Configuration

A server reads its settings from environment variables once, at start. Flags take 1/true or 0/false.

Models and devices

Variable	Default
TABCTX_BACKEND	tabicl	Comma-separated models to serve: tabicl, tabpfn, tabdpt, fake (no GPU or weights; for tests). The first is the default model. tabicl,tabdpt serves both from one cache.
TABCTX_DEVICES	auto	auto: every visible GPU, else the CPU. Or a list such as cuda:0,cuda:1. Two or more devices place each table on one GPU.
TABCTX_WORKER_PROCESSES	1	With several GPUs, one worker process per GPU (0: threads of one process).
TABCTX_GPU_MEMORY_FRACTION	1.0	The share of each GPU this server may use, in (0, 1]. Lower it when replicas share a GPU.
TABPFN_TOKEN		Needed to download TabPFN weights, after accepting Prior Labs' terms.

Cache and memory

Variable	Default
TABCTX_SPILL_DIR	off	A directory for the disk tier: evicted contexts are written here and restored on their next predict. Measure on your disk first; a re-fit can be faster than a restore from a slow one.
TABCTX_SPILL_CAPACITY_BYTES	50 GiB	The disk tier's size, shared across GPUs.
TABCTX_KV_CACHE	kv	TabICL's context: kv (fastest predicts), repr (much less memory per context), off (re-encode on every predict).
TABCTX_KV_PRECISION	int8	TabICL's cached keys and values: int8 (about half the memory; identical labels and metrics on 7 datasets, probabilities within 0.01), fp16, native.
TABCTX_OFFLOAD_DIR	<spill dir>/offload	Where TabICL may stage a very large fit's intermediate results on disk.

Retention

Variable	Default
TABCTX_MAX_CONTEXT_AGE_S	no limit	No fitted table is kept longer than this after its fit: it is deleted from device memory, host memory and the disk tier, within 30 s. It applies to every table, including those fit before it was set (a table spilled by tabctx < 0.19 has no recorded fit time and counts as expired). A fit can ask for less with ttl_s; GET /v1/tabctx/limits reports it.

TabDPT

Variable	Default
TABCTX_TABDPT_CONTEXT_SIZE	2048	Training rows sampled as each predict's context. TabDPT keeps no encoding of the table, so a predict's cost grows with this, not with the table.
TABCTX_TABDPT_ENSEMBLES	4	Ensemble members per predict.

The TABCTX_KV_* and TABCTX_COMPILE* settings apply to TabICL only.

Requests

Variable	Default
TABCTX_BATCH_WINDOW_MS	5	Concurrent predicts on one context within this window share one model call. 0 turns it off.
TABCTX_MAX_UPLOAD_BYTES	4 GiB	The largest table file a request may carry.
TABCTX_UPLOAD_TTL_S	3600	How long an unused /v1/tabctx/upload is kept.

Performance

Variable	Default
TABCTX_COMPILE	1	torch.compile TabICL's attention blocks at start (slower start, faster calls).
TABCTX_COMPILE_AFTER_WARMUP	0	Also compile shapes first seen in requests, inside those requests.
TABCTX_FAST_PREPROCESSING	1	Fit TabICL's Yeo-Johnson preprocessing on the GPU.

Tenancy and routing

Variable	Default
TABCTX_REQUIRE_TENANT	false	true, 1 or yes: every request that reads or changes contexts must carry x-tabctx-tenant-id (else 401). /v1/tabctx/limits, /v1/models and the probes don't need it.
RAY_SERVE_SESSION_ID_HEADER_KEY	x-session-id	The replica-affinity header, shared with Ray Serve's router.

The tenant header is trusted as sent

tabctx does not authenticate callers: any caller can send any tenant id. Tenancy is a security boundary only behind a gateway that authenticates each caller, sets x-tabctx-tenant-id itself and strips any value the client sent. See tenant isolation.

Standalone host

python -m tabctx.serve [--host 0.0.0.0] [--port 8000]

Variable	Default
TABCTX_MAX_ONGOING_REQUESTS	8	Requests served at once.
TABCTX_MAX_QUEUED_REQUESTS	16	Requests waiting for a slot; beyond that, 503. Health probes are never shed.

Clients

Variable
TABCTX_BASE_URL	Where TabctxClient() and the estimators send requests.
TABCTX_API_KEY	Sent as Authorization: Bearer (a gateway's key).
TABCTX_MCP_HOME	Where tabctx-mcp keeps fitted models (default .tabctx-mcp/).
TABCTX_MCP_KEY_FILE	The key tabctx-mcp signs saved models with (default ~/.config/tabctx-mcp/signing.key).