SUMIN/HuggingFace_transformer
1
0
Fork 0
Code Issues Pull Requests Actions 7 Packages Projects Releases Wiki Activity

Use Accelerate in from_pretrained for big model inference (#17341)

Browse Source 
* Initial work

* More or less finished with first draft

* Update src/transformers/modeling_utils.py

Co-authored-by: Stas Bekman <stas00@users.noreply.github.com>

* Update src/transformers/modeling_utils.py

Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com>

* Fix randomly initialized weights

* Update src/transformers/modeling_utils.py

Co-authored-by: Lysandre Debut <lysandre.debut@reseau.eseo.fr>

* Address review comments

* Rename DeepSpeed folder to temporarily fix the test issue?

* Revert to try if Accelerate fix works

* Use latest Accelerate release

* Quality and fixes

* Style

* Quality

* Add doc

* Test + fix

* More blocks

Co-authored-by: Stas Bekman <stas00@users.noreply.github.com>
Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com>
Co-authored-by: Lysandre Debut <lysandre.debut@reseau.eseo.fr>
This commit is contained in:
Sylvain Gugger
2022-05-23 14:32:21 -04:00
committed by GitHub
parent 2e7e4280aa
commit 56f50590d5
9 changed files with 270 additions and 41 deletions
Download Patch File Download Diff File Expand all files Collapse all files

69
docs/source/en/main_classes/model.mdx
View File

@@ -38,6 +38,75 @@ for text generation, [`~generation_utils.GenerationMixin`] (for the PyTorch mode
<a id='from_pretrained-torch-dtype'></a>
### Large model loading
In Transformers 4.20.0, the [`~PreTrainedModel.from_pretrained`] method has been reworked to accommodate large models using [Accelerate](https://huggingface.co/docs/accelerate/big_modeling). This requires Accelerate >= 0.9.0 and PyTorch >= 1.9.0. Instead of creating the full model, then loading the pretrained weights inside it (which takes twice the size of the model in RAM, one for the randomly initialized model, one for the weights), there is an option to create the model as an empty shell, then only materialize its parameters when the pretrained weights are loaded.
This option can be activated with `low_cpu_mem_usage=True`. The model is first created on the Meta device (with empty weights) and the state dict is then loaded inside it (shard by shard in the case of a sharded checkpoint). This way the maximum RAM used is the full size of the model only.
```py
from transformers import AutoModelForSeq2SeqLM
t0pp = AutoModelForSeq2SeqLM.from_pretrained("bigscience/T0pp", low_cpu_mem_usage=True)
```
Moreover, you can directly place the model on different devices if it doesn't fully fit in RAM (only works for inference for now). With `device_map="auto"`, Accelerate will determine where to put each layer to maximize the use of your fastest devices (GPUs) and offload the rest on the CPU, or even the hard drive if you don't have enough GPU RAM (or CPU RAM). Even if the model is split across several devices, it will run as you would normally expect.
When passing a `device_map`, `low_cpu_mem_usage` is automatically set to `True`, so you don't need to specify it:
```py
from transformers import AutoModelForSeq2SeqLM
t0pp = AutoModelForSeq2SeqLM.from_pretrained("bigscience/T0pp", device_map="auto")
```
You can inspect how the model was split across devices by looking at its `hf_device_map` attribute:
```py
t0pp.hf_device_map
```
```python out
{'shared': 0,
'decoder.embed_tokens': 0,
'encoder': 0,
'decoder.block.0': 0,
'decoder.block.1': 1,
'decoder.block.2': 1,
'decoder.block.3': 1,
'decoder.block.4': 1,
'decoder.block.5': 1,
'decoder.block.6': 1,
'decoder.block.7': 1,
'decoder.block.8': 1,
'decoder.block.9': 1,
'decoder.block.10': 1,
'decoder.block.11': 1,
'decoder.block.12': 1,
'decoder.block.13': 1,
'decoder.block.14': 1,
'decoder.block.15': 1,
'decoder.block.16': 1,
'decoder.block.17': 1,
'decoder.block.18': 1,
'decoder.block.19': 1,
'decoder.block.20': 1,
'decoder.block.21': 1,
'decoder.block.22': 'cpu',
'decoder.block.23': 'cpu',
'decoder.final_layer_norm': 'cpu',
'decoder.dropout': 'cpu',
'lm_head': 'cpu'}
```
You can also write your own device map following the same format (a dictionary layer name to device). It should map all parameters of the model to a given device, but you don't have to detail where all the submosules of one layer go if that layer is entirely on the same device. For instance, the following device map would work properly for T0pp (as long as you have the GPU memory):
```python
device_map = {"shared": 0, "encoder": 0, "decoder": 1, "lm_head": 1}
```
Another way to minimize the memory impact of your model is to instantiate it at a lower precision dtype (like `torch.float16`).
### Model Instantiation dtype
Under Pytorch a model normally gets instantiated with `torch.float32` format. This can be an issue if one tries to