GitLab

* docs: ko: tasks/mask_generation.md

* feat: nmt draft

* fix : toc local

* fix : manual edits

* fix : ko-toctree

* fix: resolve suggestions
Co-authored-by: boyunJang <gobook1234@naver.com>
Co-authored-by: Chaewon Song <chaewon1019@ewhain.net>

* fix: resolve suggestions
Co-authored-by: boyunJang <gobook1234@naver.com>
Co-authored-by: Chaewon Song <chaewon1019@ewhain.net>

* fix: resolve suggestions

* fix: resolve suggestions

* fix: resolve suggestions

---------
Co-authored-by: boyunJang <gobook1234@naver.com>
Co-authored-by: Chaewon Song <chaewon1019@ewhain.net>

Revert "fixes to properly shard FSDP across cpu and meta for cpu_effcient_loading for prequantized 4bit (#32276)" (#32477)

* Revert "fixes to properly shard FSDP across cpu and meta for cpu_efficient_loading for prequantized 4bit (#32276)"

This reverts commit 62c60a30

.

We uncovered an issue with this change that caused our training runs to hang.

* `is_torchdynamo_compiling` -- cast a wide exception net (#32476)

* cast a wide net

* make fix-copies with a few manual changes

* add copied from

---------
Co-authored-by: Joao Gante <joaofranciscocardosogante@gmail.com>

* cast a wide net

* make fix-copies with a few manual changes

* add copied from

Update nllb.md

* Migrate import checks to secondary accelerate calls

* better errs too

* Revert, just keep the import checks + remove accelerate-specific things

* Rm extra'

* Empty commit for ci

* Small nits

* Final

* add new model like

* draft cuda forward - mismatched keys (sharding on conv1)

* match keys successfully

* fix split

* get generation/forward running (wrong gens, norm?)

* :update

* some refactoring

* fixes

* works up until copy to cache

* fix

* update

* NON WORKING VERSION

* version that work?

* nit

* fix config

* fix conversion script

* working cuda forward

* nit

* update

* simplifcation

* make mamba slow simple work

* no einops

* todo

* fix style

* no einops

* update fix no einsum

* nit

* remove einops

* bug: scan_output differs strongly

* add rms norm option

* fix fast + slow generation with and w/o cache ✔



* draft integration tests

* remove a big chunk of the einsum

* fix slow, fast generations, without any einsum

* fix copies

* fix structure

* fix up modeling and tests

* fix tests

* clamping is indeed worse

* recover mamba2 cache test

* fix copies

* no cache position (yet)

* fix tf tests

* fix matmul for generate

* fixup

* skip cache tests for now

* [run-slow]mamba2

* tune out hidden states for padding

* test batched generation

* propagate attention mask changes

* fix past length

* fix integration test

* style

* address comments

* update readme

* add mamba2 version check

* fix tests

* [run-slow]mamba2

* skip edge tests

* [run-slow]mamba2

* last fixup

* [run-slow]mamba2

* update README

---------
Co-authored-by: Arthur Zucker <arthur.zucker@gmail.com>

* Add nemotron support

* fix inference

* add unit test

* add layernorm1p as a class to avoid meta device mismatch

* test fixed

* Add copied_from statements

* remove pretraining_tp args

* remove nemotronlayernorm

* force LN computation done in FP32

* remove nemotrontokenizer and use llamatokenizer

* license update

* add option for kv_channels for minitron8b

* remove assert

* o_proj fixed

* o_proj reshape

* add gated_proj option

* typo

* remove todos

* fix broken test after merging latest main

* remove nezha/nat after meging main

* chnage default config to 15b model

* add nemo conversion script

* rename conversion script

* remove gate_proj option

* pr comment resolved

* fix unit test

* rename kv_channels to head_dim

* resolve PR issue

* add nemotron md

* fix broken tests

* refactor rope for nemotron

* test fix

* remove linearscaling

* whitespace and import

* fix some copied-from

* code style fix

* reformatted

* add position_embedding to nemotronattention

* rope refactor to only use config, copied-from fix

* format

* Run make fix-copies

* nemotron md with autodoc

* doc  fix

* fix order

* pass check_config_docstrings.py

* fix config_attributes

* remove all llama BC related code

* Use PreTrainedTokenizerFast

* ruff check examples

* conversion script update

* add nemotron to toctree

deps_2

* BLIP preprocess

* BIT preprocess

* BRIDGETOWER preprocess

* CHAMELEON preprocess

* CHINESE_CLIP preprocess

* CONVNEXT preprocess

* DEIT preprocess

* DONUT preprocess

* DPT preprocess

* FLAVA preprocess

* EFFICIENTNET preprocess

* FUYU preprocess

* GLPN preprocess

* IMAGEGPT preprocess

* INTRUCTBLIPVIDEO preprocess

* VIVIT preprocess

* ZOEDEPTH preprocess

* VITMATTE preprocess

* VIT preprocess

* VILT preprocess

* VIDEOMAE preprocess

* VIDEOLLAVA

* TVP processing

* TVP fixup

* SWIN2SR preprocess

* SIGLIP preprocess

* SAM preprocess

* RT-DETR preprocess

* PVT preprocess

* POOLFORMER preprocess

* PERCEIVER preprocess

* OWLVIT preprocess

* OWLV2 preprocess

* NOUGAT preprocess

* MOBILEVIT preprocess

* MOBILENETV2 preprocess

* MOBILENETV1 preprocess

* LEVIT preprocess

* LAYOUTLMV2 preprocess

* LAYOUTLMV3 preprocess

* Add test

* Update tests

* add flash attention check

* fix

* fix

* add the missing marker

* bug fix

* add one more

* remove order

* add one more

fix: add new llava like model bug

* draft

* updates

* works?

