monomind 1.10.0 → 1.10.2

package/packages/@monomind/cli/dist/src/services/ruvector-training.js

@@ -0,0 +1,696 @@
+/**
+ * RuVector Training Service
+ * Real WASM-accelerated neural training using @ruvector packages
+ *
+ * Features:
+ * - MicroLoRA: <1µs adaptation with rank-2 LoRA (2.3M ops/s)
+ * - SONA: Self-Optimizing Neural Architecture (624k learn/s, 60k search/s)
+ * - Flash Attention: 2.49x-7.47x speedup (9k ops/s)
+ * - Trajectory Buffer: Learning from success/failure
+ * - Contrastive Learning: InfoNCE loss
+ *
+ * Backward Compatible: All v1 APIs preserved, SONA adds new capabilities
+ *
+ * github.com/nokhodian/monomind
+ */
+const ALLOWED_RUVECTOR_PACKAGES = new Set([
+    '@ruvector/learning-wasm',
+    '@ruvector/attention',
+    '@ruvector/sona',
+]);
+/**
+ * ESM/CJS interop helper — handles `.default` for CJS modules.
+ * Uses `'default' in mod` check which is safer than `mod.default || mod`.
+ */
+async function importWithInterop(packageName) {
+    if (!ALLOWED_RUVECTOR_PACKAGES.has(packageName)) {
+        throw new Error(`Disallowed dynamic import: ${packageName}`);
+    }
+    const mod = await import(packageName);
+    return ('default' in mod) ? mod.default : mod;
+}
+// Lazy-loaded WASM modules
+let microLoRA = null;
+let scopedLoRA = null;
+let trajectoryBuffer = null;
+let flashAttention = null;
+let moeAttention = null;
+let hyperbolicAttention = null;
+let optimizer = null;
+let contrastiveLoss = null;
+let curriculum = null;
+let hardMiner = null;
+// SONA engine (optional enhancement)
+let sonaEngine = null;
+let sonaAvailable = false;
+// Training state
+let initialized = false;
+let totalAdaptations = 0;
+let totalForwards = 0;
+let totalSonaLearns = 0;
+let totalSonaSearches = 0;
+let lastBenchmark = null;
+// Backend tracking
+let activeBackend = 'js-fallback';
+/**
+ * Get which backend is active for training
+ */
+export function getActiveBackend() {
+    return activeBackend;
+}
+/**
+ * Pure-JS fallback implementations for when WASM is unavailable.
+ * These provide the same API surface with basic linear algebra.
+ */
+class JsMicroLoRA {
+    _dim;
+    _alpha;
+    _lr;
+    _adaptCount = 0n;
+    _forwardCount = 0n;
+    _deltaNorm = 0;
+    _A; // Low-rank A (rank x dim)
+    _B; // Low-rank B (dim x rank)
+    RANK = 2;
+    constructor(dim, alpha, lr) {
+        this._dim = dim;
+        this._alpha = alpha;
+        this._lr = lr;
+        this._A = new Float32Array(this.RANK * dim);
+        this._B = new Float32Array(dim * this.RANK);
+        // Xavier initialization
+        const scale = Math.sqrt(2 / (dim + this.RANK));
+        for (let i = 0; i < this._A.length; i++)
+            this._A[i] = (Math.random() - 0.5) * scale;
+        for (let i = 0; i < this._B.length; i++)
+            this._B[i] = (Math.random() - 0.5) * scale;
+    }
+    adapt_array(gradient) {
+        // Simple gradient update on low-rank matrices
+        let norm = 0;
+        for (let i = 0; i < Math.min(gradient.length, this._A.length); i++) {
+            const delta = -this._lr * gradient[i % gradient.length] * this._alpha;
+            this._A[i] += delta;
+            norm += delta * delta;
+        }
+        this._deltaNorm = Math.sqrt(norm);
+        this._adaptCount++;
+    }
+    adapt_count() { return this._adaptCount; }
+    param_count() { return this._A.length + this._B.length; }
+    forward_array(input) {
+        const output = new Float32Array(this._dim);
+        // y = x + alpha * B @ A @ x  (simplified low-rank)
+        for (let i = 0; i < this._dim; i++) {
+            output[i] = input[i];
+            let sum = 0;
+            for (let r = 0; r < this.RANK; r++) {
+                let dot = 0;
+                for (let j = 0; j < this._dim; j++) {
+                    dot += this._A[r * this._dim + j] * input[j];
+                }
+                sum += this._B[i * this.RANK + r] * dot;
+            }
+            output[i] += this._alpha * sum;
+        }
+        this._forwardCount++;
+        return output;
+    }
+    forward_count() { return this._forwardCount; }
+    adapt_with_reward(improvement) {
+        const scale = improvement * this._lr * this._alpha;
+        let norm = 0;
+        for (let i = 0; i < this._A.length; i++) {
+            const delta = scale * (Math.random() - 0.5);
+            this._A[i] += delta;
+            norm += delta * delta;
+        }
+        this._deltaNorm = Math.sqrt(norm);
+        this._adaptCount++;
+    }
+    delta_norm() { return this._deltaNorm; }
+    dim() { return this._dim; }
+    reset() {
+        this._A.fill(0);
+        this._B.fill(0);
+        this._adaptCount = 0n;
+        this._forwardCount = 0n;
+        this._deltaNorm = 0;
+    }
+    free() { }
+}
+class JsScopedLoRA {
+    adapters = new Map();
+    _dim;
+    _alpha;
+    _lr;
+    _fallback = false;
+    constructor(dim, alpha, lr) {
+        this._dim = dim;
+        this._alpha = alpha;
+        this._lr = lr;
+    }
+    getAdapter(opType) {
+        if (!this.adapters.has(opType)) {
+            if (this._fallback && opType > 0 && this.adapters.has(0)) {
