d803bfe2b1
git-subtree-dir: vendor/ruvector git-subtree-split: b64c21726f2bb37286d9ee36a7869fef60cc6900
46 KiB
46 KiB
RuVector Edge-Net Model Optimization and Distribution System
Architecture Document v1.0
Executive Summary
This document defines a comprehensive model optimization and distribution system for RuVector Edge-Net, enabling browser-based AI inference with optimized models delivered via CDN and IPFS. The system supports end-user fine-tuning through MicroLoRA adapters and provides a complete CLI/SDK for model management.
1. System Architecture Overview
+==============================================================================+
| MODEL OPTIMIZATION & DISTRIBUTION SYSTEM |
+==============================================================================+
| |
| +---------------------------+ +---------------------------+ |
| | OPTIMIZATION PIPELINE | | DISTRIBUTION SYSTEM | |
| +---------------------------+ +---------------------------+ |
| | | | | |
| | +---------------------+ | | +---------------------+ | |
| | | Quantization Engine | | | | Google Cloud CDN | | |
| | | - INT8/INT4/FP16 | | | | - Primary delivery | | |
| | | - GPTQ/AWQ/GGML | |---->| | - Global edge PoPs | | |
| | +---------------------+ | | +---------------------+ | |
| | | | | | | |
| | +---------------------+ | | +---------------------+ | |
| | | Pruning & Distill | | | | IPFS Gateway | | |
| | | - Unstructured | | | | - Decentralized | | |
| | | - Knowledge distill | | | | - Content-addressed | | |
| | +---------------------+ | | +---------------------+ | |
| | | | | | | |
| | +---------------------+ | | +---------------------+ | |
| | | ONNX Export | | | | Model Registry | | |
| | | - Graph optimization| | | | - Versioning | | |
| | | - WASM compatible | | | | - Checksums/Sigs | | |
| | +---------------------+ | | +---------------------+ | |
| | | | | |
| +---------------------------+ +---------------------------+ |
| |
| +---------------------------+ +---------------------------+ |
| | MICROLORA SYSTEM | | CLI & SDK | |
| +---------------------------+ +---------------------------+ |
| | | | | |
| | +---------------------+ | | +---------------------+ | |
| | | Browser Training | | | | edge-net-models CLI | | |
| | | - WebGPU/WASM | | | | - pull/push/list | | |
| | | - Rank 1-16 adapters| |<--->| | - optimize/verify | | |
| | +---------------------+ | | +---------------------+ | |
| | | | | | | |
| | +---------------------+ | | +---------------------+ | |
| | | Adapter Merging | | | | JavaScript SDK | | |
| | | - TIES-Merging | | | | - ModelLoader | | |
| | | - DARE weights | | | | - Cache/Fallbacks | | |
| | +---------------------+ | | +---------------------+ | |
| | | | | | | |
| | +---------------------+ | | +---------------------+ | |
| | | P2P Adapter Sharing | | | | Model Loader | | |
| | | - Reputation-gated | | | | - Progressive load | | |
| | | - Domain-specific | | | | - Streaming decode | | |
| | +---------------------+ | | +---------------------+ | |
| | | | | |
| +---------------------------+ +---------------------------+ |
| |
+==============================================================================+
2. Supported Models
2.1 Language Models (Text Generation)
| Model | Params | Quantized Size | Use Case |
|---|---|---|---|
| Phi-1.5 | 1.3B | 700MB (Q4) | Code generation, reasoning |
| Phi-2 | 2.7B | 1.4GB (Q4) | General text, complex tasks |
| Qwen-0.5B | 0.5B | 280MB (Q4) | Fast inference, edge |
| Gemma-2B | 2B | 1.1GB (Q4) | General purpose |
2.2 Embedding Models (Semantic Search)
| Model | Dims | Quantized Size | Use Case |
|---|---|---|---|
| MiniLM-L6-v2 | 384 | 23MB (Q8) | Fast semantic search |
| E5-small | 384 | 34MB (Q8) | High-quality embeddings |
| BGE-small-en | 384 | 34MB (Q8) | Retrieval-optimized |
| GTE-small | 384 | 34MB (Q8) | General text embeddings |
2.3 Vision Models (Future)
| Model | Params | Quantized Size | Use Case |
|---|---|---|---|
| SigLIP-small | 86M | 45MB (Q8) | Image understanding |
| MobileCLIP | 65M | 35MB (Q8) | Fast image-text matching |
3. Model Optimization Pipeline
3.1 Quantization Engine
+--------------------------------------------------------------------------+
| QUANTIZATION PIPELINE |
+--------------------------------------------------------------------------+
| |
| Source Model (FP32/FP16) |
| | |
| v |
| +----------------+ +----------------+ +----------------+ |
| | Calibration |---->| Quantization |---->| Validation | |
| | Dataset | | Algorithm | | (Perplexity) | |
| +----------------+ +----------------+ +----------------+ |
| | | | |
| v v v |
| +-----------+ +-----------+ +-----------+ |
