welvet 0.0.6

Wrapper for Embedding Loom Via External (C-ABI) Toolchain — GPU-accelerated neural networks with transformer inference

welvet - LOOM Python Bindings

Wrapper for Embedding Loom Via External (C-ABI) Toolchain

High-performance neural network library with transformer inference for Python via C-ABI bindings. CPU-first with reliable execution.

Installation

pip install welvet

Quick Start

🎉 NEW: Simple API (Recommended)

The new simple API provides a streamlined interface with cross-platform consistency:

import welvet
import json

# Create network from JSON configuration
config = {
    "batch_size": 1,
    "grid_rows": 1,
    "grid_cols": 3,
    "layers_per_cell": 1,
    "layers": [
        {"type": "dense", "input_size": 8, "output_size": 16, "activation": "relu"},
        {
            "type": "parallel",
            "combine_mode": "grid_scatter",
            "grid_output_rows": 3,
            "grid_output_cols": 1,
            "grid_output_layers": 1,
            "grid_positions": [
                {"branch_index": 0, "target_row": 0, "target_col": 0, "target_layer": 0},
                {"branch_index": 1, "target_row": 1, "target_col": 0, "target_layer": 0},
                {"branch_index": 2, "target_row": 2, "target_col": 0, "target_layer": 0}
            ],
            "branches": [
                {"type": "parallel", "combine_mode": "add", "branches": [
                    {"type": "dense", "input_size": 16, "output_size": 8, "activation": "relu"},
                    {"type": "dense", "input_size": 16, "output_size": 8, "activation": "gelu"}
                ]},
                {"type": "lstm", "input_size": 16, "hidden_size": 8, "seq_length": 1},
                {"type": "rnn", "input_size": 16, "hidden_size": 8, "seq_length": 1}
            ]
        },
        {"type": "dense", "input_size": 24, "output_size": 2, "activation": "sigmoid"}
    ]
}

# Create network (no handle management!)
welvet.create_network_from_json(config)

# Training data
batches = [
    {"Input": [0.2] * 4 + [0.8] * 4, "Target": [1.0, 0.0]},
    {"Input": [0.9] * 4 + [0.1] * 4, "Target": [0.0, 1.0]},
    {"Input": [0.7] * 4 + [0.3] * 4, "Target": [0.0, 1.0]},
    {"Input": [0.3] * 4 + [0.7] * 4, "Target": [1.0, 0.0]},
]

# Train
training_config = {
    "Epochs": 800,
    "LearningRate": 0.15,
    "UseGPU": False,
    "PrintEveryBatch": 0,
    "GradientClip": 1.0,
    "LossType": "mse",
    "Verbose": False
}
result = welvet.train_simple(batches, training_config)

# Forward pass
output = welvet.forward_simple([0.2] * 4 + [0.8] * 4)
print(f"Output: {output}")  # [0.950, 0.050] → Class 0

# Evaluate with deviation metrics
inputs = [batch["Input"] for batch in batches]
expected = [0, 1, 1, 0]
metrics = welvet.evaluate_network_simple(inputs, expected)
print(f"Quality Score: {metrics['score']}/100")  # 100/100
print(f"Average Deviation: {metrics['avg_deviation']:.2f}%")  # 0.00%

# Save/Load model
model_json = welvet.save_model_simple("my_model")
print(f"Model size: {len(model_json)} bytes")  # ~26KB

# Load in new session
welvet.load_model_simple(model_json, "my_model")
output2 = welvet.forward_simple([0.2] * 4 + [0.8] * 4)
# output2 == output (bit-for-bit identical!)

Simple API Functions:

• create_network_from_json(config) - Create from JSON
• forward_simple(inputs) - Forward pass
• backward_simple(gradients) - Backward pass
• update_weights_simple(learning_rate) - Update weights
• train_simple(batches, config) - Train network
• save_model_simple(model_id) - Save to JSON string
• load_model_simple(json_string, model_id) - Load from JSON
• get_network_info_simple() - Get network info
• evaluate_network_simple(inputs, expected) - Evaluate with metrics

Cross-Platform Consistency: The simple API matches the API in TypeScript, C#, C, and WASM - identical function signatures and behavior!