* try adding python example in hidden section

* another try

* hwo do i render python

* format as html code?

* Update docs/source/en/kv_cache.md
Co-authored-by: Joao Gante <joaofranciscocardosogante@gmail.com>

* Update docs/source/en/kv_cache.md
Co-authored-by: Joao Gante <joaofranciscocardosogante@gmail.com>

* Update docs/source/en/kv_cache.md
Co-authored-by: Joao Gante <joaofranciscocardosogante@gmail.com>

* Update docs/source/en/kv_cache.md
Co-authored-by: Joao Gante <joaofranciscocardosogante@gmail.com>

* Update docs/source/en/kv_cache.md
Co-authored-by: Joao Gante <joaofranciscocardosogante@gmail.com>

* one more small update

* should render hidden secrtion now

* add outputs

* fix links

* check links

* update all links

* update with offloaded cache

* all cache is importable, so they appear in docs

* fix copies

* docstring...

---------
Co-authored-by: Joao Gante <joaofranciscocardosogante@gmail.com>

* Respect the config's attn if set

* Update test - can override in from_config

* Fix

Fixed tokenizertests for luke, mluke models.

* fix: persist embedding type of MBartConditonalGeneration after resize

* fix: persist embedding type of BartConditonalGeneration after resize

I always forget what typical values are, and I have to look at the paper everytime. This will be a helpful reminder.

* save total_vocab_size = vocab_size + user added tokens to speed up operation

* updating length when added_tokens_decoder is set

* add test len(tokenizer)

fix phi

* fix: SeamlessM4TFeatureExtractor stride remainder

* Added attention mask size test

* Reran ruff for style correction

Bump keras in /examples/research_projects/decision_transformer

Bumps [keras](https://github.com/keras-team/keras) from 2.8.0 to 2.13.1.
- [Release notes](https://github.com/keras-team/keras/releases)
- [Commits](https://github.com/keras-team/keras/compare/v2.8.0...v2.13.1

)

---
updated-dependencies:
- dependency-name: keras
  dependency-type: direct:production
...
Signed-off-by: dependabot[bot] <support@github.com>
Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com>

MixtralFlashAttention2: put "plus 1" inside parentheses when calculating rotary_seq_len, allowing None position_ids input. (#31500)

* Mixtral: remove unnecessary plus 1 when calculating rotary_seq_len, allowing position_ids=None (no auto position_ids generation could be unsafe)

* fix typo [:-1] to [:, -1]

* to meet formatting requirement

* to meet formatting requirement

* remove white space

* MixtralFlashAttention2: put "+ 1" inside parentheses when calculating rotary_seq_len, allowing None position_ids input. Fix format/style issue.

* propagate to startcoder2, phi3, mixtral and qwen2

* update qwen2_moe

fix: (issue #32124) Exception raised when running `transformers/examples/flax/language-modeling/t5_tokenizer_model.py`. (#32157)

fix: Exception raised when running .

* up

* style

* stopping

tests! :D

nits

* Test this zach

* Test for improper init w/o zero3

* Move back

* Apply suggestions from code review
Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>

* Get rid of stars in warning

* Make private

* Make clear

---------
Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>

* fixed hybrid cache init, added test

* Fix Test Typo

---------
Co-authored-by: Aaron Haag <aaron.haag@siemens.com>

* Initial implementation of OffloadedCache

* enable usage via cache_implementation

* Address feedback, add tests, remove legacy methods.

* Remove flash-attn, discover synchronization bugs, fix bugs

* Prevent usage in CPU only mode

* Add a section about offloaded KV cache to the docs

* Fix typos in docs

* Clarifications and better explanation of streams

* Fix conflicting key in init kwargs in PreTrainedTokenizerBase

* Update code to check for callable key in save_pretrained

* Apply PR suggestions

* Invoke CI

* Updates based on PR suggestion

empty list in defaults

* Remove TPU device map for saving tokenizer config

* Update tokenization_utils_base.py

* Fix error msg when passing non-string device into tokenizer

* Fix error message for non-string tokenizer device

* Print out tokenizer device type in error msg

* Update tokenization_utils_base.py

Co-authored-by: Guoming Zhang <37257613+nv-guomingz@users.noreply.github.com>

* Remove size check between attn_weights and kv_seq_len

* add unit tests