| | WikiText | | INT8 | | PPL < 10 | |
| | C4 | | INT4 | | Accuracy | |
| | Custom | | FP16 | | Latency | |
| +-----------+ +-----------+ +-----------+ |
| | |
| v |
| +------------------------+ |
| | Optimized ONNX Model | |
| | - Quantized weights | |
| | - Fused operators | |
| | - WASM compatible | |
| +------------------------+ |
| |
+--------------------------------------------------------------------------+
3.2 Quantization Formats
| Format | Bits | Memory Reduction | Speed | Quality |
|---|---|---|---|---|
| FP16 | 16 | 2x | 1.2x | ~100% |
| INT8 | 8 | 4x | 1.5x | 99%+ |
| INT4 | 4 | 8x | 2x | 95-98% |
| GPTQ | 4 | 8x | 1.8x | 98%+ |
| AWQ | 4 | 8x | 2x | 97%+ |
| GGML Q4_K | 4 | 8x | 2.2x | 96%+ |
3.3 Optimization Configuration
interface OptimizationConfig {
// Quantization settings
quantization: {
format: 'INT8' | 'INT4' | 'FP16' | 'GPTQ' | 'AWQ';
calibrationSamples: number; // 512 recommended
groupSize: number; // 128 for 4-bit
dampingFactor: number; // 0.01 for GPTQ
useActivationQuantization: boolean; // Dynamic INT8 activations
};
// Pruning settings (optional)
pruning?: {
method: 'magnitude' | 'movement' | 'sparsegpt';
targetSparsity: number; // 0.0-0.9
blockSize: number; // 4 or 8
};
// ONNX export settings
onnx: {
opset: number; // 17 recommended
dynamicAxes: boolean; // For variable seq length
simplify: boolean; // Graph optimization
optimizationLevel: 'basic' | 'extended' | 'all';
};
// Target runtime
target: 'browser' | 'node' | 'edge';
// Output artifacts
outputs: {
model: boolean; // Main ONNX model
tokenizer: boolean; // Tokenizer config
config: boolean; // Model config
adapters: boolean; // LoRA adapter slots
};
}
4. Distribution System
4.1 Storage Architecture
+--------------------------------------------------------------------------+
| MULTI-TIER STORAGE ARCHITECTURE |
+--------------------------------------------------------------------------+
| |
| +------------------------+ |
| | MODEL REGISTRY | (PostgreSQL + Redis) |
| | - Metadata | |
| | - Versions | |
| | - Checksums | |
| | - Signatures | |
| +------------------------+ |
| | |
| v |
| +------------------------+ +------------------------+ |
| | PRIMARY: GCS CDN | | BACKUP: IPFS | |
| | - Fastest delivery | | - Decentralized | |
| | - Global edge caching |<-->| - Content-addressed | |
| | - 99.99% SLA | | - Censorship-resistant| |
| +------------------------+ +------------------------+ |
| | | |
| v v |
| +------------------------+ +------------------------+ |
| | Browser Cache | | P2P Cache (WebRTC) | |
| | - IndexedDB | | - Peer-to-peer | |
| | - Service Worker | | - BitTorrent-style | |
| | - Cache API | | - Gossip protocol | |
| +------------------------+ +------------------------+ |
| |
+--------------------------------------------------------------------------+
4.2 CDN Configuration (Google Cloud Storage)
# bucket: edge-net-models.storage.googleapis.com
storage:
bucket: edge-net-models
region: us-central1
class: STANDARD # Multi-region for hot data
cdn:
enabled: true
cacheMode: CACHE_ALL_STATIC
defaultTtl: 86400 # 24 hours
maxTtl: 2592000 # 30 days
negativeCaching: true
signedUrls: true # For private models
compression:
enabled: true
types:
- application/octet-stream # ONNX models
- application/json # Tokenizers/configs
cors:
origins:
- "*"
methods: ["GET", "HEAD", "OPTIONS"]
responseHeaders:
- Content-Type
- Content-Length
- Content-Encoding
- X-Model-Version
- X-Model-Checksum
4.3 IPFS Integration
interface IPFSModelConfig {
// IPFS gateway endpoints (fallback order)
gateways: [
'https://cloudflare-ipfs.com/ipfs/',
'https://ipfs.io/ipfs/',
'https://gateway.pinata.cloud/ipfs/',
'https://w3s.link/ipfs/',
];
// Pinning services
pinning: {
primary: 'pinata'; // Pinata Cloud
backup: 'web3storage'; // web3.storage
};
// Content addressing
cid: {
version: 1; // CIDv1
codec: 'dag-pb'; // For chunked files
hashAlg: 'sha2-256';
};
// Chunking for large models
chunking: {
enabled: true;
chunkSize: 262144; // 256KB chunks
parallel: 4; // Parallel downloads
};
}
4.4 Model Manifest Format
{
"$schema": "https://edge-net.ruvector.dev/schemas/model-manifest-v1.json",
"id": "phi-1.5-q4",
"name": "Phi-1.5 INT4 Quantized",
"version": "1.0.0",
"type": "text-generation",
"family": "phi",
"architecture": {
"type": "transformer",
"layers": 24,
"hiddenSize": 2048,
"vocabSize": 51200,
"contextLength": 2048
},
"optimization": {
"quantization": "INT4",
"format": "GPTQ",
"groupSize": 128,
"originalSize": "2.7GB",
"optimizedSize": "700MB"
},
"artifacts": {
"model": {
"filename": "model.onnx",
"size": 734003200,
"sha256": "a3f8c2d1e4b5...",
"signature": "MEUCIQDx..."
