[Q] How to turn off only model synching in huggingface integration

[jubueche]

Hi.
I am training large models and they are logged to wandb. This happens through the artifacts. How do I only turn off this feature?
I tried googling but couldn’t find an answer.

[ArtsiomWB]

Hi @jubueche, could you please talk a bit about your workflow, and why you are interested in turning off the artifact logging ?

[jubueche]

Hi,

I don't need to turn off the artifact logging. I just don't want my model to get synced to wandb. My models are multiple GB and it takes quite some space and time when I upload them.

[ArtsiomWB]

Gotcha, could you please try setting os.environ["WANDB_LOG_MODEL"] = "false"

Here are our docs on it

[umarbutler]

Gotcha, could you please try setting os.environ["WANDB_LOG_MODEL"] = "false"

Here are our docs on it

This does not work.

[jubueche]

@ArtsiomWB I can confirm. This does not work.

[jubueche]

# # log the initial model and architecture to an artifact
# with tempfile.TemporaryDirectory() as temp_dir:
#     model_name = (
#         f"model-{self._wandb.run.id}"
#         if (args.run_name is None or args.run_name == args.output_dir)
#         else f"model-{self._wandb.run.name}"
#     )
#     model_artifact = self._wandb.Artifact(
#         name=model_name,
#         type="model",
#         metadata={
#             "model_config": model.config.to_dict() if hasattr(model, "config") else None,
#             "num_parameters": self._wandb.config.get("model/num_parameters"),
#             "initial_model": True,
#         },
#     )
#     model.save_pretrained(temp_dir)
#     # add the architecture to a separate text file
#     save_model_architecture_to_file(model, temp_dir)

#     for f in Path(temp_dir).glob("*"):
#         if f.is_file():
#             with model_artifact.new_file(f.name, mode="wb") as fa:
#                 fa.write(f.read_bytes())
#     self._wandb.run.log_artifact(model_artifact, aliases=["base_model"])

#     badge_markdown = (
#         f'[<img src="https://raw.githubusercontent.com/wandb/assets/main/wandb-github-badge'
#         f'-28.svg" alt="Visualize in Weights & Biases" width="20'
#         f'0" height="32"/>]({self._wandb.run.get_url()})'
#     )

#     modelcard.AUTOGENERATED_TRAINER_COMMENT += f"\n{badge_markdown}"

I just commented out the following in the integration_utils.py of transformers of Hugginggface.

[ArtsiomWB]

Hey @jubueche, thank you so much for the workaround. It is strange that os.environ["WANDB_LOG_MODEL"] = "false" is not working on your side. What version of wandb are you currently on? Will try reproducing this on my side.

[ArtsiomWB]

Hi there, I wanted to follow up on this request. Please let us know if we can be of further assistance or if your issue has been resolved.

[jubueche]

Hi sorry, my wandb version is

>>> wandb.__version__
'0.16.4'

For now I am just using the code with the commented out section.

[umarbutler]

# # log the initial model and architecture to an artifact
# with tempfile.TemporaryDirectory() as temp_dir:
#     model_name = (
#         f"model-{self._wandb.run.id}"
#         if (args.run_name is None or args.run_name == args.output_dir)
#         else f"model-{self._wandb.run.name}"
#     )
#     model_artifact = self._wandb.Artifact(
#         name=model_name,
#         type="model",
#         metadata={
#             "model_config": model.config.to_dict() if hasattr(model, "config") else None,
#             "num_parameters": self._wandb.config.get("model/num_parameters"),
#             "initial_model": True,
#         },
#     )
#     model.save_pretrained(temp_dir)
#     # add the architecture to a separate text file
#     save_model_architecture_to_file(model, temp_dir)

#     for f in Path(temp_dir).glob("*"):
#         if f.is_file():
#             with model_artifact.new_file(f.name, mode="wb") as fa:
#                 fa.write(f.read_bytes())
#     self._wandb.run.log_artifact(model_artifact, aliases=["base_model"])

#     badge_markdown = (
#         f'[<img src="https://raw.githubusercontent.com/wandb/assets/main/wandb-github-badge'
#         f'-28.svg" alt="Visualize in Weights & Biases" width="20'
#         f'0" height="32"/>]({self._wandb.run.get_url()})'
#     )

#     modelcard.AUTOGENERATED_TRAINER_COMMENT += f"\n{badge_markdown}"

I just commented out the following in the integration_utils.py of transformers of Hugginggface.

Do you know if this was a new addition to transformers? Maybe the problem is on their side? This problem doesn't arise for me on a different system which has an older version of transformers.

[ArtsiomWB]

Thank you so much for the follow up, @umarbutler , hey @jubueche , are you able to try a different version of transformets and see if that fixes it?

@umarbutler, what version currently works for you?