* Initial commit
* Make some fixes
* Make PT model full forward pass
* Drop TF & Flax implementation, fix copies etc
* Add Flax model and update some corresponding stuff
* Drop some TF things
* Update config and flax local attn
* Add encoder_attention_type to config
* .
* Update docs
* Do some cleansing
* Fix some issues -> make style; add some docs
* Fix position_bias + mask addition + Update tests
* Fix repo consistency
* Fix model consistency by removing flax operation over attn_mask
* [WIP] Add PT TGlobal LongT5
* .
* [WIP] Add flax tglobal model
* [WIP] Update flax model to use the right attention type in the encoder
* Fix flax tglobal model forward pass
* Make the use of global_relative_attention_bias
* Add test suites for TGlobal model
* Fix minor bugs, clean code
* Fix pt-flax equivalence though not convinced with correctness
* Fix LocalAttn implementation to match the original impl. + update READMEs
* Few updates
* Update: [Flax] improve large model init and loading #16148
* Add ckpt conversion script accoring to #16853 + handle torch device placement
* Minor updates to conversion script.
* Typo: AutoModelForSeq2SeqLM -> FlaxAutoModelForSeq2SeqLM
* gpu support + dtype fix
* Apply some suggestions from code review
Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>
Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com>
* * Remove (de)parallelize stuff
* Edit shape comments
* Update README.md
* make fix-copies
* Remove caching logic for local & tglobal attention
* Apply another batch of suggestions from code review
* Add missing checkpoints
* Format converting scripts
* Drop (de)parallelize links from longT5 mdx
* Fix converting script + revert config file change
* Revert "Remove caching logic for local & tglobal attention"
This reverts commit 2a619828f6ddc3e65bd9bb1725a12b77fa883a46.
* Stash caching logic in Flax model
* Make side relative bias used always
* Drop caching logic in PT model
* Return side bias as it was
* Drop all remaining model parallel logic
* Remove clamp statements
* Move test files to the proper place
* Update docs with new version of hf-doc-builder
* Fix test imports
* Make some minor improvements
* Add missing checkpoints to docs
* Make TGlobal model compatible with torch.onnx.export
* Replace some np.ndarray with jnp.ndarray
* Fix TGlobal for ONNX conversion + update docs
* fix _make_global_fixed_block_ids and masked neg value
* update flax model
* style and quality
* fix imports
* remove load_tf_weights_in_longt5 from init and fix copies
* add slow test for TGlobal model
* typo fix
* Drop obsolete is_parallelizable and one warning
* Update __init__ files to fix repo-consistency
* fix pipeline test
* Fix some device placements
* [wip]: Update tests -- need to generate summaries to update expected_summary
* Fix quality
* Update LongT5 model card
* Update (slow) summarization tests
* make style
* rename checkpoitns
* finish
* fix flax tests
Co-authored-by: phungvanduy <pvduy23@gmail.com>
Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>
Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com>
Co-authored-by: patil-suraj <surajp815@gmail.com>
* enable cpu distribution training using mpirun
*command like
* mpirun -n 2 python3 run_qa.py --no_cuda --xpu_backend ccl xxxx
*MASTER_ADDR and MASTER_PORT should be set as env
*export MASTER_ADDR=127.0.0.1
*export MASTER_PORT=29500
Signed-off-by: Wang, Yi A <yi.a.wang@intel.com>
* fix according to the review comment
Signed-off-by: Wang, Yi A <yi.a.wang@intel.com>
* use accelerate logic for cpu distribution training to set "RANK","LOCAL_RANK","WORLD_SIZE" environment
Signed-off-by: Wang, Yi A <yi.a.wang@intel.com>
* allow scope from trainer arg
* add ray_scope to training args
* escape double quotes
* make style && quality
* attempt to solve doc style issues
* splitting up URLs for style
* make fixup
* Update src/transformers/training_args.py
Co-authored-by: Antoni Baum <antoni.baum@protonmail.com>
* make style
Co-authored-by: Antoni Baum <antoni.baum@protonmail.com>
I'm guessing that the intention was to have the `_no_split_modules` class attribute for `GPTNeoXPreTrainedModel` to be set to `["GPTNeoXLayer"]`, akin to how its set as `["GPTJBlock"]` for `GPTJPreTrainedModel`.
If this is incorrect, please feel free to just close the PR.
Thanks!
* Raise RepoNotFoundError in case of 401
* Include changes from revert-17646-skip_repo_not_found
* Add a comment
* 💄 Code quality
* 💚 Update `get_from_cache` test
* 💚 Code quality & skip failing test
VisibleDeprecationWarning is addressed by specifying dtype=object when creating numpy array.
Update code based on review feedback.
Undo whitespace changes to tokenization_utils_base.py.
Co-authored-by: I like data <ilikedata@nym.hush.com>
When we're preparing the tensors for CPU for postprocessing, we need
to upgrade the `float16` to `float32` since CPUs don't have instructions
for `[b]float16`.
* Adding `top_k` and `sort` arguments to `text-classification` pipeline.