},
"tokenizer": {
"filename": "tokenizer.json",
"size": 2500000,
"sha256": "b4e9d3c2a1f6..."
},
"config": {
"filename": "config.json",
"size": 1024,
"sha256": "c5f0e4d3b2a7..."
}
},
"distribution": {
"primary": {
"type": "gcs",
"baseUrl": "https://storage.googleapis.com/edge-net-models/phi-1.5-q4/v1.0.0/",
"region": "us-central1"
},
"fallback": {
"type": "ipfs",
"cid": "bafybeig5yh2lj3k4m5n6o7p8q9r0s1t2u3v4w5x6y7z8a9b0c1d2e3f4g5h6",
"gateways": ["cloudflare-ipfs.com", "ipfs.io"]
}
},
"requirements": {
"minMemory": 1073741824, // 1GB
"simdRequired": false,
"webgpuOptional": true,
"browserSupport": ["chrome-88", "firefox-89", "safari-15"]
},
"adapters": {
"baseRank": 8,
"microRank": 2,
"slots": ["reasoning", "coding", "chat", "custom"]
},
"benchmarks": {
"perplexity": 9.2,
"tokensPerSecond": {
"browser": 12,
"node": 45,
"webgpu": 85
},
"memoryUsage": 850000000
},
"license": "MIT",
"homepage": "https://github.com/ruvnet/ruvector",
"publishedAt": "2026-01-03T00:00:00Z",
"publishedBy": "ruvector-team"
}
5. MicroLoRA Customization System
5.1 Architecture Overview
+--------------------------------------------------------------------------+
| MICROLORA CUSTOMIZATION |
+--------------------------------------------------------------------------+
| |
| +----------------------------+ |
| | BROWSER TRAINING | |
| +----------------------------+ |
| | | |
| | User Data | |
| | | | |
| | v | |
| | +----------------------+ | |
| | | Training Pipeline | | +-------------------+ |
| | | - WebGPU/WASM | | | Adapter Storage | |
| | | - Gradient accum | |---->| - IndexedDB | |
| | | - Low-rank updates | | | - P2P sharing | |
| | +----------------------+ | +-------------------+ |
| | | | | |
| | v | v |
| | +----------------------+ | +-------------------+ |
| | | LoRA Adapter | | | Adapter Pool | |
| | | - Rank 1-16 | | | - LRU eviction | |
| | | - ~1-10MB per adapter| |<----| - 16 slots max | |
| | +----------------------+ | +-------------------+ |
| | | | | |
| +-----------|----------------+ | |
| v v |
| +----------------------------+ +-------------------+ |
| | ADAPTER MERGING | | P2P SHARING | |
| +----------------------------+ +-------------------+ |
| | | | | |
| | +----------------------+ | | - Reputation gate | |
| | | TIES-Merging | | | - Domain-specific | |
| | | - Sign agreement | | | - Version control | |
| | | - Trim & scale | | | - Quality scores | |
| | +----------------------+ | +-------------------+ |
| | | | |
| | v | |
| | +----------------------+ | |
| | | Merged Adapter | | |
| | | - Combined expertise | | |
| | | - Optimized weights | | |
| | +----------------------+ | |
| | | |
| +----------------------------+ |
| |
+--------------------------------------------------------------------------+
5.2 MicroLoRA Training Configuration
interface MicroLoRAConfig {
// Adapter architecture
adapter: {
rank: number; // 1-16, typically 2-8
alpha: number; // Scaling factor (rank * 2 typical)
dropout: number; // 0.0-0.1
targetModules: string[]; // ['q_proj', 'v_proj', 'k_proj', 'o_proj']
};
// Training hyperparameters
training: {
batchSize: number; // 1-8 for browser
gradientAccumulation: number;// 4-16 to simulate larger batch
learningRate: number; // 1e-4 to 1e-3
warmupSteps: number; // 10% of total
maxSteps: number; // 100-1000 for quick adapt
optimizer: 'adamw' | 'sgd';
scheduler: 'cosine' | 'linear' | 'constant';
};
// Memory optimization
memory: {
gradientCheckpointing: boolean;
mixedPrecision: boolean; // FP16 forward, FP32 gradients
offloadOptimizer: boolean; // Use IndexedDB for optimizer state
};
// Compute backend
backend: 'wasm' | 'webgpu' | 'auto';
// Data settings
data: {
maxLength: number; // Max sequence length
maskProbability: number; // For MLM tasks
shuffleBuffer: number; // Data shuffling
};
}
5.3 Adapter Format
interface LoRAAdapter {
// Metadata
id: string; // UUID
name: string; // Human-readable name
description: string;
version: string;
createdAt: string;
// Architecture
baseModel: string; // e.g., 'phi-1.5-q4'
rank: number;
alpha: number;
targetModules: string[];
// Weights (per target module)
weights: {
[moduleName: string]: {
