`load_checkpoint_and_dispatch` fails for GPTNeoX 20B

johnPertoft commented 1 year ago

System Info

- `Accelerate` version: 0.15.0
- Platform: Linux-4.19.0-22-cloud-amd64-x86_64-with-glibc2.35
- Python version: 3.10.6
- Numpy version: 1.23.5
- PyTorch version (GPU?): 1.13.0+cu117 (True)
- `Accelerate` default config:
        Not found

- 1xNvidia T4, 510.47.03, cuda 11.7
- 30gb ram

Information

[ ] The official example scripts
[X] My own modified scripts

Tasks

[ ] One of the scripts in the examples/ folder of Accelerate or an officially supported no_trainer script in the examples folder of the transformers repo (such as run_no_trainer_glue.py)
[ ] My own task or dataset (give details below)

Reproduction

Run the following script (slightly adapted from https://huggingface.co/docs/accelerate/usage_guides/big_modeling) but with the GPTNeoX model.

from accelerate import init_empty_weights
from accelerate import load_checkpoint_and_dispatch
from huggingface_hub import snapshot_download
from transformers import GPTNeoXTokenizerFast
from transformers import AutoConfig
from transformers import AutoModelForCausalLM

model_path = snapshot_download("EleutherAI/gpt-neox-20b")
tokenizer = GPTNeoXTokenizerFast.from_pretrained(model_path)

prompt = "GPTNeoX20B is a 20B-parameter autoregressive Transformer model developed by EleutherAI."
inputs = tokenizer(prompt, return_tensors="pt")
inputs = inputs.to("cuda")

print("Creating empty model")
config = AutoConfig.from_pretrained(model_path)
tokenizer = GPTNeoXTokenizerFast.from_pretrained(model_path)
with init_empty_weights():
    model = AutoModelForCausalLM.from_config(config)

print("Loading checkpoint")
model = load_checkpoint_and_dispatch(
    model,
    checkpoint=model_path,
    device_map="auto",
    no_split_module_classes=["GPTNeoXLayer"],
    offload_folder="/tmp/gpt-neox-20b-offload-accelerate",
)

for k, v in model.hf_device_map.items():
    print(k, v)

print("Running forward pass")
outputs = model(**inputs)

It crashes in load_checkpoint_and_dispatch with the following trace

Traceback (most recent call last):
  File "/workspaces/run-big-model/.experimental/run-large-model/repro.py", line 29, in <module>
    model = load_checkpoint_and_dispatch(
  File "/workspaces/run-big-model/.venv/lib/python3.10/site-packages/accelerate/big_modeling.py", line 375, in load_checkpoint_and_dispatch
    load_checkpoint_in_model(
  File "/workspaces/run-big-model/.venv/lib/python3.10/site-packages/accelerate/utils/modeling.py", line 697, in load_checkpoint_in_model
    offload_weight(param, param_name, state_dict_folder, index=state_dict_index)
  File "/workspaces/run-big-model/.venv/lib/python3.10/site-packages/accelerate/utils/offload.py", line 37, in offload_weight
    array = weight.numpy()
TypeError: can't convert cuda:0 device type tensor to numpy. Use Tensor.cpu() to copy the tensor to host memory first.

I then tried to update https://github.com/huggingface/accelerate/blob/cf22df91899dd57ead1df6c30983261596da7c99/src/accelerate/utils/modeling.py#L677 to always load the state_dict into cpu (map_location="cpu") which got me past that error.

But I subsequently ran into this crash instead

Traceback (most recent call last):
  File "/workspaces/run-big-model/.experimental/run-large-model/repro.py", line 29, in <module>
    model = load_checkpoint_and_dispatch(
  File "/workspaces/run-big-model/.venv/lib/python3.10/site-packages/accelerate/big_modeling.py", line 385, in load_checkpoint_and_dispatch
    return dispatch_model(
  File "/workspaces/run-big-model/.venv/lib/python3.10/site-packages/accelerate/big_modeling.py", line 290, in dispatch_model
    attach_align_device_hook_on_blocks(
  File "/workspaces/run-big-model/.venv/lib/python3.10/site-packages/accelerate/hooks.py", line 486, in attach_align_device_hook_on_blocks
    attach_align_device_hook_on_blocks(
  File "/workspaces/run-big-model/.venv/lib/python3.10/site-packages/accelerate/hooks.py", line 486, in attach_align_device_hook_on_blocks
    attach_align_device_hook_on_blocks(
  File "/workspaces/run-big-model/.venv/lib/python3.10/site-packages/accelerate/hooks.py", line 486, in attach_align_device_hook_on_blocks
    attach_align_device_hook_on_blocks(
  File "/workspaces/run-big-model/.venv/lib/python3.10/site-packages/accelerate/hooks.py", line 465, in attach_align_device_hook_on_blocks
    attach_align_device_hook(
  File "/workspaces/run-big-model/.venv/lib/python3.10/site-packages/accelerate/hooks.py", line 378, in attach_align_device_hook
    attach_align_device_hook(
  File "/workspaces/run-big-model/.venv/lib/python3.10/site-packages/accelerate/hooks.py", line 369, in attach_align_device_hook
    add_hook_to_module(module, hook, append=True)
  File "/workspaces/run-big-model/.venv/lib/python3.10/site-packages/accelerate/hooks.py", line 146, in add_hook_to_module
    module = hook.init_hook(module)
  File "/workspaces/run-big-model/.venv/lib/python3.10/site-packages/accelerate/hooks.py", line 254, in init_hook
    set_module_tensor_to_device(module, name, self.execution_device)
  File "/workspaces/run-big-model/.venv/lib/python3.10/site-packages/accelerate/utils/modeling.py", line 117, in set_module_tensor_to_device
    raise ValueError(f"{tensor_name} is on the meta device, we need a `value` to put in on {device}.")
ValueError: bias is on the meta device, we need a `value` to put in on 0.

Expected behavior

I expected this to run without crashes

sgugger commented 1 year ago

Thanks for reporting the issue. I remember having the same problem for the second issue (which comes from the buffers in the attention layer of GPT-NeoX) in Transformers but I don't recall right now how I fixed it. I'm on vacation until the beginning of Jan, so will look deeper then!

In the meantime, a workaround is to just load the model with Transformers:

from transformers import AutoModelForCausalLM

model = AutoModelForCausalLM.from_pretrained(
    "EleutherAI/gpt-neox-20b",
    device_map="auto",
    offload_folder="/tmp/gpt-neox-20b-offload-accelerate",
)

You can then run the forward pass as long as you put it under a torch.no_grad() (otherwise it OOMs because of the activations saved for the backward pass).

sgugger commented 1 year ago

Thanks for your patience! I dug into this a bit more and the PR linked above should fix your initial code sample (as long as you add a torch.no_grad() for the forward pass, otherwise you will probably get OOM).

huggingface / accelerate