- Deprecate `return_all_scores` as `top_k` is more uniform with other
pipelines, and a superset of what `return_all_scores` can do.
BC is maintained though.
`return_all_scores=True` -> `top_k=None`
`return_all_scores=False` -> `top_k=1`
- Using `top_k` will imply sorting the results, but using no argument
will keep the results unsorted for backward compatibility.
* Remove `sort`.
* Fixing the test.
* Remove bad doc.
* Use shape_list to safely get shapes
* Add relevant test
* Tidy and add metrics
* Resolve dynamic shaping issues and move test
* Tidy up and all samples in batch
* Formatting
* adding template
* update model
* model update
* update conf for debug model
* update conversion
* update conversion script
* update conversion script
* fix missing keys check
* add tests to test the tokenizer in the local machine
* Change variable name
* add tests on xnli dataset
* add more description
* add descriptions + clearer code
* clearer code
* adding new tests + skipping few tests because of env problems
* change comment
* add dtype on the configuration
* add test embeddings
* add hardcoded test
* fix dtype issue
* adding torch.float16 to config
* adding more metrics (min, max, mean)
* add sum
* now the test passes with almost equal
* add files for conversion - test passes on cpu gpu
* add final changes
* cleaning code
* add new args in the docstring
* fix one liner function
* remove macros
* remove forward attention
* clean up init funtion
* add comments on the issue
* rm scale mask softmax
* do make style
* fix dtype in init
* fixing for loop on att probs
* fix style with black
* fix style + doc error
* fix and debug CI errors (docs + style)
* some updates
- change new operations
- finally add scaled softmax
- added new args in the config
* make use cache working
* add changes
- save sharded models
- final changes on the modeling script
* add changes
- comment on alibi
- add TODO on seq length
* test commit
- added a text to test the commit
Co-authored-by: thomasw21 <24695242+thomasw21@users.noreply.github.com>
* final changes
- attention mask change
- generation works on BS176b
Co-authored-by: thomasw21 <24695242+thomasw21@users.noreply.github.com>
* changes - model + conversion
* move to correct dir
* put ,
* fex fixes
* fix tokenizer autodoc
* fix minor CI issues
* fix minor CI issues
* fix minor CI issues
* fix style issue
* fix minor import issues
* fix few issues
* remove def main on the test
* add require torch
* replace decorator with 'with'
* fix style
* change to bloom
* add quick fix tokenizer
* fix tokenizer file
* fix tokenizer
- merge tests
- small fixes
* fix import issue
* add bloom to readme
* fix consistency
* Update docs/source/en/model_doc/bloom.mdx
Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>
* Apply suggestions from code review
fix comment issues on file headers
Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>
* fix doc issue
* small fix - modeling test
* some changes
- refactor some code
- taking into account reviews
- more tests should pass
- removed pruning tests
* remove useless division
* more tests should pass
* more tests should pass
* more tests should pass
* let's try this one
-add alibi offset
- remove all permutes to make the grad operations work
- finger crossed
* refactor
- refactor code
- style changes
- add new threshold for test
* major changes
- change BLOOM to Bloom
- add quick doc on bloom.mdx
- move embeddings test on modeling test
* modify readme
* small fixes
* small fix
- better threshold for a test
* remove old test file from fetcher
* fix small typo
* major change
- change BloomLMHead to BloomForCausalLM
* remove onnx config
* major changes
- refactor the code
- remove asserts
- change tol for test
* make style
* small change
* adding a slow test + commenting old ones for now
* make style
* Apply suggestions from code review
Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>
* make style
* fix duplicates
* cleaning comments on config
* clean a bit conversion file
* refacor a bit modeling file
* refactor tokenizer file
* fix tokenization test issue
* fix tokenization issue #2
* fix tokenization issue second try
* fix test issue
* make style + add suggestions
* change test fetcher
* try this one
- slow tests should pass
- finger crossed
* possible final changes
* make style
* try fix padding side issue
* fix side
* fix padding issue
* fix ko-readme
* fix config auto
* cleaning modeling file
* keep bloom in caps in ko
* update config docs
* remove pretraining_pp
* remove model parallel
* update config
- add correct config files
* fix duplicates
* fix fetcher
* fix refactor issue
- remove divide function
* try to remove alibi
* small fixes
- fix alibi
- remove seq length
- refactor a bit the code
* put correct values
- fix bos and eos token ids
* fix attention mask loop
Co-authored-by: thomasw21 <24695242+thomasw21@users.noreply.github.com>
* small fixes:
- remove skip bias add
* small fixes
- fix typo in readme
- fix typos in config
* small changes
- remove a test
- add reconstruction test
- change config
* small changes
- change Scaled Softmax to BloomScaledSoftmax
* small fixes
- fix alibi dtype
* major changes
- removing explicit dtype when loading modules
- fixing test args (torch_dtype=auto)
- add dosctring
* fix readmes
* major changes
- now bloom supports alibi shifting
- refactor a bit the code
