ARTI

AI x RT: composable latent tensor layers for PyTorch.

ARTI is a domain-independent neural-network library for transforming hidden tensors at runtime. Its layers work with ordinary tensors and can optionally use coordinates, masks, visibility, latent recall, and compact workspaces.

hidden tensor -> ARTI layer or block -> transformed latent tensor

ARTI does not define a tokenizer, task head, data schema, or business model. Applications remain responsible for encoding their context into tensors.

Version 3.0.6 remains the Stable Candidate baseline. Version 3.0.9a2 is a prerelease for new versioned composition contracts under arti.alpha; it does not promote those components to the stable surface. See Stability and Security.

Install

Add ARTI to a project with uv:

uv add arti-fit

To evaluate the current alpha line explicitly:

uv add --prerelease allow "arti-fit==3.0.9a2"

ARTI requires Python 3.10 or newer and PyTorch 2.2 or newer. The consuming project chooses the appropriate CPU or CUDA build of PyTorch.

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

Optional integrations can be installed as needed:

uv sync --extra jax
uv sync --extra qwen
uv sync --extra peft
uv sync --extra sd
uv sync --extra web

The alpha browser runtime is published separately:

pnpm add @arti-fit/web@alpha

What Is New In 3.0.9 Alpha

Alpha 2 keeps the Python-owned Web parity generator aligned with the experimental stateful API namespace. Recall semantics are unchanged from Alpha 1.

New arti.nn.Recall instances use versioned Recall@4 K-wide execution by default. One query preserves up to eight candidate routes, refines them as independent trajectories, and forwards exactly one winning trajectory. The forward choice is hard; a soft routing surrogate preserves useful gradients. Candidate states are not averaged unless breadth_aggregation="route_weighted" is explicitly requested.

recall = arti.nn.Recall(dim=768, slots=64)  # K=8 when capacity permits
y = recall(x)

wide = arti.nn.Recall(dim=768, slots=64, group_topk=16, breadth=16)
y, branches = wide(x, return_branches=True)
y = wide(x, active_k=4)  # reduce this run's active candidate count

Eight is the portable default; 8-32 is the recommended starting range when Bank capacity and runtime budget permit. breadth=1 keeps one trajectory, and breadth_mode="mixed" explicitly requests the historical mix-before-refine path. Resolving Recall@2 or Recall@3 also preserves those historical mixed semantics. The K-wide bridge is eager-only in this alpha. See Batched Refine.

The alpha runtime also exposes bounded branch batches, Formula application, GPU-resident hot-page binding, explicit tensor transactions, and runtime checkpoint helpers. These remain optional infrastructure contracts; this release makes no hardware-speed or task-quality claim.

What Is New In 3.0.8 Alpha

FormulaFabric is an alpha, fixed-capacity tensor executor for bounded Formula programs. It keeps routing, execution, and commit strength explicit: programs run through the existing Formula implementation, Bank-driven route sources can select operands, and iterative routing can re-query after each complete program as the workspace changes.

ObjectiveExposureBank can provide bounded commit strength to a Formula Fabric stage using an explicit current-or-past query. It does not choose routes, primitives, topology, execution depth, persistence, or host commit authority. Future targets and losses remain outside the forward graph.

Reversible Fold can bind a caller-owned topology source through a checked provenance contract, and Formula Fabric can run inside AdaptivePulse without reimplementing Formula mathematics. These APIs remain optional and alpha. See Formula Fabric for the execution contract.

This prerelease makes no task-quality, hardware-speed, autonomous-memory, or scientific-superiority claim. The arti.st format remains version 1 and the Stable Candidate remains 3.0.6.

What Is New In 3.0.7 Alpha

arti.alpha.AdaptivePulse (arti/pulse@2) is a manifest-bound composition of optional observation, Half, reversible Fold/UnFold, Formula intervention, selective compute, Bank update, and reunion aggregation stages. Disabled stages are true identity operations and are omitted from the component dependency graph.

AdaptiveObservation provides bounded fixed, learned, or Bank-conditioned observation trajectories. Identity, bounded state-affine, and Fourier shift operators are independently selectable. Fourier observation supports eager training and compiled forward; compiled FFT backward is deliberately not claimed. See Adaptive observation.

The vNext contracts keep value transport, support masks, typed operands, component provenance, and execution budgets explicit. These APIs are alpha: the release establishes composable tensor contracts, not a task-performance claim. See Component provenance for the canonical identity and dependency rules.

