Issues

• Concurrent large fits are slower on 4 GPUs than on 1 (host contention?)

  VectorInstitute/tabctx#41 · Filter by author amrit110 opened on Oct 18, 2026

• Serving architecture for tabctx: placement, process model, fit/predict disaggregation (decide by measurement)

  enhancementroadmap
  VectorInstitute/tabctx#39 · Filter by author amrit110 opened on Oct 17, 2026
  4

• Per-device TabICL calibration grids (L4, A40) for admission

  enhancementroadmap
  VectorInstitute/tabctx#37 · Filter by author amrit110 opened on Oct 17, 2026

• Latency: CUDA graphs for small predicts

  enhancementroadmap
  VectorInstitute/tabctx#36 · Filter by author amrit110 opened on Oct 17, 2026

• Tables larger than one GPU: context parallelism across a replica's GPUs

  enhancementroadmap
  VectorInstitute/tabctx#35 · Filter by author amrit110 opened on Oct 17, 2026

• Capacity: INT8 context KV cache with in-kernel dequantization

  enhancementroadmap
  VectorInstitute/tabctx#34 · Filter by author amrit110 opened on Oct 17, 2026
  1

• Throughput: paged context KV + cross-context varlen predict kernel

  enhancementroadmap
  VectorInstitute/tabctx#33 · Filter by author amrit110 opened on Oct 17, 2026

• Large tables: tiled/fused column+row stage kernels to cut the fit peak

  enhancementroadmap
  VectorInstitute/tabctx#32 · Filter by author amrit110 opened on Oct 17, 2026

• GPU profiling harness: per-stage time and peak-memory breakdown (GKE first, then A40)

  enhancementroadmap
  VectorInstitute/tabctx#31 · Filter by author amrit110 opened on Oct 17, 2026

• Docs site: split the README into a navigable site with an API reference

  documentationroadmap
  VectorInstitute/tabctx#9 · Filter by author amrit110 opened on Sep 1, 2026

• TabPFN memory calibration grid for the admission estimator

  enhancementroadmap
  VectorInstitute/tabctx#8 · Filter by author amrit110 opened on Sep 1, 2026
  1

• Post-release A100 stress run + researcher deployment on inference-platform

  roadmap
  VectorInstitute/tabctx#7 · Filter by author amrit110 opened on Sep 1, 2026

• Application catalog + persistent context store (serve many applications from pre-computed contexts)

  enhancementroadmap
  VectorInstitute/tabctx#6 · Filter by author amrit110 opened on Sep 1, 2026