feat: TableFormer v2

[PeterStaar-IBM]

No description provided.

[github-actions]

❌ DCO Check Failed

Hi @PeterStaar-IBM, your pull request has failed the Developer Certificate of Origin (DCO) check.

This repository supports remediation commits, so you can fix this without rewriting history — but you must follow the required message format.

---

🛠 Quick Fix: Add a remediation commit

Run this command:

git commit --allow-empty -s -m "DCO Remediation Commit for Ahmed Nassar AHN@zurich.ibm.com <AHN@zurich.ibm.com>

I, Ahmed Nassar AHN@zurich.ibm.com <AHN@zurich.ibm.com>, hereby add my Signed-off-by to this commit: 58494d4f07d97006f60cc24e288c29dac4079436
I, Ahmed Nassar AHN@zurich.ibm.com <AHN@zurich.ibm.com>, hereby add my Signed-off-by to this commit: 8753ab943d6269bf4a4f843defa62518453bc391
I, Christoph Auer <cau@zurich.ibm.com>, hereby add my Signed-off-by to this commit: 107142a955c336b398cc3d47c6283b29b156793d"
git push

---

🔧 Advanced: Sign off each commit directly

For the latest commit:

git commit --amend --signoff
git push --force-with-lease

For multiple commits:

git rebase --signoff origin/main
git push --force-with-lease

More info: DCO check report

[mergify]

Merge Protections

Your pull request matches the following merge protections and will not be merged until they are valid.

🟢 Enforce conventional commit

Wonderful, this rule succeeded.

Make sure that we follow https://www.conventionalcommits.org/en/v1.0.0/

• title ~= ^(fix|feat|docs|style|refactor|perf|test|build|ci|chore|revert)(?:\(.+\))?(!)?:

🟢 Require two reviewer for test updates

Wonderful, this rule succeeded.

When test data is updated, we require two reviewers

• #approved-reviews-by >= 2

[codecov]

Codecov Report

❌ Patch coverage is 85.98131% with 45 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
docling_ibm_models/tableformer_v2/model.py	85.89%	45 Missing ⚠️

📢 Thoughts on this report? Let us know!

[ameykpatel]

Hi @PeterStaar-IBM,
Thanks for the TableFormerV2 work in #149 - the architecture improvements look great (EfficientNetV2-S backbone, depthwise separable convolutions replacing the transformer encoder, and proper batched decoding). The batch inference support is particularly exciting as single-image decoding in V1 has been a throughput bottleneck for us.

A few questions on readiness:

1. Weights availability: The test file references docling-project/TableFormerV2 (v0.1.0) on HuggingFace, but I wasn't able to find this repo, are the weights published yet, or is this still pending?
2. Predictor pipeline: V1's TFPredictor handles the full pipeline whereas V2's model.generate() currently returns raw token IDs and bbox tensors. Is the plan to build a dedicated V2 predictor, or extend TFPredictor (which currently hardcodes TableModel04_rs at instantiation)?
3. Docling integration: TableStructureModel is wired to V1 at multiple levels. Are there plans to add a V2 mode, or will this be a separate model class?

Is there a roadmap for when V2 will be usable through the standard docling pipeline? Happy to contribute if there are pieces we can help with.

[cau-git]

@ameykpatel the hugging face weights should be public now. We are working on docling Integration here: docling-project/docling#3013

[ameykpatel]

That's great! I look forward to the docling integration