welvet 0.80.0

M-POLY-VTD AI Engine (Loom v0.80.0) — 21 Numerical Types, native .entity checkpoints, WebGPU, and 100% Determinism

welvet — Loom Python Bindings

M-POLY-VTD AI Engine (Loom v0.80.0) — Python bindings: 21 numerical types, volumetric grids, CPU/GPU training, DNA/NEAT, JSON wire + native .entity checkpoints.

welvet wraps the Loom C-ABI with zero runtime Python dependencies. The PyPI wheel ships prebuilt native libraries for every supported OS/arch; at import time only the matching binary is loaded (linux_amd64/welvet.so, windows_amd64/welvet.dll, etc.).

Bedrock validation (v0.80): seven-layer CPU suite (10 layer types × 21 dtypes × train × JSON serialize × .entity roundtrip). See docs/bedrock_validation.md.

---

Install

pip install welvet

import welvet
print(welvet.__version__)  # 0.80.0

Supported platforms (64-bit): Linux (x86-64, ARM64), macOS (x86-64, ARM64, universal fallback), Windows (x86-64, ARM64), Android (ARM64, x86-64), iOS (device / simulator / XCFramework when built into the wheel).

Build from source (monorepo)

PyPI wheels ship prebuilt .so / .dylib / .dll. To run latest main against your checkout:

Option A — build + copy in one step

cd welvet/cabi/internal/build
./build_unix.sh linux amd64    # native Linux x86_64 (+ --test for CABI smoke)
# or: ./build_unix.sh all       # every platform you have cross-toolchains for

build_unix.sh already copies dist/* → welvet/python/src/welvet/.

Option B — you already compiled into dist/ (or copied builds there)

cd welvet/cabi/internal/build
./copy_to_python.sh              # default: ./dist → python/src/welvet/
# or if your tree is elsewhere:
./copy_to_python.sh ../../dist   # e.g. welvet/cabi/dist from build_macos.sh

One-liner from the Python folder:

cd welvet/python
./copy_from_cabi.sh              # copy + pip install -e .

Then install / verify

cd welvet/python
pip install -e .                 # editable install picks up src/welvet/*.so
python3 -m welvet.cabi_verify    # C-ABI symbols + smoke
python3 examples/run_all.py      # README examples
python3 benchmark_seven_layer.py --layer Dense

Artifacts land as python/src/welvet/linux_amd64/welvet.so (etc.). Without that, import welvet fails with “native library not found”.

Publish to PyPI (maintainers)

pip install build twine
cd welvet/cabi/internal/build && ./copy_to_python.sh   # all platforms → src/welvet/
cd ../../../python
pip install -e . && python3 examples/run_all.py        # smoke before release
./publish.sh                                           # python3 -m build + twine upload

The wheel is multi-platform: it contains every */welvet.{so,dylib,dll} you copied; each machine only loads its own folder.

---

Examples (runnable)

Scripts in examples/ mirror the snippets below. Run one file or verify all:

cd welvet/python
pip install -e .
python3 examples/01_dense_forward.py
python3 examples/run_all.py          # runs 01–05

Script	What it shows
01_dense_forward.py	Volumetric JSON → forward_polymorphic + forward
02_morph_and_train.py	morph(INT8), CPU MC train() with shapes
03_save_reload.py	JSON wire + .entity checkpoint roundtrip
04_mha_forward.py	MHA with [batch, seq, d_model]
05_dna_compare.py	dna() + compare_dna()

---

Quick start

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

1. Build and forward

from welvet import Network

net = Network({
    "id": "demo",
    "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"},
    ],
})

inp = [0.1] * 16
out = net.forward_polymorphic(inp, [1, 16])   # preferred: explicit shape
print(out)                                    # length 4

net.free()

2. Morph precision (21 types)

from welvet import DType, Network

# ... same net as above ...
net.morph(0, DType.INT8)          # layer index 0 only
out_q = net.forward_polymorphic(inp, [1, 16])

Use net.morph_all(DType.INT8) to morph every layer that has weights (skips Residual / Softmax).