See examples/grid_scatter_demo.py for a complete working example.

🚀 Transformer Inference (LLMs)

Run LLaMA, SmolLM, GPT-2, and other transformers with streaming support!

import welvet

# Load tokenizer and model
with open('models/SmolLM2-135M-Instruct/tokenizer.json', 'rb') as f:
    welvet.load_tokenizer_from_bytes(f.read())

with open('models/SmolLM2-135M-Instruct/config.json', 'rb') as f:
    config = f.read()
with open('models/SmolLM2-135M-Instruct/model.safetensors', 'rb') as f:
    weights = f.read()

welvet.load_transformer_from_bytes(config, weights)

# Generate text with streaming!
for token in welvet.generate_stream("Once upon a time", max_tokens=50):
    print(token, end='', flush=True)

# Or generate all at once
text = welvet.generate_text("Once upon a time", max_tokens=50, temperature=0.7)
print(text)

Web Interface Example

cd examples
./transformer_web_interface.py ../../models/SmolLM2-135M-Instruct 8080
# Open http://localhost:8080/inference.html

See examples/test_transformer.py for a complete example.

✨ Neural Network Training - Load Complete Models

import welvet

# Load a complete model (structure + all weights) in ONE LINE!
network = welvet.load_model_from_string(model_json, "my_model")

# That's it! Network is ready to use
output = welvet.forward(network, input_data)

# Train it
welvet.backward(network, gradient)
welvet.update_weights(network, learning_rate=0.01)

# Save it
model_json = welvet.save_model_to_string(network, "my_model")

Building Networks from Scratch

import welvet

# Create a neural network with all 6 layer types
network = welvet.create_network(
    input_size=32,
    grid_rows=1,
    grid_cols=1,
    layers_per_cell=7,
    use_gpu=True
)

# Initialize layers using registry-based system
dense1 = welvet.call_layer_init("InitDenseLayer", [32, 32, welvet.Activation.LEAKY_RELU])
conv2d = welvet.call_layer_init("InitConv2DLayer", [4, 4, 2, 4, 3, 2, 1, welvet.Activation.LEAKY_RELU])
attention = welvet.call_layer_init("InitMultiHeadAttentionLayer", [4, 4, 2, welvet.Activation.TANH])
rnn = welvet.call_layer_init("InitRNNLayer", [4, 8, 4, 32])
lstm = welvet.call_layer_init("InitLSTMLayer", [8, 4, 4, 16])
parallel = welvet.call_layer_init("InitParallelLayer", [0, 2])  # concat mode, 2 branches
dense2 = welvet.call_layer_init("InitDenseLayer", [16, 2, welvet.Activation.SIGMOID])

# Set layers in network
welvet.set_layer(network, 0, 0, 0, dense1)
welvet.set_layer(network, 0, 0, 1, conv2d)
welvet.set_layer(network, 0, 0, 2, attention)
welvet.set_layer(network, 0, 0, 3, rnn)
welvet.set_layer(network, 0, 0, 4, lstm)
welvet.set_layer(network, 0, 0, 5, parallel)
welvet.set_layer(network, 0, 0, 6, dense2)

# Prepare training data
batches = [
    {"Input": [0.8] * 16 + [0.2] * 16, "Target": [1.0, 0.0]},
    {"Input": [0.2] * 16 + [0.8] * 16, "Target": [0.0, 1.0]},
]

# Train using high-level API
result = welvet.train(
    network,
    batches,
    epochs=10,
    learning_rate=0.003,
    gradient_clip=1.0,
    loss_type="mse"
)

print(f"Final Loss: {result['FinalLoss']:.6f}")
print(f"Throughput: {result['AvgThroughput']:.0f} samples/sec")

# Clean up
welvet.cleanup_gpu(network)
welvet.free_network(network)

Complete Example: All Layers Test

See examples/all_layers_test.py for a comprehensive test that:

1. Downloads a complete model from localhost:3123
2. Loads it with load_model_from_string() - ONE line!
3. Runs inference and compares outputs
4. Trains to verify weights are mutable

# Start the file server (serves test.json)
cd ../../examples
./serve_files.sh

# Run the test (in another terminal)
cd ../python/examples
python3 all_layers_test.py

Output:

✅ test.json loaded (26.4 KB)
✅ ✨ Model loaded completely! (handle: 1)
✅ All 16 layers with weights loaded automatically!
✅ Outputs match with small differences (expected with softmax)
✅ Weights successfully changed!