lora_A: Float32Array; // Down projection
lora_B: Float32Array; // Up projection
};
};
// Training info
training: {
steps: number;
finalLoss: number;
samples: number;
duration: number; // ms
};
// Quality metrics
quality: {
perplexityDelta: number; // Change from base
taskAccuracy?: number; // If evaluated
humanScore?: number; // 1-5 rating
};
// Sharing
sharing: {
public: boolean;
domain: string; // 'general' | 'code' | 'legal' | etc.
license: string;
};
// Integrity
checksum: string; // SHA-256 of weights
signature?: string; // Ed25519 signature
}
5.4 Adapter Merging Algorithms
// TIES-Merging: Trim, Elect Sign, Disjoint Merge
interface TIESMergeConfig {
method: 'ties';
trimFraction: number; // 0.1-0.3, remove low-magnitude deltas
electSign: 'majority' | 'weighted';
scalingCoef: number; // 0.5-1.0
}
// DARE: Drop And Rescale
interface DAREMergeConfig {
method: 'dare';
dropRate: number; // 0.1-0.5
rescale: boolean;
}
// Task Arithmetic
interface TaskArithmeticConfig {
method: 'task_arithmetic';
weights: { [adapterId: string]: number }; // Weighted combination
}
function mergeAdapters(
adapters: LoRAAdapter[],
config: TIESMergeConfig | DAREMergeConfig | TaskArithmeticConfig
): LoRAAdapter {
// Implementation based on method
// Returns merged adapter with combined expertise
}
6. CLI: edge-net-models
6.1 Command Reference
# Installation
npm install -g @ruvector/edge-net-models
# ======== Model Management ========
# List available models
edge-net-models list
edge-net-models list --type embedding
edge-net-models list --quantization INT4
# Pull a model
edge-net-models pull phi-1.5-q4
edge-net-models pull phi-1.5-q4 --version 1.0.0
edge-net-models pull phi-1.5-q4 --source ipfs # Use IPFS instead of CDN
edge-net-models pull phi-1.5-q4 --output ./models/
# Verify model integrity
edge-net-models verify phi-1.5-q4
edge-net-models verify phi-1.5-q4 --check-signature
# Show model info
edge-net-models info phi-1.5-q4
edge-net-models info phi-1.5-q4 --benchmarks
edge-net-models info phi-1.5-q4 --requirements
# ======== Model Optimization ========
# Quantize a model
edge-net-models optimize ./my-model.onnx \
--quantization INT4 \
--calibration-data ./calibration.jsonl \
--output ./my-model-q4.onnx
# Full optimization pipeline
edge-net-models optimize ./my-model.onnx \
--quantization INT4 \
--pruning magnitude --sparsity 0.3 \
--onnx-simplify \
--target browser
# ======== Model Publishing ========
# Publish to registry
edge-net-models publish ./my-model-q4.onnx \
--name "my-custom-model" \
--version 1.0.0 \
--type text-generation
# Sign model
edge-net-models sign ./my-model-q4.onnx \
--key ./private-key.pem
# ======== Adapter Management ========
# List local adapters
edge-net-models adapters list
# Export adapter
edge-net-models adapters export my-adapter --output ./adapter.json
# Import adapter
edge-net-models adapters import ./adapter.json
# Share adapter to P2P network
edge-net-models adapters share my-adapter --domain coding
# Merge adapters
edge-net-models adapters merge adapter1 adapter2 \
--method ties \
--output merged-adapter
# ======== Cache Management ========
# Show cache info
edge-net-models cache info
# Clear cache
edge-net-models cache clear
edge-net-models cache clear --older-than 30d
# Prune unused models
edge-net-models cache prune
# ======== Configuration ========
# Show config
edge-net-models config show
# Set default source
edge-net-models config set defaultSource gcs
edge-net-models config set cacheDir ~/.edge-net/models
6.2 CLI Configuration File
# ~/.edge-net/models.yaml
version: 1
# Default model source
defaultSource: gcs
# Cache settings
cache:
directory: ~/.edge-net/models
maxSize: 10GB
autoCleanup: true
# CDN settings
cdn:
endpoint: https://storage.googleapis.com/edge-net-models
timeout: 30000
retries: 3
# IPFS settings
ipfs:
gateways:
- https://cloudflare-ipfs.com/ipfs/
- https://ipfs.io/ipfs/
parallel: 4
timeout: 60000
# Model preferences
preferences:
preferQuantization: INT4
preferSmallModels: true
autoUpdate: true
# Signing key (for publishing)
signing:
keyPath: ~/.edge-net/keys/private.pem
keyId: ruvector-team