3. Training (CPU, shape-aware)

from welvet import Network, train

net = Network({...})  # single dense 16→8
inp, tgt = [0.1] * 16, [0.5] * 8
in_shape, out_shape = [1, 16], [1, 8]

net.set_training_mode(2)   # 1 = CPU SC, 2 = CPU MC (multicore on native C-ABI)
losses = train(
    net, [[inp]], [[tgt]],   # [[batch of input rows]], [[batch of target rows]]
    epochs=10, learning_rate=0.05, mode=2,
    input_shape=in_shape, target_shape=out_shape,
)
print(losses[-1])
net.free()

4. Save / reload (JSON wire + .entity)

JSON wire (human-readable checkpoint, same as v0.79):

wire = net.serialize()
copy = Network.deserialize(wire)
# ... forward on copy, then copy.free()

Native .entity blob (compact binary checkpoint, v0.80+):

blob = net.serialize_entity()
copy = Network.deserialize_entity(blob)
copy.sync_inference_weights()   # after training reload, before inference
# ... forward on copy, then copy.free()

layer_persistence_from_entity(blob, layer_index) inspects per-layer weight blobs in a checkpoint without loading a full network.

Full scripts: examples/03_save_reload.py.

---

Core Concepts

Supported Layer Types

Type	Description
dense	Fully connected / linear layer
mha	Multi-Head Attention (with RoPE, GQA/MQA, Causal Masking)
swiglu	SwiGLU gated MLP (LLaMA-style)
rmsnorm	Root Mean Square Normalization
layernorm	Layer Normalization
cnn1 / cnn2 / cnn3	1D / 2D / 3D Convolution
convtransposed1d / 2d / 3d	Transposed Convolution
rnn / lstm	Recurrent layers
embedding	Token embedding lookup
kmeans	Differentiable K-Means clustering
softmax	10 softmax variants (Standard, Gumbel, Masked, Entmax, ...)
parallel	MoE / ensemble branching
sequential	Nested sequential sub-graph
residual	Residual / skip connection

21 Numerical Types

float64, float32, float16, bfloat16, int64/32/16/8, uint64/32/16/8, fp8_e4m3, fp8_e5m2, int4, uint4, fp4_e2m1, int2, uint2, ternary, binary

Morph a layer's precision at runtime (zero realloc when cached):

from welvet import DType, morph_layer

morph_layer(net.handle, layer_index=0, target_dtype=DType.INT8)
# or: net.morph(0, DType.INT8)

---

WebGPU Acceleration

import welvet
from welvet import Network

net = Network({...})

# Upload weights to GPU
welvet.init_wgpu(net._handle)
welvet.sync_to_gpu(net._handle)

# GPU forward pass
output = welvet.forward_wgpu(net._handle, inputs)

net.free()

Numerical tiling (SC vs MC)

v0.80+ uses specialized tiling profiles to maximize throughput:

• SC (Single-Core): Optimized for Edge/WASM/Small NPUs.
• MC (Multi-Core): Optimized for high-bandwidth L1/L2 caches (Ryzen, RTX, M4).

GPU backward training is live for Dense, RMSNorm, CNN 1D/2D/3D — 17x–65x speedup over CPU on real workloads.

---

DNA & Network Comparison

Extract a topological fingerprint and compare networks:

from welvet import Network, compare_dna

dna_a = net_a.dna()
dna_b = net_b.dna()
result = compare_dna(dna_a, dna_b)
print(f"Overlap: {result['OverallOverlap']:.4f}")
print(f"Logic shifts: {len(result.get('LogicShifts', []))}")

---

NEAT Evolution

Genetically evolve a population of networks:

from welvet import (
    default_neat_config, neat_mutate,
    new_neat_population, neat_population_size,
    neat_population_get_network, neat_population_evolve,
    neat_population_best, neat_population_best_fitness,
    neat_population_summary, free_neat_population,
    build_network, free_network,
)

# Create a seed network
seed = Network({
    "id": "seed", "depth": 1, "rows": 1, "cols": 1, "layers_per_cell": 2,
    "layers": [
        {"z": 0, "y": 0, "x": 0, "l": 0, "type": "dense", 
         "input_height": 32, "output_height": 32},
        {"z": 0, "y": 0, "x": 0, "l": 1, "type": "dense", 
         "input_height": 32, "output_height": 1},
    ]
})

cfg = default_neat_config(32)
pop = seed.create_population(size=16, config=cfg)

for gen in range(5):
    fitnesses = [0.5 + 0.1 * i for i in range(pop.size())]
    pop.evolve(fitnesses)
    print(pop.summary(gen))

best = pop.best()
print(f"Best fitness: {pop.best_fitness():.6f}")
best.free()
pop.free()
seed.free()

