* Improve CTRL doctests
* Fix `CTRLForSequenceClassification` flakiness with inconsistent losses
* Remove unused
* Fixup
* Add CTRL to documentation_tests.txt
* Fix control code not being first
* Add output assertions
* Change from sshleifer/tiny-ctrl -> ctrl
* Run `make fixup`
* apply `list` to output logits shape for clarity
* Reduce output loss precision to make assertion more robust
* Add assertion of control code being first
* Fix docstyle
* upper case sentence following control code
* Weird bug fixes
* Add a better generation example
Co-authored-by: Yih-Dar <2521628+ydshieh@users.noreply.github.com>
* Required the values GPTJ unfortunately cannot run the model =)
* Added the file to the doc tests
* Run Fixup and Style
* Fixed with the test versions of gptj. Ran Style and Fixup.
* Trigger ci
* A Minor Change to License
* Fixed spacing added to the benchmark_utils. Then refactored tests to const variables.
* Removed strings that were included as default parameters anyways.
Co-authored-by: ArEnSc <xx.mike.chung.xx@gmail.com>
* Fix setters of *_token_id properties of SpecialTokensMixin
* Test setters of common tokens ids
* Move to a separate test checks of setters of tokens ids
* Add independent test for ByT5
* Add Canine test
* Test speech to text
* Change the chunk_iter function to handle
the subtle cases where the last chunk gets ignored since all the
data is in the `left_strided` data.
We need to remove the right striding on the previous item.
* Remove commented line.
This avoids an unnecessary call and avoids problems during
initialization of class hierarchies.
Co-authored-by: Samuel Melm <samuel.melm@stud.uni-heidelberg.de>
* First Pass All Tests Pass
* WIP
* Adding file to documentation tests
* Change the base model for the example in the doc test.
* Fix Code Styling by running
make fixup
* Called Style
* Reverted to gpt2 model rather than distill gpt2
Then used a token classification model over a sequence model for an example.
* Fix Styling Issue
* Hopefully ignores the formatting issue.
Co-authored-by: ArEnSc <xx.mike.chung.xx@gmail.com>
* Fix t5 shard on TPU Pods
The current script doesn't work properly on a TPU pod because the global batch is not divided correctly per host.
This pull request fixes this issue by dividing the global batch to each host before it is shared on each host.
* fix style
Co-authored-by: ahmed-elnaggar <ahmed.elnaggar@allianz.com>
I create an archive of older checkpoints during training the checkpoint has a name with `f"{checkpoint_prefix}-*.zip/.tar `
previously `glob(f"{checkpoint_prefix}-*")` takes all files/folders starting with the name checkpoint, and later `shutil.rmtree(checkpoint)` takes a folder name; since at some point it my get a zip file; it crashes training; adding this `if os.path.isdir(x)` allows only folders on `glob_checkpoints`
* add simple multi gpu complet
* add human_eval_multi_gpu
* use copy strategy to distribute across gpu, to avoid padding
* add doc string
* update code style
* use task id to arrange output
* truncate input to avoid zero pad
* Stop the copy mechanism
* update style
* restore copies to scale better in distributed mode
* update style
* replace human eval
* Apply suggestions from code review
1. Tokenize all input at the same time
2. use attention_mask to get the input length
3. other small fixes
Co-authored-by: Leandro von Werra <lvwerra@users.noreply.github.com>
* correct typo and update docstring
* update code style
* remove num sample division constraint
* remove max len calculation
* use accelerator.gather once to speed up
* use accelerate set_seed; update accelerate version
* correct gather bug
Co-authored-by: Leandro von Werra <lvwerra@users.noreply.github.com>
* update proto sentencepiece model
* Revert "update proto sentencepiece model"
This reverts commit b07f671747fec35773d0b3d4788b8b15aefa0229.
* add check
* add test
* Revert "Revert "update proto sentencepiece model""
This reverts commit 46108257b8927b73627ec8f4f3eed53a95fc700d.
* test for log level
* test for log level 2
* warning at the warning level
* clean
* format
* add explanation in docstring
* Fixed some bugs involving saving during epochs
* Added tests mimicking the existing examples tests
* Added in json exporting to all `no_trainer` examples for consistency