7. JavaScript SDK
7.1 ModelLoader API
import { ModelLoader, ModelConfig } from '@ruvector/edge-net-models';
// ======== Basic Usage ========
// Create loader with default settings
const loader = new ModelLoader();
// Load a model
const model = await loader.load('phi-1.5-q4');
// Use the model
const output = await model.generate('Hello, world!', {
maxTokens: 100,
temperature: 0.7,
});
// ======== Advanced Configuration ========
const loader = new ModelLoader({
// Cache settings
cache: {
enabled: true,
storage: 'indexeddb', // 'indexeddb' | 'memory' | 'filesystem'
maxSize: 2 * 1024 * 1024 * 1024, // 2GB
},
// Source preferences
sources: {
primary: 'gcs', // Google Cloud Storage CDN
fallback: ['ipfs', 'p2p'], // Fallback sources
timeout: 30000,
},
// Download settings
download: {
parallel: 4, // Parallel chunk downloads
retries: 3,
progressCallback: (progress) => {
console.log(`${progress.loaded}/${progress.total} bytes`);
},
},
// Compute backend
backend: 'auto', // 'wasm' | 'webgpu' | 'auto'
// Memory management
memory: {
maxModels: 2, // Max models in memory
evictionPolicy: 'lru', // 'lru' | 'fifo' | 'priority'
},
});
// ======== Model Loading ========
// Load with specific version
const model = await loader.load('phi-1.5-q4', { version: '1.0.0' });
// Load with adapter
const model = await loader.load('phi-1.5-q4', {
adapter: 'coding-assistant',
});
// Load multiple models
const [llm, embedder] = await Promise.all([
loader.load('phi-1.5-q4'),
loader.load('minilm-l6-v2-q8'),
]);
// ======== Caching ========
// Check if model is cached
const isCached = await loader.isCached('phi-1.5-q4');
// Get cached models
const cachedModels = await loader.getCachedModels();
// Preload model
await loader.preload('phi-1.5-q4');
// Clear specific model from cache
await loader.clearCache('phi-1.5-q4');
// Clear all cache
await loader.clearAllCache();
// ======== Model Info ========
// Get model manifest
const manifest = await loader.getManifest('phi-1.5-q4');
// Get available models
const models = await loader.listModels({
type: 'text-generation',
quantization: 'INT4',
});
// Check requirements
const compatible = await loader.checkCompatibility('phi-1.5-q4');
7.2 InferenceSession API
import { InferenceSession, GenerateConfig } from '@ruvector/edge-net-models';
// ======== Text Generation ========
const session = await InferenceSession.create('phi-1.5-q4');
// Simple generation
const text = await session.generate('Write a haiku about AI:');
// Streaming generation
const stream = session.generateStream('Once upon a time');
for await (const token of stream) {
process.stdout.write(token);
}
// Generation with config
const text = await session.generate('Explain quantum computing:', {
maxTokens: 200,
temperature: 0.7,
topP: 0.9,
topK: 40,
repetitionPenalty: 1.1,
stopSequences: ['\n\n', '###'],
});
// Chat completion
const response = await session.chat([
{ role: 'system', content: 'You are a helpful assistant.' },
{ role: 'user', content: 'What is 2+2?' },
]);
// ======== Embeddings ========
const embedder = await InferenceSession.create('minilm-l6-v2-q8');
// Single embedding
const embedding = await embedder.embed('Hello world');
// Batch embeddings
const embeddings = await embedder.embedBatch([
'Hello world',
'How are you?',
'Nice to meet you',
]);
// Similarity
const similarity = await embedder.similarity(
'king is to queen',
'man is to woman'
);
// ======== MicroLoRA Integration ========
// Enable adapter
await session.loadAdapter('coding-assistant');
// Train adapter on custom data
const trainer = session.createTrainer({
rank: 4,
learningRate: 1e-4,
batchSize: 2,
});
await trainer.train([
{ input: 'Write a Python function', output: 'def hello():...' },
{ input: 'Create a class', output: 'class MyClass:...' },
]);
// Save adapter
const adapter = await trainer.getAdapter();
await adapter.save('my-coding-adapter');
// ======== Memory Management ========
// Get memory usage
const memoryInfo = session.getMemoryInfo();
console.log(`Model memory: ${memoryInfo.modelBytes / 1024 / 1024}MB`);
console.log(`KV cache: ${memoryInfo.kvCacheBytes / 1024 / 1024}MB`);