---

DNA Splice / Genetic Crossover

Combine two parent networks into a child:

from welvet import default_splice_config, splice_dna, splice_dna_with_report

cfg = default_splice_config()
cfg["CrossoverMode"] = "blend"   # "blend" | "point" | "uniform"
cfg["FitnessA"] = 0.8
cfg["FitnessB"] = 0.5

child_handle = splice_dna(parent_a._handle, parent_b._handle, cfg)

# Or get a full diagnostic report
report = splice_dna_with_report(parent_a._handle, parent_b._handle, cfg)
print(f"Layers blended: {report['blended_count']}")
child_handle = report["child_handle"]

---

Step mesh (online learning)

The volumetric 3D grid supports clock-cycle accurate propagation with spatial feedback loops:

state = welvet.create_step_state(net._handle)
welvet.set_input(state, inputs)
welvet.mesh_step(net._handle, state)
output = welvet.get_output(state, layer_idx=-1)
welvet.free_step_state(state)

---

Training

• train(net, batches, …) — poly LoomTrain, shape-aware (used in seven-layer suite). See examples/02_morph_and_train.py.
• train_network(net, inputs, targets, …) — step-mesh clock-cycle path.

GPU backward dispatch and benchmarks: benchmark_training.py, benchmark_seven_layer.py.

---

Tween (neural target propagation)

An alternative to backpropagation using localized Hebbian gap-based learning. We call this tween in APIs; papers often say target propagation.

from welvet import (
    create_tween_state,
    get_default_tween_config,
    tween_forward,
    tween_backward,
)

_ = get_default_tween_config()
handle = create_tween_state(net.handle)
tween_forward(net.handle, handle, inputs)
tween_backward(net.handle, handle, targets)
net.free()

---

LLM Inference

Load a SafeTensors model and run token generation:

from welvet import Network, Tokenizer, sequential_forward

net = Network.from_file("path/to/model.safetensors")
tok = Tokenizer("path/to/tokenizer.json")

ids = tok.encode("Hello, world!")
output = net.forward([float(i) for i in ids])
tok.free()
net.free()

---

Seven-layer validation (Python → CABI)

Same bedrock gate as Lucy and @openfluke/welvet (WASM). Logic in seven_layer_spec.py; engine work stays in the .so.

Each layer × dtype run: forward → train → JSON save/reload → .entity save/reload → sync_inference_weights() → native persistence check.

cd welvet/python
pip install -e .
python3 benchmark_seven_layer.py --layer Dense
python3 benchmark_seven_layer.py --layer Embedding
python3 benchmark_seven_layer.py --layer Residual
# full suite (slow): python3 benchmark_seven_layer.py

---

Platform Support

Platform	Architecture	Binary
Windows	x86-64	welvet.dll
Windows	ARM64	welvet.dll
Linux	x86-64	welvet.so
Linux	ARM64	welvet.so
Linux	ARM (v7)	welvet.so
macOS	ARM64 (M-series)	welvet.dylib
macOS	x86-64	welvet.dylib
macOS	Universal	welvet.dylib
Android	ARM64	welvet.so
Android	x86-64	welvet.so
iOS	ARM64 (device)	welvet.dylib
iOS	Simulator (x86-64)	welvet.dylib
iOS	Simulator (ARM64)	welvet.dylib
iOS	XCFramework (all slices)	.xcframework

At runtime, import welvet resolves welvet/<platform>_<arch>/welvet.{so,dylib,dll} for the current machine (see src/welvet/utils.py).

---

Version alignment

Component	Version
Loom engine (C-ABI / poly)	0.80.0 — Bedrock Validation + .entity
PyPI welvet	0.80.0
npm @openfluke/welvet	0.80.0

---

Links

• GitHub: github.com/openfluke/loom
• Python package: welvet/python
• Engine docs: poly/README.md · docs index
• TypeScript / WASM: @openfluke/welvet
• Issues: github.com/openfluke/loom/issues

---

License

Apache 2.0 — see LICENSE.

Loom: Universal precision. Volumetric freedom. Bedrock performance.