What Is New In 3.0

The 3.0.6 maintenance release adds versioned reversible topology operations under arti.alpha, alongside the target-addressable forward memory updater, and keeps the reversible pretrained-model workflow: after fit and arti.st export, a fresh model can reload the artifact and the workflow can detach() without losing the native model class, methods, or original trainability settings.

Half now makes sampling an explicit activation option. Half(stochastic=True) (the default) samples each feature using its survival probability in both train and eval modes; Half(stochastic=False) returns the deterministic q * x path. Set learnable=True to train the threshold, base, and scale of the survival curve. The learned q curve is available through half.survival(x); the stochastic learned path uses a straight-through estimator.

TargetBankUpdater treats the Bank being changed as an addressable Recall partition. Each bounded write-refine step reads the current Bank, applies a Formula transition, updates the Bank, and lets the next step address the new state. An optional private partition may participate in the read, but only the target Bank is changed. The API remains alpha and is deliberately separate from optimizers, event logs, and task-specific training loops:

from arti.alpha import TargetBankUpdater, WriteRefinePolicy

updater = TargetBankUpdater(
    hidden_dim=64,
    slots=32,
    target_coupling="required_after_bootstrap",
    policy=WriteRefinePolicy.adaptive(max_steps=8, min_steps=2),
)
next_bank = updater(trace, bank, exposure=1.0)

arti.alpha.Fold (arti/fold@2) selects a fixed-size active workspace by reordering tensor instances and records the complete permutation. The paired arti.alpha.UnFold (arti/unfold@2) restores the original topology exactly; it never predicts or reconstructs discarded values because no values are discarded:

from arti.alpha import Fold, UnFold

fold = Fold(active_count=16)
unfold = UnFold(active_count=16)
state = fold(x, mask)
state = state.replace(active=block(state.active))
result = unfold(state)

Learned and Bank Formula topology policies can decide which instances enter the active workspace without mixing tensor values. The existing arti.nn.Fold and arti.nn.UnFold remain the @1 contracts. See Reversible topology.

ARTI 3.0 makes the current Recall architecture the public default. Recall now uses a fixed query basis, host-dimensional Bank values, versioned Formula contracts, explicit per-Bank composition, and bounded iterative refinement. The package also exposes reusable Recall artifact, expert, policy, workspace, and value-transition primitives without coupling them to a training loop.

The experimental RecallTTTSession API and the 2.x Formula symbols have been removed. They mixed optimization policy with the tensor layer and are not part of the 3.x replacement. Applications should compose arti.nn.Recall, Formula contracts, and explicit artifact/state APIs instead.

The arti.st container remains version 1, but the Recall Formula contract and manifest schema are version 2. ARTI 3.0 deliberately rejects 2.x Formula contracts, manifests, and locks instead of silently interpreting them under a different execution contract.

Choose The Smallest Useful Surface

ARTI mechanisms are independent. A project can use one tensor layer, attach Recall to selected model boundaries, or compose separately trained Bank assets. Coordinates, masks, visibility, Recall, Pulse, and the other mechanisms do not need to be enabled together.

Attach At Explicit Tensor Boundaries

ARTI can scan a real sample forward, select module input or output tensor boundaries, and preview the exact parameter cost before changing the model. Placement and scale remain application choices:

import arti

project = (
    arti.project(model)
    .at(
        ["model.layers.*"],
        exclude=["*.lm_head"],
        positions="output",
        scale_pattern={"model.layers.0": "small", "model.layers.*": "medium"},
    )
)

preview = project.preview(sample_batch)  # no model mutation
print(preview.insertion_plan.to_dict())
project.insert()

This is not a model-specific patch list. ARTI temporarily packs the selected tensor into [B, D] or [B, N, D], applies the configured tensor layer, and restores the original rank and output container.

For provider-backed pretrained workflows, keep the lifecycle explicit:

workflow = arti.pretrained(model, provider="transformers")
workflow.scan(sample_batch).plan(where="mlp", scale="tiny")
workflow.apply()
workflow.fit(train_data)
exported = workflow.export("arti.st")

restored = arti.pretrained(fresh_model, provider="transformers")
restored.scan(sample_batch).plan(where="mlp", scale="tiny")
restored.apply()
restored.load_weights(exported.saved.weights_path)
tokens = restored.generate(**inputs)
restored.detach()

detach() removes only the adapters owned by that workflow, restores the pre-attachment requires_grad settings, and leaves the host model's native API available.