// Clear KV cache
session.clearKVCache();
// Dispose session
session.dispose();
7.3 Web Worker Integration
import { ModelWorker } from '@ruvector/edge-net-models/worker';
// ======== Main Thread ========
const worker = new ModelWorker();
// Load model in worker
await worker.load('phi-1.5-q4');
// Generate in worker (non-blocking)
const result = await worker.generate('Hello!', { maxTokens: 50 });
// Streaming from worker
worker.onToken((token) => console.log(token));
await worker.generateStream('Once upon a time');
// Batch processing in worker
const results = await worker.batch([
{ id: '1', prompt: 'Hello' },
{ id: '2', prompt: 'World' },
]);
// Terminate worker
worker.terminate();
// ======== Worker Pool ========
import { ModelWorkerPool } from '@ruvector/edge-net-models/worker';
const pool = new ModelWorkerPool({
workers: 4, // Number of workers
model: 'phi-1.5-q4', // Pre-load model
});
// Parallel inference
const results = await Promise.all([
pool.generate('Task 1'),
pool.generate('Task 2'),
pool.generate('Task 3'),
pool.generate('Task 4'),
]);
pool.terminate();
8. File Structure
examples/edge-net/
|-- models/ # Model optimization & distribution
| |-- src/
| | |-- optimization/
| | | |-- mod.rs # Optimization pipeline
| | | |-- quantization.rs # INT8/INT4/FP16 quantization
| | | |-- pruning.rs # Magnitude/movement pruning
| | | |-- distillation.rs # Knowledge distillation
| | | |-- onnx_export.rs # ONNX conversion & optimization
| | | |-- calibration.rs # Calibration dataset handling
| | | `-- validation.rs # Quality validation
| | |
| | |-- distribution/
| | | |-- mod.rs # Distribution system
| | | |-- gcs.rs # Google Cloud Storage CDN
| | | |-- ipfs.rs # IPFS integration
| | | |-- registry.rs # Model registry
| | | |-- manifest.rs # Model manifest handling
| | | |-- integrity.rs # Checksums & signatures
| | | `-- p2p.rs # P2P model distribution
| | |
| | |-- lora/
| | | |-- mod.rs # MicroLoRA system
| | | |-- adapter.rs # LoRA adapter types
| | | |-- training.rs # Browser-based training
| | | |-- merging.rs # TIES/DARE merging
| | | |-- pool.rs # Adapter pool management
| | | `-- sharing.rs # P2P adapter sharing
| | |
| | |-- loader/
| | | |-- mod.rs # Model loader
| | | |-- cache.rs # Caching system
| | | |-- streaming.rs # Streaming decode
| | | `-- fallback.rs # Fallback handling
| | |
| | `-- lib.rs # Library entry point
| |
| |-- cli/
| | |-- src/
| | | |-- main.rs # CLI entry point
| | | |-- commands/
| | | | |-- mod.rs # Command handlers
| | | | |-- pull.rs # Pull model
| | | | |-- push.rs # Push model
| | | | |-- list.rs # List models
| | | | |-- optimize.rs # Optimize model
| | | | |-- verify.rs # Verify integrity
| | | | |-- adapters.rs # Adapter management
| | | | `-- cache.rs # Cache management
| | | |-- config.rs # CLI configuration
| | | `-- output.rs # Output formatting
| | |-- Cargo.toml
| | `-- README.md
| |
| |-- sdk/
| | |-- package.json
| | |-- tsconfig.json
| | |-- src/
| | | |-- index.ts # SDK entry point
| | | |-- loader.ts # ModelLoader class
| | | |-- session.ts # InferenceSession class
| | | |-- cache.ts # Browser caching
| | | |-- sources/
| | | | |-- gcs.ts # GCS CDN source
| | | | |-- ipfs.ts # IPFS source
| | | | `-- p2p.ts # P2P source
| | | |-- adapters/
| | | | |-- trainer.ts # Browser training
| | | | |-- merger.ts # Adapter merging
| | | | `-- pool.ts # Adapter pool
| | | |-- worker/
| | | | |-- index.ts # Worker exports
| | | | |-- model-worker.ts # Model worker class
| | | | |-- pool.ts # Worker pool
| | | | `-- inference.worker.ts # Worker implementation
| | | `-- types.ts # TypeScript types
| | `-- README.md
| |
| |-- manifests/ # Model manifests
| | |-- phi-1.5-q4.json
| | |-- phi-2-q4.json
| | |-- qwen-0.5b-q4.json
| | |-- gemma-2b-q4.json
| | |-- minilm-l6-v2-q8.json
| | |-- e5-small-q8.json
| | `-- bge-small-q8.json
| |
| |-- Cargo.toml
| `-- README.md
|
|-- src/
| |-- ai/
| | |-- models/
| | | |-- mod.rs # Model integration
| | | |-- inference.rs # Inference engine
| | | `-- tokenizer.rs # Tokenization
| | |-- sona/
| | | |-- lora.rs # MicroLoRA (existing)
| | | `-- ...
