Train Loss:

[HideLord]

Here is the logs difference if you need it:
https://www.diffchecker.com/YdIhGqzq/

[hiyouga]

Could you try again with 3bcd41b ?

[HideLord]

Absolute legend 🤩
That fixed it! Blue is the new run:

[hiyouga]

Thank you very much for your rapid test, we are still working to check the reason leading to the loss regression. We will post a bug report once we find it out.