Conflict between Lightning and Huggingface Transformers (device_map).

[richarddwang]

Hugginface will add AlignDeviceHook under certain cases, intercept nn.modules' forward, and send input tensors to some devices that is automatically decided by their logics. It, in many cases, will conflict with devices we set for L.Trainer.

For example,

When

• Use quantization:
  e.g.

AutoModelForCausalLM.from_pretrained(model_name, quantization_config=BitsAndBytesConfig(
                load_in_4bit=True,
                bnb_4bit_quant_type="nf4",
                bnb_4bit_use_double_quant=True,
                bnb_4bit_compute_dtype=torch.bfloat16,
), device_map='auto')

• or usse PeftModel:
  e.g.

from peft import get_peft_config, get_peft_model, TaskType
class lightningmodule
  def __init__
    model = AutoModelForCausalLM.from_pretrained(...)
    sef.model = get_peft_model(model , peft_config)

In these cases huggingface secretly send input tensors to cuda:0 or cuda:1, which conflicts with L.trainer(devices=[2]) that send model weights to cuda:2. All of this results in error of that tensors are at differnet devices.

This conflict make lightning hard to train hugginface transformers in many cases, currently I manually remove huggingface's hook that send tensors to differnet device by

def remove_hook_from_module(module: torch.nn.Module, recurse=False, hook_cls=AlignDevicesHook):

    if hasattr(module, "_hf_hook") and isinstance(module._hf_hook, hook_cls):
        module._hf_hook.detach_hook(module)
        delattr(module, "_hf_hook")

        if hasattr(module, "_old_forward"):
            module.forward = module._old_forward
            delattr(module, "_old_forward")

    if recurse:
        for child in module.children():
            remove_hook_from_module(child, recurse)

    return module

wonder if there is better way for use lightning and huggingface together...

[leeglg]

how can i use this function?

[richarddwang]

from accelerate.hooks import AlignDevicesHook

def remove_hook_from_module(module: torch.nn.Module, recurse=False, hook_cls=AlignDevicesHook):

    if hasattr(module, "_hf_hook") and isinstance(module._hf_hook, hook_cls):
        module._hf_hook.detach_hook(module)
        delattr(module, "_hf_hook")

        if hasattr(module, "_old_forward"):
            module.forward = module._old_forward
            delattr(module, "_old_forward")

    if recurse:
        for child in module.children():
            remove_hook_from_module(child, recurse)

    return module

model = auto_cls.from_pretrained(model_name, **config_kwargs)
remove_hook_from_module(model, recurse=True)

[AndriiZelenko]

Legend

[qvuer7]

Question, how can i make this one works with peft model as:

model = auto_cls.from_pretrained(model_name, **config_kwargs)
model = get_peft_model(model, peft_config)
remove_hook_from_module(model, recurse=True)

as for now getting this error:

AttributeError: 'PeftModelForCausalLM' object has no attribute '_hf_hook'

[GoldenStain]

So if we specify the device with a number instead of a list, this problem can be avoided?