welvet 0.80.4

Loom M-POLY-VTD AI engine for Flutter/Dart — 21 numerical types, WebGPU, native .entity checkpoints, and Welvet C-ABI FFI bindings.

welvet

M-POLY-VTD AI Engine (Loom v0.80.0) — Flutter/Dart FFI bindings for the Loom deterministic neural VM: 21 numerical types, volumetric 3D grids, CPU/GPU training paths, DNA evolution, JSON + native .entity checkpoints.

Loom core: README · Bedrock validation: docs/bedrock_validation.md

This is the pub.dev package page content (same role as @openfluke/welvet on npm or welvet on PyPI).

What this package is

welvet ships prebuilt Welvet C-ABI natives (libwelvet.so, welvet.dll, Welvet.xcframework, …) inside a Flutter FFI plugin. Your Dart/Flutter app calls Loom through dart:ffi — same engine as Go poly/, Python ctypes, and TypeScript WASM.

Binding	Best for
This package (Dart FFI)	Flutter apps, Dart CLI on desktop/mobile with bundled natives
@openfluke/welvet npm	Browser, Node.js — WASM
welvet PyPI	Servers, notebooks — ctypes C-ABI
Go poly/	Reference, Lucy harness

Features

• Flutter FFI plugin — natives bundled per platform (Android, iOS, Linux, macOS, Windows).
• 21 DTypes — runtime morphLayer() per layer index.
• Volumetric grid — depth × rows × cols cells, Lucy-style JSON networks.
• Training — loomLib.train() with CPU SC/MC modes (configureTrainingMode).
• Polymorphic forward/backward — shape-aware tensors (MHA [batch, seq, d_model], CNN, etc.).
• Persistence — JSON wire (serializeNetwork) and .entity binary when the native build exports entity APIs.
• LLM — createLLM, llmGenerate, streaming llmPollToken (SoulGlitch-style apps).
• DNA — extractDNA, compareDNA, spliceDNA, NEAT configs.

Installation

dependencies:
  welvet: ^0.80.0

import 'package:welvet/loom_ffi.dart';
import 'package:welvet/welvet.dart'; // welvetVersion

print(welvetVersion); // 0.80.0

Supported platforms (64-bit): Linux (x86_64, ARM64), macOS (universal/arm64), Windows (x86_64, ARM64), Android (arm64-v8a, x86_64), iOS (device + simulator via XCFramework).

Monorepo / path dependency

dependencies:
  welvet:
    path: ../loom/welvet/dart

Build natives from source

PyPI/npm ship prebuilt binaries. To refresh from your Loom checkout:

cd loom/welvet/cabi/internal/build
./build_unix.sh all          # or platform-specific target

cd ../../../dart
bash tool/copy_native.sh --all
flutter test

Without native/ artifacts, loomAvailable is false — run copy_native.sh first.

Examples (runnable)

Scripts in tool/examples/ mirror Python welvet/python/examples/ and the npm README quick-start.

cd loom/welvet/dart
bash tool/copy_native.sh

cd tool/examples
dart pub get
dart run 01_dense_forward.dart
dart run run_all.dart          # runs 01–05

Script	What it shows
01_dense_forward.dart	Volumetric JSON → forwardPolymorphic + sequentialForward
02_morph_and_train.dart	morphLayer(INT8), CPU MC train()
03_save_reload.dart	JSON wire + .entity roundtrip
04_mha_forward.dart	MHA with [batch, seq, d_model]
05_dna_compare.dart	extractDNA + compareDNA

Flutter demo app: example/ (plugin template — flutter run from example/).

Quick start

Layers live on a 3D grid (z, y, x, l) — see docs/overview.md.

1. Load library and build a network

import 'dart:convert';
import 'package:welvet/loom_ffi.dart';

if (!loomAvailable) {
  throw StateError('Welvet missing: $loomLibLastError');
}

final json = jsonEncode({
  'id': 'demo-dense',
  'depth': 1,
  'rows': 1,
  'cols': 1,
  'layers_per_cell': 2,
  'layers': [
    {
      'z': 0, 'y': 0, 'x': 0, 'l': 0,
      'type': 'dense',
      'dtype': 'float32',
      'input_height': 16,
      'output_height': 8,
      'activation': 'relu',
    },
    {
      'z': 0, 'y': 0, 'x': 0, 'l': 1,
      'type': 'dense',
      'dtype': 'float32',
      'input_height': 8,
      'output_height': 4,
      'activation': 'linear',
    },
  ],
});

final handle = loomLib.createNetwork(json);

2. Forward (shape-aware)

final input = List<double>.generate(16, (i) => 0.1);
final inShape = [1, 16];

loomLib.configureTrainingMode(handle, 2); // CPU MC
final outJson = loomLib.forwardPolymorphic(handle, input, inShape);
final output = loomParseFloatArray(outJson); // List<double>?

