Serve more open tabular foundation models, so tabctx is the vendor-neutral server for them. Today it serves `tabicl-v2` and `tabpfn-3.5`. ## Candidates - **Mitra** (AutoGluon, Amazon) - **TabDPT** (Layer 6) ## Per model - A backend class next to `backends/tabicl.py` and `backends/tabpfn.py`: fit to a cached context, predict against it. - Memory calibration for admission, as for TabICL and TabPFN. - Its own id in `GET /v1/models`, plus a row in the README's models table with the weights license. - Accuracy on the real-dataset suite (`benchmarks/eval_real_datasets.py`) matches the model's own package. ## Accept - Fit and predict through the server give the same predictions as the model's own package, within tolerance. - A fit that would run the GPU out of memory is refused before it runs.
Activity
TabDPT: a third model, with Apache-2.0 code and weights (#102) (#110)