Build Reusable Recall Experts

A Recall expert artifact can contain only trainable Bank tensors. The host and shared reader are frozen and fingerprinted by an immutable contract:

import arti
import torch

attached = arti.ARTI.attach(
    model,
    recall={"layers": "model.layers.*", "rank": 16, "slots": 32},
)

contract = attached.arti.expert_contract(
    "qwen-recall-v1",
    model_id="Qwen/Qwen3-0.6B",
)
attached.arti.freeze_expert_banks()
optimizer = torch.optim.AdamW(
    attached.arti.parameters("expert_banks"),
    lr=1e-3,
)

# Run the application-owned training loop, then export only the Banks.
attached.arti.save_expert(
    "style.recall.arti.st",
    expert_id="style",
    contract=contract,
)

Compatible immutable experts can be rebuilt into one native Bank assembly:

experts = attached.arti.experts(contract)
experts.replace(["style.recall.arti.st", "domain.recall.arti.st"])
print(experts.expert_ids)

For fit-exported adapters, the equivalent lower-level composition keeps each Bank independently controllable:

arti.concatenate_adapter_banks(
    model,
    ["style.recall.arti.st", "domain.recall.arti.st"],
    bank_names=["style", "domain"],
    weights={"style": 2.0, "domain": 1.0},
)
arti.set_adapter_bank_weights(model, {"style": 1.0, "domain": 3.0})
arti.set_adapter_bank_influences(model, {"style": 1.0, "domain": -0.5})

Weights change routing priors. Signed influences change write direction and strength. Neither operation rewrites the source artifacts. See Recall artifacts.

Compose And Run Efficiently

Independent adapters can also be loaded in a hash-checked declared order:

results = arti.apply_adapter_stack(model, "arti-stack.json", sample_batch=sample)

After attachment, runtime controls do not rewrite weights:

arti.set_recall_refine_steps(model, 4)
arti.set_adapter_scale(model, 0.75)
compiled = arti.compile_adapter_hotpaths(model)

compile_adapter_hotpaths compiles ARTI write paths without compiling the host model. Eager artifacts remain portable and unchanged.

What Was New In 1.9

ARTI 1.9 adds runtime control over the exact number of Recall refinement steps without changing or rewriting adapter weights:

arti.set_recall_refine_steps(model, 6)
arti.set_recall_refine_steps(model, 0)  # exact Recall bypass

arti.set_recall_refine_schedule(model, [1, 1, 3, 3, 6, 6])
arti.set_recall_refine_schedule(
    model,
    {
        "model.layers.0": 1,
        "model.layers.1": 3,
        "model.layers.2": 6,
    },
)

Sequence schedules follow model.named_modules() registration order. Named schedules must cover every attached adapter exactly and are recommended for persistent configuration. ARTI validates the complete schedule before changing any layer, so an invalid depth cannot leave a partially updated model.

The controls only change runtime refinement depth. They do not change adapter parameters, artifact format, optimizer state, or the separately versioned Web runtime. A positive depth cannot enable an adapter that was initialized without a Recall field.

What Was New In 1.8

ARTI provides arti.nn.Recall, a standalone tensor-in/tensor-out layer with an extensible Formula API:

current state + routed Bank factors -> Formula -> next state

The Bank owns trainable tensors and routing. The Formula only defines how the current state and a fixed, named set of factors produce the next state. This separation lets applications change Recall mathematics without rebuilding routing, serialization, masking, or iterative execution.

The release includes:

• canonical versioned arti/delta@1, arti/affine@1, and arti/state@1 formulas;
• local custom formulas implemented as ordinary torch.nn.Module objects;
• explicit process-local registration for trusted application formulas;
• stable factor ordering and passive manifest metadata;
• masked [B, D] and [B, N, D] execution, optional iterative steps, and diagnostics.

Recall formulas do not own optimizers, gradient policy, files, network access, or training schedules. Third-party formula code is never imported from an artifact. Each Formula can expose a pure-data contract and an instance lock; the lock binds the declared formula to its factor layout, hidden dimension, slot count, and execution backend before weights are loaded.

