Add Mistral GPT-2 Stability Tweaks (#13573)

* Add layer-wise scaling

* Add reorder & upcasting argument

* Add OpenAI GPT-2 weight initialization scheme

* start `layer_idx` count at zero for consistency

* disentangle attn and reordered and upscaled attn function

* rename `scale_attn_by_layer` to `scale_attn_by_layer_id`

* make autocast from amp compatible with pytorch<1.6

* fix docstring

* style fixes

* Add fixes from PR feedback, style tweaks

* Fix doc whitespace

* Reformat

* First pass scale_attn_by_layer_idx and reorder_and_upcast_attn tests

* Rename scale_attn_by_layer_idx, add tip

* Remove extra newline

* add test for weight initialization

* update code format

* add assert check weights are fp32

* remove assert

* Fix incorrect merge

* Fix shape mismatch in baddbmm

* Add generation test for Mistral flags

Co-authored-by: leandro <leandro.vonwerra@spoud.io>
Co-authored-by: Keshav Santhanam <keshav2@stanford.edu>
Co-authored-by: J38 <jebolton@stanford.edu>

docs/source/model_doc/gpt2.rst

@@ -41,6 +41,8 @@ Tips:
   pre-computed values in the context of text generation. For PyTorch, see `past_key_values` argument of the
   :meth:`~transformers.GPT2Model.forward` method, or for TF the `past` argument of the
   :meth:`~transformers.TFGPT2Model.call` method for more information on its usage.
+- Enabling the `scale_attn_by_inverse_layer_idx` and `reorder_and_upcast_attn` flags will apply the training stability
+  improvements from `Mistral <https://github.com/stanford-crfm/mistral/>`__ (for PyTorch only).
 
 `Write With Transformer <https://transformer.huggingface.co/doc/gpt2-large>`__ is a webapp created and hosted by
 Hugging Face showcasing the generative capabilities of several models. GPT-2 is one of them and is available in five