Thank you and Ali for doing this great work!

Chen Li

On Tue, Oct 6, 2026 at 2:22 PM Tanishq Prashant Gandhi <[email protected]>
wrote:

> Hi all,
>
> ML model support tracked under 6494
> <https://github.com/apache/texera/issues/6494> is now on main! A model is
> a
> resource of its own, and a Python UDF can read one at runtime without
> downloading it. Big thanks to Ali, who built the mounting half and reviewed
> most of this work, to Meng for reviewing, and to Professor Chen for his
> advising!
>
> Three parts landed together, because running a model needs all three:
>
>    - Models. Create one under Your Work -> Models, upload files or whole
> folders, and commit them as an immutable version. Share it with people,
> publish it to the Hub, and find it through unified search.
>
>    - Mounting (6606 <https://github.com/apache/texera/issues/6606>). A UDF
> declares a parameter with value=Resource.MODEL; the property panel then
> offers a model browser instead of a text box, and at runtime the UDF
> receives a local directory. Nothing is downloaded - file contents travel
> only when the code reads them, so loading one shard of a large model
> transfers one shard.
>
>    - Curated computing-unit images (8466). An administrator registers a
> published image, Texera checks it is a computing-unit image and pins its
> digest, and it appears as a dropdown when a user creates a computing unit.
> This is what makes a model whose runtime the stock image does not carry -
> xgboost, or something like AlphaFold 3 - actually runnable.
>
> Models are off by default behind gui.tabs.models_enabled and
> gui.tabs.model_enabled, both toggleable in Admin -> Settings.
>
> Ali and I wrote up the design and what the three parts look like in use on
> the blog, go here if you want to learn more!:
> https://texera.apache.org/blog/2026/10/04/ml-model-support/
>
> Feedback is welcome. Next up is agreeing on a standard layout for a model
> so Texera can load one without being told how, and a no-code inference
> operator on top of that.
>
> Thanks,
> Tanishq
>

Reply via email to