Revert default behavior of get_state_dict_from_offload

[kylesayrs]

What does this PR do?

Purpose

• Fix bug introduced by slight behavior change of the get_state_dict_from_offload introduced starting in accelerate 1.1.0
• This bug affects models which are executed on the GPU with partial cpu/disk offloading only

Background

Starting in accelerate 1.1.0, get_state_dict_from_offload moves state_dict tensors to the CPU by default.

In the case that a module has parameters on the GPU, get_state_dict_from_offload implements the following procedure:

• The module starts with a reference to the tensors on the GPU (1)
• When get_state_dict_from_offload is called, copies of the tensors are created on the CPU (2) which are returned later
• After the tensors are copied to the CPU, the reference to the GPU tensors (1) is dropped
• After the CPU tensors (2) are moved into the state_dict, the CPU tensors (2) are copied back to the GPU (3)

This procedure ensures memory efficiency due to how reference to GPU tensors (1) is dropped before GPU tensors (3) are allocated.

However, in the downstream case of transformers, PretrainedModel.save_pretrained keeps a copy of the original GPU tensors (1) for longer than they are needed. This long-lived reference to (1) results in references to both (1) and (3) being alive at the same time, leading to increased memory usage.

Changes

• Change default behavior of get_state_dict_from_offload to not move parameter tensors by default
  • This ensures compatibility with transformers versions which have been released prior to and after Fix save_pretrained for partially offloaded models transformers#34890

Testing

save_offloaded.py

import torch
from transformers import AutoModelForCausalLM

model = AutoModelForCausalLM.from_pretrained("nvidia/Llama-3.1-Nemotron-70B-Instruct-HF", torch_dtype="auto", device_map="auto")
print(model.hf_device_map)
"""
{'model.embed_tokens': 0, 'model.layers.0': 0, 'model.layers.1': 0, 'model.layers.2': 0, 'model.layers.3': 0, 'model.layer
s.4': 0, 'model.layers.5': 0, 'model.layers.6': 0, 'model.layers.7': 0, 'model.layers.8': 0, 'model.layers.9': 0, 'model.l
ayers.10': 0, 'model.layers.11': 0, 'model.layers.12': 0, 'model.layers.13': 0, 'model.layers.14': 0, 'model.layers.15': 0
, 'model.layers.16': 0, 'model.layers.17': 0, 'model.layers.18': 0, 'model.layers.19': 0, 'model.layers.20': 0, 'model.lay
ers.21': 0, 'model.layers.22': 0, 'model.layers.23': 0, 'model.layers.24': 0, 'model.layers.25': 0, 'model.layers.26': 0, 
'model.layers.27': 0, 'model.layers.28': 0, 'model.layers.29': 0, 'model.layers.30': 0, 'model.layers.31': 0, 'model.layer
s.32': 0, 'model.layers.33': 0, 'model.layers.34': 0, 'model.layers.35': 0, 'model.layers.36': 0, 'model.layers.37': 0, 'm
odel.layers.38': 0, 'model.layers.39': 0, 'model.layers.40': 0, 'model.layers.41': 0, 'model.layers.42': 'cpu', 'model.lay
ers.43': 'cpu', 'model.layers.44': 'cpu', 'model.layers.45': 'cpu', 'model.layers.46': 'cpu', 'model.layers.47': 'cpu', 'm
odel.layers.48': 'cpu', 'model.layers.49': 'cpu', 'model.layers.50': 'cpu', 'model.layers.51': 'cpu', 'model.layers.52': '
cpu', 'model.layers.53': 'cpu', 'model.layers.54': 'cpu', 'model.layers.55': 'cpu', 'model.layers.56': 'cpu', 'model.layer
s.57': 'cpu', 'model.layers.58': 'cpu', 'model.layers.59': 'cpu', 'model.layers.60': 'cpu', 'model.layers.61': 'cpu', 'mod
el.layers.62': 'cpu', 'model.layers.63': 'cpu', 'model.layers.64': 'cpu', 'model.layers.65': 'cpu', 'model.layers.66': 'cp
u', 'model.layers.67': 'cpu', 'model.layers.68': 'cpu', 'model.layers.69': 'cpu', 'model.layers.70': 'cpu', 'model.layers.71': 'cpu', 'model.layers.72': 'cpu', 'model.layers.73': 'cpu', 'model.layers.74': 'cpu', 'model.layers.75': 'cpu', 'model.layers.76': 'cpu', 'model.layers.77': 'cpu', 'model.layers.78': 'cpu', 'model.layers.79': 'cpu', 'model.norm': 'cpu', 'model.rotary_emb': 'cpu', 'lm_head': 'cpu'}
"""

torch.cuda.memory._record_memory_history()
model.save_pretrained("save_dir")
torch.cuda.memory._dump_snapshot(f"align.pickle")
torch.cuda.memory._record_memory_history(enabled=None)

transformers <= 4.46.0 | accelerate < 1.1.0

• Original references are not moved

transformers <= 4.46.0 | accelerate == 1.1.0, 1.1.1

• Results in bug
• Original references are not garbage collected
• Calling save_pretrained results in OOM (>80GiB)

transformers <= 4.46.0 | accelerate > 1.1.1 (this branch)

• Original references are not moved

transformers > 4.46.0 | accelerate > 1.1.1 (this branch)

• Original references are not moved

Who can review?

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[muellerzr]

Oh dear, thanks for the follow up! We have a release planned shortly after thanksgiving so this can be included in it :)

[SunMarc]

Thanks for the bug fix !

[muellerzr]

cc @ArthurZucker for final