Features

• 🧠 8 Layer Types (All CPU): Dense, Conv2D, Multi-Head Attention, LayerNorm, RNN, LSTM, Softmax (10 variants), Parallel (4 combine modes)
• ✅ Full CPU Implementation: Every layer works on CPU with complete forward/backward passes
• 🚀 Reliable CPU Execution: All layers fully tested on CPU with complete forward/backward passes (GPU code exists but untested)
• 🎯 Registry-based Initialization: Dynamic layer creation via call_layer_init() for any layer type
• ⚡ High-Level Training API: Built-in train() function with automatic gradients and loss tracking
• 🎯 Cross-Platform: Pre-compiled binaries for Linux, macOS, Windows, Android
• 📦 Easy Integration: Simple Python API with high-level helpers
• 🔧 Low-Level Access: Direct control over layers and training loop via C-ABI
• 🏗️ Grid Architecture: Flexible grid-based neural network topology
• 📊 Comprehensive Activations: ReLU, Sigmoid, Tanh, Softplus, LeakyReLU, Linear

API Reference

Network Management

load_model_from_string(model_json, model_id="loaded_model") ✨

The Easy Way! Load a complete model (structure + all weights) from JSON string.

Parameters:

• model_json (str): JSON string containing the complete model
• model_id (str): Model identifier (default: "loaded_model")

Returns: Network handle (int)

Example:

# Load from file
with open('model.json', 'r') as f:
    model_json = f.read()

network = welvet.load_model_from_string(model_json, "my_model")
# Done! All layers + weights loaded, ready to use

save_model_to_string(handle, model_id="saved_model")

Save a complete model (structure + all weights) to JSON string.

Parameters:

• handle (int): Network handle
• model_id (str): Model identifier (default: "saved_model")

Returns: JSON string containing the complete model

Example:

model_json = welvet.save_model_to_string(network, "my_model")

# Save to file
with open('model.json', 'w') as f:
    f.write(model_json)

create_network(input_size, grid_rows=2, grid_cols=2, layers_per_cell=3, use_gpu=False)

Creates a new grid-based neural network.

Parameters:

• input_size (int): Number of input features
• grid_rows (int): Grid rows (default: 2)
• grid_cols (int): Grid columns (default: 2)
• layers_per_cell (int): Layers per grid cell (default: 3)
• use_gpu (bool): Enable GPU acceleration (default: False) - Note: GPU code exists but is untested; use CPU for reliable execution

Simplified API:

• create_network(input_size, hidden_size, output_size, use_gpu=False) - Auto-calculates grid

Returns: Network handle (int)

free_network(handle)

Frees network resources.

Parameters:

• handle (int): Network handle

Layer Configuration

Activation (Class)

Activation function constants:

• Activation.RELU (0) - Scaled ReLU (1.1x) activation
• Activation.SIGMOID (1) - Sigmoid activation
• Activation.TANH (2) - Tanh activation
• Activation.SOFTPLUS (3) - Softplus activation
• Activation.LEAKY_RELU (4) - LeakyReLU (0.1x negative slope)
• Activation.LINEAR (5) - Linear (no activation)

Layer Initialization (Registry-based)

call_layer_init(function_name, params)

Dynamically create any layer type using the registry system.

