trainfer

Train and infer at the same time. A single-process FastAPI daemon that shares one set of weights between inference and training.

One mutable model. Always serving. Always trainable. Any objective, any time, via API.

Feedback you POST lands on the next inference request under a typed contract, not a best-effort: /v1/train returns a commit_token, and /v1/chat/completions accepts after_commit_token to block until that training step is reflected in the forward pass. That contract is what makes continual learning something you can build on rather than hope for.

trainfer is the serving substrate. The continual-learning research that runs on top of it — teachers, RLVR/RLAIF loops, benchmark runners, the eval harness, campaign journals — lives in cont and drives this daemon over HTTP.

Where to look first

You want to…	Read
Get the daemon running	OPERATING.md — env vars, ports, data layout, smoke test
Understand what trainfer does and why	trainfer/PLAN.md (north-star spec) + trainfer/DESIGN.md (one-pager)
See what's actually shipped	trainfer/STATUS.md — every claim is cited by a test
Learn the in-house vocabulary	trainfer/GLOSSARY.md
Move off an older lile / agi checkout	MIGRATION.md
Touch the daemon HTTP surface	trainfer/README.md

Layout

trainfer/                     # the daemon — server, training engine, objectives, queue, state
trainfer/console/             # web console (launch.py, dashboard.html, metrics.html, …)
trainfer/ttrl_mv.py           # idle-time TTRL majority-vote pseudo-label scheduler
trainfer/objectives/verifiers # verifier registry + in-tree corpora + plugin entry point
trainfer/tests/               # pytest suite (cpu_only + gpu markers)
compose.trainfer-dev.yaml     # dev-mode docker compose (daemon + studio)
.claude/skills/trainfer/      # Claude Code skill for working in this repo
pyproject.toml                # trainfer package + cross-repo unsloth dep

Cross-repo dependencies

trainfer consumes unsloth (the heiervang fork) as a pinned git dependency. The single load-bearing coupling is a runtime monkeypatch of unsloth.kernels.utils.matmul_lora in trainfer/state.py; TrainferMatmulRebindError guards against upstream signature drift.

The HTTP-side companion (Studio frontend + studio/backend/routes/lile.py proxy) lives in ht-unsloth and talks to the daemon over LILE_DAEMON_URL. That side has not been renamed yet — see MIGRATION.md for the pending follow-up.

cont depends on trainfer, never the other way round. Benchmark-shaped verifiers that carry their own corpora register into the daemon through the trainfer.verifiers entry-point group rather than being imported by it.

Quickstart

pip install -e .
python -m trainfer.console.launch    # default: Qwen3-8B on :8768

# OpenAI-compatible chat
curl -sS http://127.0.0.1:8768/v1/chat/completions \
  -H 'content-type: application/json' \
  -d '{"messages":[{"role":"user","content":"hi"}],"max_tokens":64}' | jq

# Train in the same process — returns commit_token N
curl -sS http://127.0.0.1:8768/v1/train \
  -H 'content-type: application/json' \
  -d '{"objective":"sft","samples":[{"prompt":"2+2?","response":"4."}]}' | jq

# Next chat that MUST see batch N
curl -sS http://127.0.0.1:8768/v1/chat/completions \
  -H 'content-type: application/json' \
  -d '{"messages":[{"role":"user","content":"2+2?"}],"after_commit_token":N}' | jq

Tests

pytest -m cpu_only           # no GPU required
pytest                       # full suite (GPU needed)