PyPI release (1.0.0) is incompatible with the published Hugging Face checkpoint(for kaggle); loader fix exists on main but has not been released

[Samarth-ITM]

The current PyPI release (tabfm==1.0.0) cannot load the official Hugging Face PyTorch checkpoint (google/tabfm-1.0.0-pytorch).

After investigating the issue, it appears that this is a release mismatch rather than an architecture or checkpoint problem. The current GitHub source already contains a rewritten loader (TabFM_HF.from_pretrained() via PyTorchModelHubMixin), but PyPI still distributes the older loader implementation.

The current PyPI package (tabfm==1.0.0) cannot load the official
google/tabfm-1.0.0-pytorch checkpoint.

The published Hugging Face repository contains

classification/
model.safetensors

while the released PyPI package expects

classification/pytorch_model.bin

and calls torch.load().

This causes

FileNotFoundError:
Weights not found at ...

I noticed the current repository already contains a rewritten loader
based on PyTorchModelHubMixin (version reports 1.0.1 in the source tree),
which correctly delegates loading through Hugging Face.

Therefore, this appears to be a packaging/release mismatch rather than
an implementation bug.

The GitHub source contains the fix, but the published package used by
Kaggle and pip users still contains the older loader.

Would it be possible to publish a new PyPI release matching the current
repository so users can load the official checkpoint without patching
the loader manually?

[Samarth-ITM]

Environment

Python 3.12
tabfm==1.0.0
huggingface_hub
torch

Reproduction

from tabfm import tabfm_v1_0_0_pytorch as tabfm

model = tabfm.load(
model_type="classification"
)

Result

FileNotFoundError:

Weights not found at

classification/pytorch_model.bin

Investigation

The official Hugging Face repository contains

classification/
model.safetensors

There is no

classification/pytorch_model.bin

The released PyPI loader manually searches for

classification/pytorch_model.bin

and calls torch.load().

I manually loaded the published checkpoint using

safetensors.torch.load_file()

and verified

Missing Keys: []
Unexpected Keys: []

which confirms the published checkpoint exactly matches the released model
architecture.

Additional observation

The current GitHub source already contains a rewritten loader based on

TabFM_HF.from_pretrained()

using Hugging Face's PyTorchModelHubMixin.

My local checkout reports 'Version: 1.0.1' while pip index versions tabfm shows 'Available versions: 1.0.0'