- better test tolerance now
* refactor a bit
* refactor a bit
* put correct name on test
* change docstring
* small changes
- fix docstring modeling
- fix test tolerance
* fix small nit
- take dtype from tensors in the conversion script
* minor fix
- fix mdx issue
* minor fix
- change config docstring
* forward contrib credits from PR14084
* Apply suggestions from code review
Co-authored-by: Stas Bekman <stas00@users.noreply.github.com>
* apply modifications
Co-authored-by: Stas Bekman <stas00@users.noreply.github.com>
* resolve softmax upcast
* Apply suggestions from code review
Co-authored-by: Stas Bekman <stas00@users.noreply.github.com>
* Update src/transformers/models/bloom/modeling_bloom.py
Co-authored-by: Niklas Muennighoff <n.muennighoff@gmail.com>
* final changes modeling
Co-authored-by: Stas Bekman <stas00@users.noreply.github.com>
* Merge commit 'd156898f3b9b2c990e5963f5030a7143d57921a2'
* merge commit
* Apply suggestions from code review
Co-authored-by: Stas Bekman <stas00@users.noreply.github.com>
* apply suggestions
Apply suggestions from Stas comments
Co-authored-by: Stas Bekman <stas00@users.noreply.github.com>
* Fix gradient checkpointing
Co-authored-by: Stas Bekman <stas00@users.noreply.github.com>
* add slow but exact
* add accelerate compatibility
Co-authored-by: Nicolas Patry <Narsil@users.noreply.github.com>
* forward contrib credits
Co-authored-by: thomasw21 <thomasw21@users.noreply.github.com>
Co-authored-by: sgugger <sgugger@users.noreply.github.com>
Co-authored-by: patrickvonplaten <patrickvonplaten@users.noreply.github.com>
Co-authored-by: Niklas Muennighoff <n.muennighoff@gmail.com>
Co-authored-by: LysandreJik <LysandreJik@users.noreply.github.com>
* Apply suggestions from code review
Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com>
* fix torch device on tests
* make style
* Apply suggestions from code review
Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com>
* fix nits
Co-authored-by: patrickvonplaten<patrickvonplaten@users.noreply.github.com>
* remove final nits
* fix doc
- add more details on the doc
- add links to checkpoints
* Update src/transformers/__init__.py
Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>
* Update src/transformers/models/bloom/modeling_bloom.py
Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>
* apply suggestions
Co-authored-by: sgugger <sgugger@users.noreply.github.com>
* put test torchscript to false
* Update src/transformers/models/bloom/modeling_bloom.py
Co-authored-by: justheuristic <justheuristic@gmail.com>
* fix alibi
- create alibi only once
* add small doc
* make quality
* replace torch.nn
* remove token type emb
* fix fused op + output bias
* add fused op
- now can control fused operation from config
* remove fused op
* make quality
* small changes
- remove unsed args on config
- removed bias gelu file
- make the model torchscriptable
- add torchscript slow tests
* Update src/transformers/models/bloom/modeling_bloom.py
* fix slow
* make style
* add accelerate support
* add bloom to deepspeed tests
* minor changes
* Apply suggestions from code review
Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com>
* minor change
* slow tests pass
* Apply suggestions from code review
Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>
* Update docs/source/en/model_doc/bloom.mdx
Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>
* minor changes:
- change docstring
- add link to paper
Co-authored-by: Thomwolf <thomwolf@gmail.com>
Co-authored-by: Thomas Wolf <thomas@huggingface.co>
Co-authored-by: thomasw21 <24695242+thomasw21@users.noreply.github.com>
Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>
Co-authored-by: sIncerass <sheng.s@berkeley.edu>
Co-authored-by: Stas Bekman <stas00@users.noreply.github.com>
Co-authored-by: Niklas Muennighoff <n.muennighoff@gmail.com>
Co-authored-by: Nicolas Patry <Narsil@users.noreply.github.com>
Co-authored-by: thomasw21 <thomasw21@users.noreply.github.com>
Co-authored-by: sgugger <sgugger@users.noreply.github.com>
Co-authored-by: patrickvonplaten <patrickvonplaten@users.noreply.github.com>
Co-authored-by: LysandreJik <LysandreJik@users.noreply.github.com>
Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com>
Co-authored-by: justheuristic <justheuristic@gmail.com>
Co-authored-by: Stas Bekman <stas@stason.org>
* Fix link for community notebooks
This fixes the link for community notebooks due to reorganization.
* Replace old link with fully link to the doc page
Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>
Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>
* feat: initial implementation of data2vec segmentation model in TF.
* chore: minor corrections to make the segmenter work.
* chore: removed unncessary files.
* chore: add tests and other modifications.
* fix: loss computation for segmentation.
* chore: remove unused variable.
* chore: formatting.
* added a dummy adaptive pooling layer.
* removed unnecessary file.
* potentially add identifiers to layer names.
* fix: layer naming.
* chore: removed unnecessary print.
* Skipping unneeded test
* chore: add logging to debug tolerance.
* fix: segmentation tests for tfdata2vecvision
* chore: make style.
* fix: layer names, assertion to be resolved.
* Bumping test tolerance a bit
* chore: bump the tol in PT test.
Co-authored-by: matt <rocketknight1@gmail.com>
