[BUG] Zero3 causes AttributeError: 'NoneType' object has no attribute 'numel' in continual training

[thkimYonsei]

I was training LLaVA model using deepspeed zero3. What I want to do is continually train the model to different datasets. I create LLaVA model and in the for-loop, I create new dataset and new trainer, then calls trainer.train(). At the first iteration of the for-loop, the training works properly. However, at the second iteration, I got 1 warning and 1 error: warning: "Invalidate trace cache @ step XX: expected module XX, but got module XX" error: AttributeError: 'NoneType' object has no attribute 'numel' at the same location as this issue

Screenshot:

But when I simply change the deepspeed config to use zero2 instead of zero3, no error occurs. I want to use zero3 for larger batch size training. Can you help me out with this?

System info

• OS: Ubuntu 20.04
• GPU count and types: single node 4 RTX A6000 gpus
• Python version: 3.10.0

zero3.json that I used:

{
    "fp16": {
        "enabled": "auto",
        "loss_scale": 0,
        "loss_scale_window": 1000,
        "initial_scale_power": 16,
        "hysteresis": 2,
        "min_loss_scale": 1
    },
    "bf16": {
        "enabled": "auto"
    },
    "train_micro_batch_size_per_gpu": "auto",
    "train_batch_size": "auto",
    "gradient_accumulation_steps": "auto",
    "zero_optimization": {
        "stage": 3,
        "overlap_comm": true,
        "contiguous_gradients": true,
        "sub_group_size": 1e9,
        "reduce_bucket_size": "auto",
        "stage3_prefetch_bucket_size": "auto",
        "stage3_param_persistence_threshold": "auto",
        "stage3_max_live_parameters": 1e9,
        "stage3_max_reuse_distance": 1e9,
        "stage3_gather_16bit_weights_on_model_save": true
    }
}

zero2.json

{
    "fp16": {
        "enabled": "auto",
        "loss_scale": 0,
        "loss_scale_window": 1000,
        "initial_scale_power": 16,
        "hysteresis": 2,
        "min_loss_scale": 1
    },
    "bf16": {
        "enabled": "auto"
    },
    "train_micro_batch_size_per_gpu": "auto",
    "train_batch_size": "auto",
    "gradient_accumulation_steps": "auto",
    "zero_optimization": {
        "stage": 2,
        "overlap_comm": true,
        "contiguous_gradients": true,
        "sub_group_size": 1e9,
        "reduce_bucket_size": "auto"
    }
}

[Xirid]

I have the same issue and solved it by just reloading the whole model on every iteration. Though now I am getting oom due to the memory not freeing after the second epoch, but I guess that is a different issue.

[tjruwase]

@Xirid, please try the following API to free engine memory https://deepspeed.readthedocs.io/en/latest/zero3.html#gpu-memory-management

[apToll]

我也遇到了在 for 循环中创建新的数据集和新的训练器，然后调用trainer.train()。 在 for 循环的第一次迭代中，训练工作正常。然而，在第二次迭代中，报错为错误： AttributeError: 'NoneType' object has no attribute 'numel' 尝试释放GPU，但没有用，释放GPU如下：# Free GPU memory consumed by model parameters ds_engine.empty_partition_cache()

[tjruwase]

Closing this issue due to lack of response. Please reopen if needed.