Parameters:

• function_name (str): Name of the layer init function
  • "InitDenseLayer" - Fully-connected layer
  • "InitConv2DLayer" - 2D Convolutional layer
  • "InitMultiHeadAttentionLayer" - Multi-head attention layer
  • "InitRNNLayer" - Recurrent Neural Network layer
  • "InitLSTMLayer" - Long Short-Term Memory layer
• params (list): Parameters for the layer (varies by type)

Returns: LayerConfig dictionary

Examples:

# Dense layer: [inputSize, outputSize, activation]
dense = welvet.call_layer_init("InitDenseLayer", [128, 64, welvet.Activation.RELU])

# Conv2D: [height, width, channels, filters, kernelSize, stride, padding, activation]
conv = welvet.call_layer_init("InitConv2DLayer", [28, 28, 1, 32, 3, 1, 1, welvet.Activation.RELU])

# Attention: [seqLength, dModel, numHeads, activation]
attn = welvet.call_layer_init("InitMultiHeadAttentionLayer", [10, 64, 8, welvet.Activation.TANH])

# RNN: [inputSize, hiddenSize, seqLength, outputSize]
rnn = welvet.call_layer_init("InitRNNLayer", [32, 64, 10, 640])

# LSTM: [inputSize, hiddenSize, seqLength, outputSize]
lstm = welvet.call_layer_init("InitLSTMLayer", [32, 64, 10, 640])

list_layer_init_functions()

Get metadata about all available layer initialization functions.

Returns: List of dictionaries with function metadata

functions = welvet.list_layer_init_functions()
for func in functions:
    print(f"{func['Name']}: {func['Parameters']}")

init_dense_layer(input_size, output_size, activation=0)

Initialize a dense layer configuration.

Parameters:

• input_size (int): Input neurons
• output_size (int): Output neurons
• activation (int): Activation function (use Activation constants)

Returns: Layer configuration dict

set_layer(handle, row, col, layer_index, layer_config)

Set a layer in the network grid.

Parameters:

• handle (int): Network handle
• row (int): Grid row (0-indexed)
• col (int): Grid column (0-indexed)
• layer_index (int): Layer index in cell (0-indexed)
• layer_config (dict): Layer config from init_dense_layer()

configure_sequential_network(handle, layer_sizes, activations=None)

High-level helper to configure a simple feedforward network.

Parameters:

• handle (int): Network handle (must have 1x1 grid)
• layer_sizes (List[int]): Layer sizes [input, hidden1, ..., output]
• activations (List[int], optional): Activation for each layer. Defaults to ReLU for hidden, Sigmoid for output.

Example:

net = create_network(input_size=784, grid_rows=1, grid_cols=1, layers_per_cell=2)
configure_sequential_network(net, [784, 128, 10])  # MNIST classifier

get_network_info(handle)

Get network information.

Returns: Dict with type, gpu_enabled, grid_rows, grid_cols, layers_per_cell, total_layers

Operations

forward(handle, input_data)

Performs forward pass through the network.

Parameters:

• handle (int): Network handle
• input_data (List[float]): Input vector

Returns: Output vector (List[float])

backward(handle, target_data)

Performs backward pass for training.

Parameters:

• handle (int): Network handle
• target_data (List[float]): Target/label vector

update_weights(handle, learning_rate)

Updates network weights using computed gradients.

Parameters:

• handle (int): Network handle
• learning_rate (float): Learning rate for gradient descent

Training Helpers

train_epoch(handle, inputs, targets, learning_rate=0.01)

Train the network for one epoch.

Parameters:

• handle (int): Network handle
• inputs (List[List[float]]): List of input vectors
• targets (List[List[float]]): List of target vectors
• learning_rate (float): Learning rate (default: 0.01)

Returns: Average loss for the epoch (float)

Example:

loss = train_epoch(net, train_inputs, train_targets, learning_rate=0.1)
print(f"Epoch loss: {loss:.4f}")

GPU Management

initialize_gpu(handle)

Explicitly initialize GPU resources.

Returns: True if successful, False otherwise

cleanup_gpu(handle)

Release GPU resources.

Parameters:

• handle (int): Network handle

get_version()

Get LOOM library version string.

