GitLab

BrideTower Model parallelism logits device for loss calculation

* Add model with cli tool

* Remove unwanted stuff

* Add new code

* Remove inference runner

* Style

* Fix checks

* Test updates

* make fixup

* fix docs

* fix doc

* fix test

* hopefully fix pipeline tests

* refactor

* fix CIs

* add comment

* rename to `GPTBigCodeForCausalLM`

* correct readme

* make fixup + docs

* make fixup

* fixes

* fixes

* Remove pruning

* Remove import

* Doc updates

* More pruning removal

* Combine copies

* Single MQA implementation, remove kv cache pre-allocation and padding

* Update doc

* Revert refactor to match gpt2 style

* Merge back key and value caches, fix some type hints

* Update doc

* Fix position ids pith padding (PR 21080)

* Add conversion script temporarily

* Update conversion script

* Remove checkpoint conversion

* New model

* Fix MQA test

* Fix copies

* try fix tests

* FIX TEST!!

* remove  `DoubleHeadsModel`

* add MQA tests

* add slow tests

* clean up

* add CPU checker

* final fixes

* fixes

- fix GPU issue
- fixed slow tests
- skip disk offload

* fix final issue

* Simplify and comment baddbmm fix

* Remove unnecessary code

* Transpose tweaks

* Use beta=1 on cpu, improve tests

---------
Co-authored-by: younesbelkada <younesbelkada@gmail.com>

moved labels to the same device as logits for BLOOM, GPT Neo, GPT NeoX, RoBERTa and VIT models (#22663)

moved labels to the same device as logits

add API warning

* Fix default attention mask size

* fixup

* add a test to make sure that even if attention mask are not provided, works

* style

* do not push special file

* Update src/transformers/tokenization_utils_base.py
Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>

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

* Small nit,
Fixes #21986

* Update src/transformers/pipelines/__init__.py

docs: feat: Korean pipeline_tutorial
Co-authored-by: Jungnerd <46880056+jungnerd@users.noreply.github.com>
Co-authored-by: Hyeonseo Yun <0525_hhgus@naver.com>
Co-authored-by: gabrielwithappy <102908949+gabrielwithappy@users.noreply.github.com>
Co-authored-by: Na Yeon Han <nayeon2.han@gmail.com>

* fix

---------
Co-authored-by: ydshieh <ydshieh@users.noreply.github.com>

* (feat): Move labels to the same device as logits

* Trigger CI

* Trigger CI

* Trigger CI

* (feat): Making changes for Blip2

translate the autoclass_tutorial and fix the typo of the quicktour

fix fsdp

* Update tiny model summary file for recent models

---------
Co-authored-by: ydshieh <ydshieh@users.noreply.github.com>

fix slow tests and doctests

add safety checker

* Auto. add and update pipeline_model_mapping

* Fix style and quality

* Finalize (comments)

* fix

---------
Co-authored-by: ydshieh <ydshieh@users.noreply.github.com>

* Add TFBlipForConditionalGeneration

* update pipeline_model_mapping

* Add import

* Revert changes in GPTSanJapaneseTest

---------
Co-authored-by: ydshieh <ydshieh@users.noreply.github.com>

fix broken link

* Final Tiny things

---------
Co-authored-by: ydshieh <ydshieh@users.noreply.github.com>

* Add out_indices to backbones, deprecate out_features

* Update - can specify both out_features and out_indices but not both

* Add backbone mixin tests

* Test tidy up

* Add test_backbone for convnext

* Remove redefinition of method

* Update for Dinat and Nat backbones

* Update tests

* Smarter indexing

* Add checks on config creation for backbone

* PR comments

* Adding Llama FastTokenizer support.

- Requires https://github.com/huggingface/tokenizers/pull/1183 version
- Only support byte_fallback for llama, raise otherwise (safety net).
- Lots of questions are special tokens

How to test:

```python

from transformers.convert_slow_tokenizer import convert_slow_tokenizer
from transformers import AutoTokenizer
from tokenizers import Tokenizer

tokenizer = AutoTokenizer.from_pretrained("huggingface/llama-7b")

if False:
    new_tokenizer = Tokenizer.from_file("tok.json")
else:
    new_tokenizer = convert_slow_tokenizer(tokenizer)
    new_tokenizer.save("tok.json")

strings = [
    "This is a test",
    "生活的真谛是",
    "生活的真谛是[MASK]。",
    # XXX: This one is problematic because of special tokens
    # "<s> Something something",
]

for string in strings:
    encoded = tokenizer(string)["input_ids"]
    encoded2 = new_tokenizer.encode(string).ids

    assert encoded == encoded2, f"{encoded} != {encoded2}"

    decoded = tokenizer.decode(encoded)
    decoded2 = new_tokenizer.decode(encoded2)

    assert decoded.strip() == decoded2, f"{repr(decoded)} != {repr(decoded2)}"
```

The converter + some test script.

The test script.

Tmp save.

Adding Fast tokenizer + tests.

Adding the tokenization tests.

Correct combination.

Small fix.

Fixing tests.

Fixing with latest update.

Rebased.

fix copies + normalized added tokens  + copies.

Adding doc.

TMP.

Doc + split files.

Doc.

Versions + try import.

Fix Camembert + warnings -> Error.

Fix by ArthurZucker.

Not a decorator.

* Fixing comments.

* Adding more to docstring.

* Doc rewriting.

feat(model parallelism): moving the labels to the same device as the logits for gpt2 and bart (#22591)

* Use native TF checkpoints for the TF tests

* Remove unneeded exceptions

* add deplot + matcha on `transformers`

* more docs

* correct path

* Update docs/source/en/model_doc/deplot.mdx
Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>

* fix

* use auto processor

* Update docs/source/en/model_doc/matcha.mdx
Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>

* make fixup

* Update docs/source/en/model_doc/deplot.mdx
Co-authored-by: NielsRogge <48327001+NielsRogge@users.noreply.github.com>

* add correct names

---------
Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>
Co-authored-by: NielsRogge <48327001+NielsRogge@users.noreply.github.com>

* Adding support for BPE merge creation from scores instead of ids.

* Revert warn -> raise.

* Update src/transformers/convert_slow_tokenizer.py

* Quality.

Fixes a typo in one of the BLIP pretrained checkpoint names

* Update run_speech_recognition_ctc.py

Make sure all processes wait until data is saved before loading the processor from the output_dit

* Make sure all processes wait until data is saved before loading the processor from the output_dit

* Update run_speech_recognition_ctc.py

* Update run_speech_recognition_seq2seq.py

Co-authored-by: gabrielwithappy <102908949+gabrielwithappy@users.noreply.github.com>

The logger prints a summary at the beginning of training that displays some info such as number of examples, number of parameters, total number of steps, etc. Those numbers can be quite large and difficult to read. I added a thousand separator to improve readability for the following:
- num_examples
- num_train_epochs
- per_device_train_batch_size
- total_train_batch_size
- max_steps
- num_trainable_params

* Re-enable skipped test and fix the hidden state shape issue

* Actually fix the bug instead of just doing something wrong

* Fix inverted conditional in TF common test!

* Make the same change in the PT tests file

* Make sure hidden states for GPT2 have the same output shape in PT/TF

* Minor fix to PT implementation of token classification loss

* Skip loss equivalence test for TFHubert because it keeps overflowing to inf

* Compute LM loss for TF the (weird) way it's computed in PT

* Skip loss equivalence test for Wav2Vec2 for the same reason as Hubert

* Fix - don't try to access the hidden states property when output is a tuple