Use ARTI As A Layer

The smallest API behaves like a normal PyTorch layer:

import arti
import torch

layer = arti.nn.Layer(dim=32)
x = torch.randn(4, 16, 32)
mask = torch.ones(4, 16, dtype=torch.bool)

out = layer(x, mask=mask)

assert out.y.shape == (4, 16, 32)
assert out.pooled.shape == (4, 32)
print(out.diagnostics.keys())

For [B, D] inputs, ARTI treats each row as a single token and restores the original rank on output.

Capabilities are opt-in. Enable only the structure carried by the data:

recall_layer = arti.nn.Layer(dim=32, profile="recall")
multisource = arti.nn.Layer(dim=32, profile="multisource", coord_dim=4)

Use Recall As A Layer

Recall can be inserted anywhere a shape-preserving PyTorch layer is useful:

import arti
import torch

recall = arti.nn.Recall(
    dim=64,
    slots=32,
    formula="arti/affine@1",
    steps=2,
)

h = torch.randn(2, 32, 64)
mask = torch.ones(2, 32, dtype=torch.bool)

h, info = recall(h, mask=mask, return_info=True)

assert h.shape == (2, 32, 64)
print(info["recall_steps_executed"])

Recall routes trainable Bank factors and applies a versioned formula to the current state. Its default activation is the public Half policy; choose activation="none" to disable it, or use an explicit Half(stochastic=False) when a deterministic survival path is required. Module train() / eval() does not silently change the selected Half policy. Built-in formulas use canonical IDs: arti/delta@1, arti/affine@1, and arti/state@1. Legacy short names are rejected rather than silently mapped to a different implementation.

slots is the total Bank slot count and must be divisible by the selected formula's factor count. steps, min_steps, and tolerance control bounded iterative execution. Set activation="none" when a Recall application should not use the default Half survival activation.

Define A Local Formula

Applications can pass a trusted local torch.nn.Module. A custom Formula receives one state vector [D] and its ordered factors [F, D], then returns the complete next state [D]:

import torch
import arti


class SignedGate(torch.nn.Module):
    factor_names = ("content", "gate")

    def forward(
        self,
        state: torch.Tensor,
        factors: torch.Tensor,
    ) -> torch.Tensor:
        content, gate = factors.unbind(dim=0)
        return state + torch.tanh(gate) * content


recall = arti.nn.Recall(
    dim=64,
    slots=32,
    formula=SignedGate(),
)
output = recall(torch.randn(2, 16, 64))

Custom formulas run independently for every latent vector, so Formula-side reductions cannot couple batch items or tokens. Their parameters participate in normal autograd and state_dict() handling. For reusable process-local names, register a trusted factory explicitly with arti.register_formula(...).

See Custom Recall formulas for the complete contract, initialization rules, validation checklist, registration model, and portability boundaries.

Half, Fold, and Recall remain independently usable tensor layers.

Expand And Rearrange With UnFold

UnFold exposes values queried from an input tensor and learns a hard, sample-conditioned layout while preserving every original input instance:

import arti
import torch

x = torch.randn(4, 16, 64)
unfold = arti.nn.UnFold(dim=64, exposed=8)
y, exposed_mask = unfold(x, return_exposed_mask=True)

assert y.shape == (4, 24, 64)
assert exposed_mask.shape == (4, 24)

Original values may move, but they are not averaged, interpolated, projected, or discarded. The queried region and layout remain trainable and support masks, optional guide tensors, autograd, CUDA, and arti.st serialization. UnFold is unrelated to torch.nn.Unfold, which extracts image patches. See the UnFold guide.

One UnFold capacity can serve different runtime workspace sizes by passing target_length. Only the required prefix of exposed query parameters is active for that call.

Fuse Compact Workspaces With FusionPulse

FusionPulse is an alpha layer for combining several already compact Pulse workspaces. It learns feature-wise salience in their joint context, applies Half, and lets one shared UnFold query a fixed-size fused workspace:

left = arti.nn.Pulse(k=8, dim=64)(left_fragments)
right = arti.nn.Pulse(k=8, dim=64)(right_fragments)

fusion = arti.nn.FusionPulse(k=8, dim=64)
z = fusion.concat(left, right)

assert z.shape == (left.shape[0], 8, 64)

Inputs may have different slot counts and the number of sources may change between calls. For balanced consolidation during training, request diagnostics and add info["structural_loss"] to the task loss. See the FusionPulse guide.