| | `-- ...
| `-- ...
|
`-- ...
9. API Contracts
9.1 Model Registry API
openapi: 3.0.3
info:
title: Edge-Net Model Registry API
version: 1.0.0
paths:
/v1/models:
get:
summary: List available models
parameters:
- name: type
in: query
schema:
type: string
enum: [text-generation, embedding, vision]
- name: quantization
in: query
schema:
type: string
enum: [FP16, INT8, INT4]
responses:
'200':
description: List of models
content:
application/json:
schema:
type: array
items:
$ref: '#/components/schemas/ModelSummary'
/v1/models/{modelId}:
get:
summary: Get model details
parameters:
- name: modelId
in: path
required: true
schema:
type: string
responses:
'200':
description: Model manifest
content:
application/json:
schema:
$ref: '#/components/schemas/ModelManifest'
/v1/models/{modelId}/download:
get:
summary: Get download URLs
parameters:
- name: modelId
in: path
required: true
schema:
type: string
- name: source
in: query
schema:
type: string
enum: [gcs, ipfs]
default: gcs
responses:
'200':
description: Download URLs
content:
application/json:
schema:
$ref: '#/components/schemas/DownloadUrls'
/v1/adapters:
get:
summary: List public adapters
parameters:
- name: baseModel
in: query
schema:
type: string
- name: domain
in: query
schema:
type: string
responses:
'200':
description: List of adapters
content:
application/json:
schema:
type: array
items:
$ref: '#/components/schemas/AdapterSummary'
post:
summary: Share adapter
requestBody:
content:
application/json:
schema:
$ref: '#/components/schemas/AdapterUpload'
responses:
'201':
description: Adapter shared
content:
application/json:
schema:
$ref: '#/components/schemas/AdapterSummary'
components:
schemas:
ModelSummary:
type: object
properties:
id:
type: string
name:
type: string
type:
type: string
quantization:
type: string
size:
type: integer
version:
type: string
ModelManifest:
type: object
# Full manifest as defined in section 4.4
DownloadUrls:
type: object
properties:
model:
type: string
format: uri
tokenizer:
type: string
format: uri
config:
type: string
format: uri
signature:
type: string
format: uri
AdapterSummary:
type: object
properties:
id:
type: string
name:
type: string
baseModel:
type: string
rank:
type: integer
domain:
type: string
quality:
type: number
downloads:
type: integer
9.2 P2P Adapter Protocol
syntax = "proto3";
package edgenet.adapters;
// Adapter discovery message
message AdapterAnnounce {
string adapter_id = 1;
string name = 2;
string base_model = 3;
uint32 rank = 4;
string domain = 5;
float quality_score = 6;
bytes checksum = 7;
bytes signature = 8;
string peer_id = 9;
uint64 timestamp = 10;
}
// Adapter request
message AdapterRequest {
string adapter_id = 1;
string requester_id = 2;
uint64 timestamp = 3;
}
// Adapter response (chunked for large adapters)
message AdapterChunk {
string adapter_id = 1;
uint32 chunk_index = 2;
uint32 total_chunks = 3;
bytes data = 4;
bytes chunk_hash = 5;
}
// Adapter quality report
message AdapterReport {
string adapter_id = 1;
string reporter_id = 2;
float quality_score = 3;
string feedback = 4;
uint64 timestamp = 5;
bytes signature = 6;
}
10. Security Considerations
10.1 Model Integrity
+--------------------------------------------------------------------------+
| MODEL INTEGRITY CHAIN |
+--------------------------------------------------------------------------+
| |
| Publisher |
| | |
| v |
| +------------------------+ |
| | 1. Generate Checksums | |
| | - SHA-256 per file | |
| | - Merkle root | |
| +------------------------+ |
| | |
| v |
| +------------------------+ |
| | 2. Sign Manifest | |
| | - Ed25519 signature | |
| | - Publisher identity| |
| +------------------------+ |
| | |
| v |
| +------------------------+ |
| | 3. Upload to CDN/IPFS | |
| | - Model artifacts | |
| | - Signed manifest | |
| +------------------------+ |
| | |
| v |
| Consumer |
| | |
| v |
| +------------------------+ |
| | 4. Download & Verify | |
| | - Check checksums | |
| | - Verify signature | |
| | - Validate manifest | |
| +------------------------+ |
| | |
| v |
| +------------------------+ |
| | 5. Load Model | |
| | - If valid: proceed | |
| | - If invalid: abort | |
