ARTI

ARTI = AI x RT.

ARTI is a programmable neural runtime for PyTorch and JAX. Compose neural programs with Formula Fabric, explicit tensor resources, learned connections, local iteration and forward plasticity. Adapt an existing model or build a standalone model, then train, resume, evaluate and deploy the same asset.

The PyPI distribution is arti-fit; the Python import is arti.

Install

pip install "arti-fit==3.3.0"
pip install "arti-fit[qwen,serve]==3.3.0"  # language recipes and optional ASGI serving
pip install "arti-fit[jax,flax]==3.3.0"    # native JAX and optional Flax integration

Python 3.10+ is supported. PyTorch is a core dependency; model, JAX, Flax, documentation and serving integrations are optional extras.

For an unpublished candidate, install its reviewed local wheel with python -m pip install --no-index --no-deps /path/to/arti_fit-3.3.0-py3-none-any.whl in an environment whose declared dependencies are already provisioned. A candidate version string does not imply availability on PyPI.

Start An Application Task

arti task init my-task --name "my task"
cd my-task
arti task check project.json
arti task explain project.json
python workflow.py prepare
python workflow.py train
python workflow.py package
python workflow.py predict
python workflow.py evaluate
python workflow.py runtime
python workflow.py export

The generated project owns its model, native objective, preprocessing, file producer, scorers, runtime entrypoint and result writer. Project checks and status/explain inspect existing declarations and evidence without running application callbacks. The complete installed-user tutorial is also included as TUTORIAL.md in the project. It demonstrates interrupted recovery, cache reuse, committed file outputs and an offline archive consumer.

One Training Workflow

Native models, attached programs and pretrained workflows use one session:

import arti

# model is a configured native model or a host with trained ARTI parameters.
session = arti.trainer(model, objective=task_objective)
print(session.plan(steps=1000))
assert session.check(example_batch)["ok"]
result = session.fit(
    train_loader,
    steps=1000,
    run=arti.ARTITrainingRunConfig(output_dir="outputs/my-task"),
)
metrics = session.evaluate(validation_loader)

The local Torch profile saves the first update, periodic checkpoints and a final checkpoint. Continuation retains optimizer, scheduler, supported RNG/scaler state and the declared data cursor. Evaluation preserves model and ARTI state. The host's task loss is the default; tensor alignment is an explicit alternative.

Use the training guide for native, attachment and pretrained entrances, data-resume guarantees and export.

Real Task Recipes

The same lifecycle is available from JSON/TOML:

arti recipe plan recipe.toml --output plan.json
arti recipe check recipe.toml
arti recipe train recipe.toml --run-dir outputs/first
arti recipe train recipe.toml --resume-from-checkpoint outputs/first/checkpoints/step-00000200/model.arti.st --steps 40 --run-dir outputs/continued
arti recipe evaluate recipe.toml --model outputs/continued/model --split test --output outputs/test.json
arti recipe predict outputs/continued/model --data requests.jsonl --output outputs/predictions.jsonl
arti recipe serve outputs/continued/model --port 8080

• Classification: explicit train/validation/test records, train-only preprocessing, cross-entropy, metrics and portable batched prediction without the original dataset.
• Qwen SFT: complete conversation input, final-assistant/EOS supervision, target-token-weighted accumulation, pinned base revision, checkpoint continuation and tokenizer-bound deployment.
• Optional FastAPI/Uvicorn serving loads the asset once per worker and exposes /predict, /chat, /health and OpenAPI. It is a local ASGI integration, not an authenticated gateway or a vLLM/TensorRT-LLM export.

Runnable repository examples:

uv sync --extra dev --extra qwen --extra serve
python examples/classification_workflow.py --output .tmp/digits-workflow
python examples/qwen_sft_workflow.py --prepare-only --output .tmp/qwen-sft

Qwen planning loads native configuration/tokenization and a meta-device model, not base weights or CUDA allocations. check deliberately performs a real gradient preflight. The recipe's fixed Formula adaptation profile is one usable composition, not a claim that all ARTI features or architecture search are enabled.

Compose Neural Programs

ARTILayer hosts a configured FederatedProgram or ProgramGraph. An empty layer is an exact identity, not an untrained adapter that pretends to know a task.

import torch
import arti

layer = arti.ARTILayer()
x = torch.randn(2, 8, 64)
assert torch.equal(layer(x), x)

# Configure the graph or federation before attaching a useful program.
layer = arti.ARTILayer(graph=graph, graph_program_id="forward",
                      graph_input_resource_id="input",
                      graph_output_resource_id="output")
model = arti.ARTI.attach(model, layer, layers="model.layers.*")
session = model.arti.trainer(objective="model_loss")

The developer guide and Formula Fabric guide cover typed operations, named ports, routes, parallel stages, joins, loops, tensor resources and compiler boundaries. Learning boundaries allow independently owned optimizer domains and region-specific objectives.

Reusable operations live in arti.mechanisms. Retired layered Recall APIs live only in arti.legacy. Component contracts use content-addressed provenance; package versions and component identities are independent.

Native JAX

arti.jax binds the same Formula/ProgramGraph declarations to native array execution, with explicit parameter and forward-state PyTrees. Use jit, grad, vmap and fixed-horizon scan; integrate custom native modules or Flax without running Torch operations in a compiled JAX invocation.

uv sync --extra dev --extra jax --extra flax
python examples/jax_native_training.py

Read the native JAX guide for supported atoms, resource state, native overrides and limitations. This is not full parity with every Torch dynamic search, custom VJP or optimizer-domain engine.

Assets And Support

ARTI uses SafeTensors and validated component/graph metadata instead of Python pickle for its model assets. Frozen host weights remain separate from adapter exports. Custom Python layers still require their reconstruction code.

Stability distinguishes supported workflow contracts from experimental component surfaces. Release checks cover software correctness, installation, documentation and privacy. They do not imply general task superiority or universal acceleration.

uv run --extra dev python scripts/quality_gate.py quick --fail-fast
uv run --extra dev --extra qwen --extra serve python scripts/quality_gate.py recipes --fail-fast

MIT licensed. Designed and maintained by Thiocy.