Returns: Version string (e.g., "LOOM C ABI v1.0")

Examples

Basic Training Example

import welvet

# Create network with GPU
net = welvet.create_network(
    input_size=4,
    grid_rows=1,
    grid_cols=1,
    layers_per_cell=2,
    use_gpu=True
)

# Configure architecture: 4 -> 8 -> 2
welvet.configure_sequential_network(net, [4, 8, 2])

# Training data
inputs = [[0.1, 0.2, 0.3, 0.4], [0.5, 0.6, 0.7, 0.8]]
targets = [[1.0, 0.0], [0.0, 1.0]]

# Train for 50 epochs
for epoch in range(50):
    loss = welvet.train_epoch(net, inputs, targets, learning_rate=0.1)
    if (epoch + 1) % 10 == 0:
        print(f"Epoch {epoch+1}: loss = {loss:.6f}")

# Test
output = welvet.forward(net, [0.1, 0.2, 0.3, 0.4])
print(f"Output: {output}")

# Cleanup
welvet.cleanup_gpu(net)
welvet.free_network(net)

Custom Layer Configuration

import welvet

# Create network
net = welvet.create_network(
    input_size=10,
    grid_rows=2,
    grid_cols=2,
    layers_per_cell=3,
    use_gpu=False
)

# Configure individual layers
for row in range(2):
    for col in range(2):
        # Layer 0: 10 -> 20 (ReLU)
        layer0 = welvet.init_dense_layer(10, 20, welvet.Activation.RELU)
        welvet.set_layer(net, row, col, 0, layer0)

        # Layer 1: 20 -> 15 (Tanh)
        layer1 = welvet.init_dense_layer(20, 15, welvet.Activation.TANH)
        welvet.set_layer(net, row, col, 1, layer1)

        # Layer 2: 15 -> 5 (Sigmoid)
        layer2 = welvet.init_dense_layer(15, 5, welvet.Activation.SIGMOID)
        welvet.set_layer(net, row, col, 2, layer2)

# Network is now configured
info = welvet.get_network_info(net)
print(f"Total layers: {info['total_layers']}")

welvet.free_network(net)

Transformer API Reference

Loading Models

# Load tokenizer from bytes
result = welvet.load_tokenizer_from_bytes(tokenizer_bytes)
# Returns: {'success': True, 'vocab_size': 49152}

# Load transformer model
result = welvet.load_transformer_from_bytes(config_bytes, weights_bytes)
# Returns: {'success': True, 'num_layers': 30, 'hidden_size': 576, 'vocab_size': 49152}

Text Processing

# Encode text to token IDs
ids = welvet.encode_text("Hello world", add_special_tokens=True)
# Returns: [123, 456, 789]

# Decode token IDs to text
text = welvet.decode_tokens([123, 456, 789], skip_special_tokens=True)
# Returns: "Hello world"

Generation

# Generate text all at once
text = welvet.generate_text("Once upon a time", max_tokens=50, temperature=0.7)

# Generate with streaming (yields tokens one by one)
for token in welvet.generate_stream("Once upon a time", max_tokens=50, temperature=0.7):
    print(token, end='', flush=True)

Testing

Run the included examples to verify installation:

# Test transformer inference
python examples/test_transformer.py ../../models/SmolLM2-135M-Instruct

# Run web interface
python examples/transformer_web_interface.py ../../models/SmolLM2-135M-Instruct 8080

# Basic GPU training test (neural networks)
python examples/train_gpu.py

Or test programmatically:

import welvet

# Test basic functionality
net = welvet.create_network(input_size=2, grid_rows=1, grid_cols=1,
                             layers_per_cell=1, use_gpu=False)
welvet.configure_sequential_network(net, [2, 4, 2])

# Verify forward pass works
output = welvet.forward(net, [0.5, 0.5])
assert len(output) == 2, "Forward pass failed"

# Verify training works
inputs = [[0.0, 0.0], [1.0, 1.0]]
targets = [[1.0, 0.0], [0.0, 1.0]]
loss = welvet.train_epoch(net, inputs, targets, learning_rate=0.1)
assert loss > 0, "Training failed"

welvet.free_network(net)
print("✅ All tests passed!")

Platform Support

Pre-compiled binaries included for:

• Linux: x86_64, ARM64
• macOS: ARM64 (Apple Silicon)
• Windows: x86_64
• Android: ARM64

Building from Source

See the main LOOM repository for building the C ABI from source.

License

Apache License 2.0

Links

• GitHub Repository
• C ABI Documentation
• Issue Tracker