| +------------------------+ |
| |
+--------------------------------------------------------------------------+
10.2 Adapter Security
- Reputation-gated sharing: Only nodes with reputation > 0.7 can share adapters
- Quality thresholds: Adapters must meet minimum quality scores
- Signature verification: All shared adapters must be signed
- Sandboxed training: Browser training runs in isolated context
- Rate limiting: Max 10 adapter shares per hour per node
10.3 WASM Sandbox
- Models run in isolated WASM sandbox
- No network access from inference
- Memory limits enforced
- Execution time limits
- No filesystem access
11. Performance Targets
| Metric | Target | Notes |
|---|---|---|
| Model Load Time | < 5s (cached), < 30s (download) | For 1GB model |
| First Token Latency | < 100ms (WASM), < 50ms (WebGPU) | After model loaded |
| Tokens/Second | 10-15 (WASM), 50-100 (WebGPU) | Phi-1.5 Q4 |
| Memory Usage | < 1.5x model size | Including KV cache |
| Adapter Training | 100 samples in < 5 min | Rank-4 adapter |
| Adapter Load Time | < 100ms | From IndexedDB |
| Cache Hit Rate | > 95% | After initial load |
| CDN Latency | < 50ms | To nearest PoP |
12. Future Enhancements
Phase 2 (Q2 2026)
- Vision model support (SigLIP, MobileCLIP)
- WebGPU acceleration for all models
- Federated adapter training across P2P network
- Model sharding for very large models
Phase 3 (Q3 2026)
- Mixture-of-Experts (MoE) support
- Speculative decoding
- Cross-model adapter transfer
- On-device model distillation
Phase 4 (Q4 2026)
- Multimodal models (text + image)
- Audio models (Whisper variants)
- Real-time streaming inference
- Hardware acceleration (WebNN)
Appendix A: Quantization Details
A.1 INT4 Quantization Algorithm
def quantize_int4_gptq(weights, calibration_data):
"""
GPTQ-style INT4 quantization with optimal rounding.
1. Compute Hessian approximation from calibration data
2. For each weight column:
a. Quantize to nearest INT4 value
b. Compute quantization error
c. Update remaining weights to minimize error
"""
H = compute_hessian(weights, calibration_data)
for i in range(weights.shape[1]):
# Optimal rounding
q = round(weights[:, i] / scale)
q = clip(q, -8, 7) # INT4 range
# Error compensation (Cholesky update)
error = (weights[:, i] - q * scale) / H[i, i]
weights[:, i+1:] -= error * H[i, i+1:]
return quantized_weights, scales, zeros
A.2 Calibration Dataset Format
{"text": "The quick brown fox jumps over the lazy dog."}
{"text": "Machine learning models can be quantized for efficiency."}
{"text": "WebAssembly enables near-native performance in browsers."}
...
Appendix B: IPFS Content Addressing
B.1 Model CID Computation
// Compute CID for model file
import { CID } from 'multiformats/cid';
import { sha256 } from 'multiformats/hashes/sha2';
import * as dagPB from '@ipld/dag-pb';
async function computeModelCID(modelBytes) {
// Chunk the model (256KB chunks)
const chunks = chunkify(modelBytes, 262144);
// Build Merkle DAG
const links = [];
for (const chunk of chunks) {
const hash = await sha256.digest(chunk);
const cid = CID.create(1, dagPB.code, hash);
links.push({ Hash: cid, Tsize: chunk.length });
}
// Create root node
const node = dagPB.createNode(new Uint8Array(), links);
const rootHash = await sha256.digest(dagPB.encode(node));
return CID.create(1, dagPB.code, rootHash);
}
Appendix C: Signature Verification
C.1 Ed25519 Signature Format
interface ModelSignature {
// Signed data
manifest: {
id: string;
version: string;
artifacts: {
[name: string]: {
sha256: string;
size: number;
};
};
timestamp: string;
};
// Signature
signature: string; // Base64-encoded Ed25519 signature
publicKey: string; // Base64-encoded public key
keyId: string; // Key identifier (e.g., 'ruvector-team')
}
// Verification
async function verifyModelSignature(sig: ModelSignature): Promise<boolean> {
const publicKey = base64ToBytes(sig.publicKey);
const signature = base64ToBytes(sig.signature);
const message = new TextEncoder().encode(JSON.stringify(sig.manifest));
return ed25519.verify(signature, message, publicKey);
}
Document Version: 1.0 Last Updated: 2026-01-03 Authors: RuVector Team