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

[models] respect dtype of the model when instantiating it (#12316)

Browse Source 
* [models] respect dtype of the model when instantiating it

* cleanup

* cleanup

* rework to handle non-float dtype

* fix

* switch to fp32 tiny model

* improve

* use dtype.is_floating_point

* Apply suggestions from code review

Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>

* fix the doc

* recode to use explicit torch_dtype_auto_detect, torch_dtype args

* docs and tweaks

* docs and tweaks

* docs and tweaks

* merge 2 args, add docs

* fix

* fix

* better doc

* better doc

Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>
This commit is contained in:
Stas Bekman
2021-06-28 20:11:21 -07:00
committed by GitHub
parent 31c3e7e75b
commit 7682e97702
8 changed files with 221 additions and 26 deletions
Download Patch File Download Diff File Expand all files Collapse all files

33
docs/source/main_classes/model.rst
View File

@@ -1,4 +1,4 @@
..
..
Copyright 2020 The HuggingFace Team. All rights reserved.
Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
@@ -38,6 +38,37 @@ PreTrainedModel
:members:
.. _from_pretrained-torch-dtype:
Model Instantiation dtype
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
Under Pytorch a model normally gets instantiated with ``torch.float32`` format. This can be an issue if one tries to
load a model whose weights are in fp16, since it'd require twice as much memory. To overcome this limitation, you can
either explicitly pass the desired ``dtype`` using ``torch_dtype`` argument:
.. code-block:: python
model = T5ForConditionalGeneration.from_pretrained("t5", torch_dtype=torch.float16)
or, if you want the model to always load in the most optimal memory pattern, you can use the special value ``"auto"``,
and then ``dtype`` will be automatically derived from the model's weights:
.. code-block:: python
model = T5ForConditionalGeneration.from_pretrained("t5", torch_dtype="auto")
Models instantiated from scratch can also be told which ``dtype`` to use with:
.. code-block:: python
config = T5Config.from_pretrained("t5")
model = AutoModel.from_config(config)
Due to Pytorch design, this functionality is only available for floating dtypes.
ModuleUtilsMixin
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~