* Greedy is implemented.

* Add early stopping, sort beams by probability

* Pass on temperature from UI. Early stopping.

* Kill some more dead code.

* Typo

* Fix lint.

* Add Aim logging to progress_bar

* Add Aim logging arguments to validate() method progress_bar

* Add aim package to install requirements

* [fix] Add imports from base_progress_bar

* Add Aim usage mini-guide

* Add len method to AimProgressBarWrapper

* Fix linter issues

Co-authored-by: Gor Arakelyan <arakelyangor10@gmail.com>

This reverts commit 1aa40c2a.

* added changes for resharding MP parts

* changes

* changes

* changes

* 66b weights, md5sum

* add baselines logbook

* add link to arxiv, link to chronicles on project readme

* [api] Support temp=0. Return errors as json.

* Update hub_utils.py

* Save work

* Correctly handle.

Singleton checkpoint needs to include decoder.version for single-ton checkpoint to run correctly (#164)

* Singleton checkpoint needs to include decoder.version

If we don't transfer the `"decoder.version"` to the singleton checkpoint, a very sneaky bug happens which was found by @thomasw21 as part of this PR: https://github.com/huggingface/transformers/pull/17785

If the `decoder.version` param is not present in the state_dict it follows that upon loading the single-ton checkpoint the loaded layer_norm is set to `None` here: https://github.com/facebookresearch/metaseq/blob/e0c4f6b0e4c523906ad8d561f727e3f2ac3a8e73/metaseq/models/transformer.py#L932

So it's absolutely crucial that we include this variable.

I will update all of the converted HF checkpoints here later today and then I think we can be sure that OPT works correctly :partying_face: 
https://huggingface.co/models?other=opt_metasq



* Update convert_to_singleton.py

Co-authored-by: Stephen Roller <roller@fb.com>

* fix Namespace missing import

* add noqa comment to Namespace import line

* document attention masking first commit

* Fixing bugs, formatting changes

* Fixing bug

Addressing comments, adding documentation

Fixing comments

Fixing lint

Fixing positional embedding reset

Bug fix for positional embeddings

Adding comments

Co-authored-by: Ramakanth Pasunuru <ramakanth.1729@gmail.com>

* fix beam search

* add whitespace back

* Update test_sequence_generator.py

Co-authored-by: Stephen Roller <roller@fb.com>

* making prompt_len independent of batchfy implmentation

* modify for echo=True case

* returning logprobs, to support logprob input

* setting need_logprobs depends on each request to save memory

* fix progress bar exception when both tensorboard and wandb are turned on

* prioritize wandb progress bar over tensorboard progress bar when both are set

* Add eos mode for src tgt dataset

* lint

* Addressed comments

* lint

* fix docker build fails

* revert metaseq branch tomain in dockerfile

* limit max version of hydra-core

Co-authored-by: qiushuo <qiushuo@microsoft.com>

* remove unused build_progress_bar method

* remove noop/none logging as a log format option

* make log_format default to json and remove default_log_format arg

* move progress_bar to submodule

* rename to base_progress_bar

* module rename

* move out json progress bar

* move out tensorboard progress bar

* move out wandb progress wrapper

* split out helpers to utils, avoid circular import

* fix broken init

* remove module rename for progress bar

* move get_precise_epoch to higher level utils

* lint

* remove no progress bar flag

* fix more imports

* cleanup

* remove utils

* add license

* pin omegaconf to 2.1.1

* pin tensorboard to 2.8.0

* protobuf pin to 3.20.1

Based on the setup instructions

* Add maximum for example proportional sampling

* lint

* Addressed comments

* include diff with black check

* fix lint

* set defaults for scale window, and init scale

* remove more default logic for scale window

* also decrease scale window with loss scale

* remove floating point error for hitting min loss scale

* also log out scale_window

* 0.03125 -> 2 ** -5