3. Morph precision

loomLib.morphLayer(handle, 0, LoomLib.dtypeInt8);
loomLib.syncInferenceWeights(handle);

DType IDs: LoomLib.dtypeFloat32, dtypeInt8, etc. (0–20, see Loom numerical types doc).

4. Training

final target = List<double>.generate(4, (_) => 0.5);
final trainResult = loomLib.train(
  handle,
  input,
  target,
  batchSize: 1,
  inDim: 16,
  outDim: 4,
  epochs: 50,
  learningRate: 0.05,
  mode: 2, // CPU MC
  inputShape: [1, 16],
  targetShape: [1, 4],
);
final hist = (jsonDecode(trainResult)['loss_history'] as List).cast<num>();

5. Save / reload

JSON wire (always available):

final wire = loomLib.serializeNetwork(handle);
final reloaded = loomLib.deserializeNetwork(wire);
loomLib.syncInferenceWeights(reloaded);
// forward on reloaded, then loomLib.freeNetwork(reloaded)

.entity wire (when native exports LoomSerializeEntity):

final bytes = loomLib.serializeEntity(handle);
final h2 = loomLib.deserializeEntity(bytes);
loomLib.syncInferenceWeights(h2);

6. LLM (local inference)

final llm = loomLib.createLLM(
  '/path/to/snapshot',
  execMode: 3,
  precision: 4,
  useGPU: true,
);
final reply = loomLib.llmGenerate(llm, 'You are helpful.', 'Hello!');
loomLib.freeLLM(llm);

Seven-layer validation (Dart → C-ABI → Loom)

Same Lucy menu [7] logic as loom/lucy, Python benchmark_seven_layer.py, and npm test:seven-layer.

cd loom/welvet/dart
bash tool/copy_native.sh

# Full suite — 10 layer types × 21 dtypes × grids (slow)
dart run welvet:seven_layer

# One layer
dart run welvet:seven_layer -- --layer Dense
dart run welvet:seven_layer -- --layer MHA

# Single dtype row (fast)
dart run welvet:seven_layer -- --layer Dense --dtype Float32

Layer names: Dense, SwiGLU, MHA, CNN1, CNN2, CNN3, RNN, LSTM, Embedding, Residual.

Each row checks: SC/MC forward & backward parity, train, JSON save/reload, .entity roundtrip (when exported).

Automated tests (CI-friendly subset):

flutter test

Test	Mirrors
test/consumer_smoke_test.dart	Python consumer_smoke.py
test/seven_layer_suite_test.dart	Dense · Float32 · all grids
test/library_load_test.dart	C-ABI load + methods manifest

Cross-check with Python:

cd loom/welvet/python && python benchmark_seven_layer.py --layer Dense

API surface (loomLib / LoomLib)

Import: package:welvet/loom_ffi.dart

Category	Methods
Lifecycle	createNetwork, freeNetwork, getNetworkInfo, newVolumetricNetwork
Inference	sequentialForward, forwardPolymorphic, backwardPolymorphic, configureTrainingMode
Step mesh	createStepState, setInput, meshStep, getOutput, meshBackward
Training	train, applyGradients
Morph	morphLayer, syncInferenceWeights
Persistence	serializeNetwork, deserializeNetwork, serializeEntity, deserializeEntity
DNA	extractDNA, compareDNA, spliceDNA, defaultNEATConfig
GPU	initWGPU, destroyWGPU, isGPU
LLM	createLLM, llmGenerate, llmStartGenerate, llmPollToken, freeLLM
Tokenizer	loadTokenizer, tokenize, detokenize

Helpers: loomAvailable, loomLibLastError, loomParseFloatArray, loomParseResult, LlmPollEvent.

Publishing to pub.dev

Not a manual upload — use the CLI (like npm publish / twine upload):

flutter pub login
cd loom/welvet/dart
bash tool/publish.sh                         # dry-run (desktop: mac + linux + windows)
bash tool/publish.sh --publish               # upload desktop slice (~59 MB)

Binaries stay out of git. publish.sh runs copy_native.sh locally, copies the package to a temp dir outside the repo, strips iOS/Android natives, and uploads from there.

Size: pub.dev rejects uploads over 100 MB compressed. Default desktop slice (macOS + Linux + Windows) fits; iOS/Android are omitted (use copy_native.sh in the monorepo for mobile).

Version alignment

Component	Version
Loom engine (poly)	0.80.0
pub welvet	0.80.1
npm @openfluke/welvet	0.80.0
PyPI welvet	0.80.0

License

Apache-2.0 — see LICENSE.