GitLab

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

* fix

* fix

* update

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

Logging an error and continuing is probably following the principle of least surprise.

It's more idiomatic and significantly more efficient because
1) it avoids repeated `append` call that Python has to resolve on each iteration
2) can preallocate the size of the final list avoiding resizing

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

* Add a test of the built release

* Polish everything

* Trigger CI

* [automodel] fix torch_dtype=auto in from_pretrained

* add test

* fix logic

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

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

* Fix translation

add conditional generation on docstring

* Few fixes in new Tools implementation

* code quality

* docs: ko: task/asr.mdx

* feat: manual draft

* fix: resolve suggestions
Co-authored-by: Hyeonseo Yun <0525yhs@gmail.com>

---------
Co-authored-by: Hyeonseo Yun <0525yhs@gmail.com>

fix chat prompts

* fix

* fix

* fix test + add more docs

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

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

* Use `mkstemp` to replace deprecated `mktemp`

The `tempfile.mktemp` function is [deprecated](https://docs.python.org/3/library/tempfile.html#tempfile.mktemp) due to [security issues](https://cwe.mitre.org/data/definitions/377.html

).

* Update src/transformers/utils/hub.py
Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>

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

* Replace numpy operations with jax.numpy for JIT compatibility

Replaced numpy operations with their jax.numpy equivalents in the transformer library. This change was necessary to prevent errors during JIT compilation. Specifically, the modifications involve changing numpy's in-place assignments to jax.numpy's immutable update methods.

* rm numpy import

* rm numpy import and fix np->jnp

* fixed slices bug

* fixed decoder_start_tokens -> decoder_start_token_id

* fixed jnp in modleing mt5

* doc fix

* rm numpy import

* make

* Added type hints for `Graphormer` pytorch version

added type hints for graphormers pytorch , checked formating issues .

* made the code less bloated

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

[Bugfix] `OPTDecoderLayer` does not return attentions when `gradient_checkpointing` and `training` is enabled. (#23367)

Update modeling_opt.py

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

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

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

* fix

* fix

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

Update graphormer.mdx

Typo suggestion

* fix

* fix

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

Removing twice defined position_embeddings

The self.position_embeddings in LongFormerEmbeddings is defined twice.
Removing the first with padding_idx

* fix

* fix

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

replaced assert with raise ValueError for t5, switch_transformers, pix2struct, mt5, longt5, gptsan_japanese. (#23273)

* replaced assert with raise ValueError

* one liner

* reverse one liner and cache-decoder check

* Handle padding warning in generation when using `inputs_embeds`

* Simpler condition

* Black formatter

* Changed warning logic

or or and

fix kwargs

* min. version for pytest

* fix

* fix

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

When working on TorchInductor, I realised that there was a part from
`XLNetLMHeadModel` that was being compiled to CPU code.

This PR should allow to fuse this operation with other CUDA operations
in `torch.compile`. It also should be faster on eager mode, as it has a
this implementation has a lower foot-print.

If in-place operations are not allowed even in non-grad context, I still
believe that doing ones + tril rather than a ones + tril + zeros + cat
should be faster simply due to the number of memory reads/writes.

I tested that this code produces the same results for `0 <= qlen,mlen <
10` and `same_length in (True, False)`.

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