TabPFN memory calibration grid for the admission estimator

[amrit110]

Why

The factory-preloaded admission grid (memory/calibration_tabicl_a100.py) is TabICLv2 data measured on an A100-40GB. TabPFNBackend deployments therefore start from the conservative power-law formula plus whatever runtime observations accumulate (AdaptiveMemoryEstimator), so early TabPFN admissions are much tighter than the hardware allows and /v1/tabctx/limits under-reports what TabPFN can take. backends/tabpfn.py says as much in context_bytes_hint.

Since v0.10.0 the serving budget is the detected device capacity, so the grid only needs to describe the model's peak memory per shape; the card size is handled separately.

Tasks

• Run benchmarks/calibrate_memory.py-style sweeps through TabPFNBackend on an A100 for the three cache modes (kv -> fit_with_cache, repr -> fit_preprocessors, off -> low_memory), recording peak fit bytes, resident bytes, predict-time peak at N_TEST_ROWS, and the first OOM per feature count. TabPFN's pretraining limits (~10k rows / ~500 features on the v2-series checkpoints, larger on TabPFN-3) bound the grid; record where the model rejects (422) vs where the device OOMs.
• Extend benchmarks/gen_calibration_data.py to emit A100_40GB_TABPFN_* grids and teach serve/factory.py::_preloaded_observations to select the grid by backend kind (today it returns empty for anything but tabicl).
• Unit test: build_estimator(ServeSettings(backends=("tabpfn",))) reports preloaded measurements in confidence() and estimates a grid shape from measurement, not formula (mirror TestCalibrationPreload).
• Re-run benchmarks/probe_deployment.py on a TabPFN deployment and confirm max_admissible_train_rows_by_feature_count loosened accordingly.

Pointers

memory/adaptive.py (preloaded observations, margins), memory/calibration_tabicl_a100.py (generated file format), ROADMAP.md item 3.