Automatic safetensors conversion when lacking these files #29390
Conversation
src/transformers/modeling_utils.py
Outdated
| cls._auto_conversion = Thread( | ||
| target=auto_conversion, | ||
| args=(pretrained_model_name_or_path,), | ||
| kwargs=cached_file_kwargs, | ||
| ) | ||
| cls._auto_conversion.start() |
There was a problem hiding this comment.
@Wauplin curious if you have a better idea in mind to have access to the thread started here; I don't need to join it during runtime, I'm only attributing it to the class here so that I can access it within the test files (but not super keen on modifying internals just for the tests to be simpler ...)
There was a problem hiding this comment.
@LysandreJik I'm not shocked by having a cls._auto_conversion attribute TBH. Though a solution to get rid of it is to give a name to the thread. Something like that:
Thread(
target=auto_conversion,
args=(pretrained_model_name_or_path,),
kwargs=cached_file_kwargs,
name="Thread-autoconversion-{<unique id here>}",
).start()and then in the tests:
for thread in threading.enumerate():
print(thread.name)
# ...
# Thread-autoconversion-0Thread names don't have to be unique BTW (they have a thread id anyway). But I think it's best to at least assign a unique number to the name.
There was a problem hiding this comment.
But it's quite hacky IMO. In a simple case it should work fine but if you start to have several threads / parallel tests, it might get harder to be 100% sure the thread you've started is indeed the one you retrieve in the test logic.
There was a problem hiding this comment.
yeah here it's really only for testing and I don't want to depend on a flaky time.sleep or something so ensuring that the thread joins first is optimal. The thread name is actually much better IMO, I'll implement that! Thanks a lot!
|
The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update. |
| # message. | ||
|
|
||
| if resolved_archive_file is not None: | ||
| if filename in [WEIGHTS_NAME, WEIGHTS_INDEX_NAME]: |
There was a problem hiding this comment.
one thing I would be wary is just that if we convert a big checkpoint from torch to safetensors and we want to load it in Flax, sharded safetensors are not supported yet
There was a problem hiding this comment.
Flax defaults to loading flax checkpoints, not safetensors, so it won't be affected by a repo where there is sharded safetensors
| @@ -1428,7 +1429,7 @@ def test_safetensors_on_the_fly_wrong_user_opened_pr(self): | |||
| bot_opened_pr_title = None | |||
|
|
|||
| for discussion in discussions: | |||
| if discussion.author == "SFconvertBot": | |||
| if discussion.author == "SFconvertbot": | |||
There was a problem hiding this comment.
➕ on @julien-c's comment, have had feedback that this is not explicit enough.
| if discussion.author == "SFconvertbot": | |
| if discussion.author == "HuggingFaceOfficialSafetensorConverter": |
bot is scary for some 😅
There was a problem hiding this comment.
we can't change the account name now
but we will think of a way to make it clearer in the UI that it's a "official bot"
|
Thanks both for the review! |
|
I'll merge it now and will keep monitoring issues to ensure it doesn't break things in the wild. |
…e#29390) * Automatic safetensors conversion when lacking these files * Remove debug * Thread name * Typo * Ensure that raises do not affect the main thread
…ggingface#2… (huggingface#29507) Revert "Automatic safetensors conversion when lacking these files (huggingface#29390)" This reverts commit a69cbf4.
* Automatic safetensors conversion when lacking these files * Remove debug * Thread name * Typo * Ensure that raises do not affect the main thread
* Automatic safetensors conversion when lacking these files * Remove debug * Thread name * Typo * Ensure that raises do not affect the main thread
When a user calls the PyTorch
from_pretrainedon a repository that only contains PyTorch/Flax/TF files, start an auto conversion in the background so that it has a PR opened withsafetensorsfiles.