Attach To An Existing Model

ARTI can discover and attach Recall branches without changing the model class:

import arti

model = arti.ARTI.attach(
    model,
    recall={
        "layers": "model.layers.*",
        "rank": 16,
        "slots": 8,
    },
)

print(model.arti.summary())
model.arti.save("assistant.recall.arti.st")

Attachment configuration supports explicit layer paths, per-layer dimensions, independent Recall lines, Half switches, resource previews, and reversible removal. Transformers, PEFT, and Diffusers are optional integration boundaries; the core package remains PyTorch-first.

Save And Load Weights

ARTI uses SafeTensors with JSON integrity sidecars:

saved = arti.save(layer, "layer.arti.st")
loaded = arti.load("layer.arti.st", model=fresh_layer)

print(saved.weights_sha256)
print(loaded.missing_keys, loaded.unexpected_keys)

ARTI 3.x reads compatible format-version 1 artifacts produced by the pre-public 0.x and public 1.x lines. Legacy .pt migration uses PyTorch's restricted tensor-only loader:

arti.migrate_pt("legacy-state.pt", "layer.arti.st")

Artifact hashes detect modification relative to their lock files; they are not publisher signatures. Obtain models and weights from trusted sources.

Inspect Composite Components

ARTI keeps composition as ordinary PyTorch composition. The optional component graph records nested ARTI modules, mount paths, typed application bindings, shared parameter objects, and a closure fingerprint without importing code from an artifact:

graph = arti.component_graph(model)
arti.validate_component_graph(graph)
saved = arti.save(model, "model.arti.st")

The graph is an inspectable architecture contract; it does not change tensor execution or require a special composite base class. Unregistered PyTorch modules remain opaque nodes, and legacy arti.st manifests without a graph remain loadable.

Public Modules

• arti.nn: Layer, Half, Fold, UnFold, Pulse, alpha Recall, alpha FusionPulse, RecallRefiner, and visual workspace modules.
• arti: complete ARTI layers, residual blocks, reference models, attachment, serialization, and diagnostics.
• arti.fit: boundary scanning, planning, attachment, artifact stacks, Bank composition, runtime scaling, and ARTI-only hotpath compilation.
• Recall expert APIs: immutable contracts, Bank-only SafeTensors artifacts, named assemblies, per-Bank routing weights, and signed influences.
• arti.torch: backend-explicit aliases for PyTorch applications.
• arti.jax: optional functional JAX subset with array-only parameter trees, JIT, whole-tree gradients, and batch/VMAP-consistent single-sample APIs.
• arti.functional: mask, visibility, pooling, coordinate-frame, and activation helpers.

Experimental and legacy APIs are identified in their docstrings and are not frozen at the same level as the supported core surface.

ARTI remains PyTorch-first. The JAX namespace does not provide attachment, training helpers, Recall, serialization, or full ARTILayer parity.

WebGPU Alpha

arti.web.export(...) calls the real Python module and compiles its named tensor inputs and outputs into a hashed artifact v2 ONNX graph. The separate @arti-fit/web package is a generic executor: it contains no Half, Fold, Pulse, Recall, q, or mask rules. It uses WebGPU and falls back to WebAssembly when device: "auto" is selected. See WebGPU Alpha.

The binding also provides a CPU-friendly predict() path, contract-aware tensor factories, structured errors, cancellable loading, Python-generated artifact-specific TypeScript clients, and a native module Worker example. The low-level run() API remains available for GPU-resident and preallocated tensor workflows.

Inspectable exports flatten tensors from the module's real forward(..., return_info=True) result into Python-declared ONNX outputs. module.inspect(...) selectively retains and downloads those outputs while reporting device and timing metadata through an explicitly disposable result. JavaScript treats workspace, diagnostic, mask, and index labels as contract metadata; it does not implement their ARTI semantics.

Stateful Recall can be exported as paired read/update artifact v3 graphs and loaded with loadArtiStateful(...). Model parameters remain read-only; mutable state is explicit, fixed-size, bounded by caller budgets, and non-persistent unless the application requests a snapshot.

Develop

git clone https://github.com/ragnarok-io/ARTI.git
cd ARTI
uv sync --extra dev
uv run --extra dev pytest
uv build

The test suite covers tensor shapes, masks, gradients, serialization, malformed artifacts, public API imports, and optional backend boundaries. Contribution guidance is in CONTRIBUTING.md.

Citation And Authorship

ARTI was initiated and designed by Thiocy. Citation metadata is provided in CITATION.cff. The project also documents authorship and AI assistance.

License

MIT