+                return this.adapters.get(0);
+            }
+            this.adapters.set(opType, new JsMicroLoRA(this._dim, this._alpha, this._lr));
+        }
+        return this.adapters.get(opType);
+    }
+    adapt_array(opType, gradient) { this.getAdapter(opType).adapt_array(gradient); }
+    adapt_count(opType) { return this.getAdapter(opType).adapt_count(); }
+    forward_array(opType, input) { return this.getAdapter(opType).forward_array(input); }
+    forward_count(opType) { return this.getAdapter(opType).forward_count(); }
+    adapt_with_reward(opType, improvement) { this.getAdapter(opType).adapt_with_reward(improvement); }
+    delta_norm(opType) { return this.getAdapter(opType).delta_norm(); }
+    set_category_fallback(enabled) { this._fallback = enabled; }
+    total_adapt_count() {
+        let total = 0n;
+        for (const a of this.adapters.values())
+            total += a.adapt_count();
+        return total;
+    }
+    total_forward_count() {
+        let total = 0n;
+        for (const a of this.adapters.values())
+            total += a.forward_count();
+        return total;
+    }
+    reset_all() { this.adapters.clear(); }
+    reset_scope(opType) { this.adapters.delete(opType); }
+    free() { this.adapters.clear(); }
+}
+class JsTrajectoryBuffer {
+    entries = [];
+    capacity;
+    constructor(capacity, _dim) {
+        this.capacity = capacity;
+    }
+    record(_embedding, _opType, _attType, executionMs, baselineMs) {
+        const improvement = baselineMs > 0 ? (baselineMs - executionMs) / baselineMs : 0;
+        if (this.entries.length >= this.capacity)
+            this.entries.shift();
+        this.entries.push({ improvement });
+    }
+    is_empty() { return this.entries.length === 0; }
+    total_count() { return BigInt(this.entries.length); }
+    success_rate() {
+        if (this.entries.length === 0)
+            return 0;
+        return this.entries.filter(e => e.improvement > 0).length / this.entries.length;
+    }
+    mean_improvement() {
+        if (this.entries.length === 0)
+            return 0;
+        return this.entries.reduce((s, e) => s + e.improvement, 0) / this.entries.length;
+    }
+    best_improvement() {
+        if (this.entries.length === 0)
+            return 0;
+        return Math.max(...this.entries.map(e => e.improvement));
+    }
+    high_quality_count(threshold) {
+        return this.entries.filter(e => e.improvement > threshold).length;
+    }
+    variance() {
+        if (this.entries.length < 2)
+            return 0;
+        const mean = this.mean_improvement();
+        return this.entries.reduce((s, e) => s + (e.improvement - mean) ** 2, 0) / (this.entries.length - 1);
+    }
+    reset() { this.entries = []; }
+    free() { this.entries = []; }
+}
+/**
+ * Initialize the RuVector training system.
+ * Attempts to load @ruvector/learning-wasm for WASM-accelerated training.
+ * Falls back to a pure-JS implementation if WASM is unavailable.
+ */
+export async function initializeTraining(config = {}) {
+    const features = [];
+    const dim = Math.min(config.dim || 256, 256); // Max 256 for WASM
+    const lr = config.learningRate || 0.01;
+    const alpha = config.alpha || 0.1;
+    // --- Attempt WASM backend first ---
+    let wasmLoaded = false;
+    try {
+        const fs = await import('fs');
+        const { createRequire } = await import('module');
+        const require = createRequire(import.meta.url);
+        const wasmPath = require.resolve('@ruvector/learning-wasm/ruvector_learning_wasm_bg.wasm');
+        const wasmBuffer = fs.readFileSync(wasmPath);
+        const learningWasm = await import('@ruvector/learning-wasm');
+        learningWasm.initSync({ module: wasmBuffer });
+        microLoRA = new learningWasm.WasmMicroLoRA(dim, alpha, lr);
+        features.push(`MicroLoRA/WASM (${dim}-dim, <1μs adaptation)`);
+        scopedLoRA = new learningWasm.WasmScopedLoRA(dim, alpha, lr);
+        scopedLoRA.set_category_fallback(true);
+        features.push('ScopedLoRA/WASM (17 operators)');
+        trajectoryBuffer = new learningWasm.WasmTrajectoryBuffer(config.trajectoryCapacity || 10000, dim);
+        features.push('TrajectoryBuffer/WASM');
+        activeBackend = 'wasm';
+        wasmLoaded = true;
+    }
+    catch (wasmError) {
+        // WASM not available - fall back to JS implementation
+        const reason = wasmError instanceof Error ? wasmError.message : String(wasmError);
+        console.warn(`[ruvector] WASM backend unavailable (${reason}), using JS fallback`);
+        microLoRA = new JsMicroLoRA(dim, alpha, lr);
+        features.push(`MicroLoRA/JS (${dim}-dim, JS fallback)`);
+        scopedLoRA = new JsScopedLoRA(dim, alpha, lr);
+        scopedLoRA.set_category_fallback(true);
+        features.push('ScopedLoRA/JS (17 operators)');
+        trajectoryBuffer = new JsTrajectoryBuffer(config.trajectoryCapacity || 10000, dim);
+        features.push('TrajectoryBuffer/JS');
+        activeBackend = 'js-fallback';
+    }
+    // --- Attention mechanisms (optional, independent of WASM) ---
+    try {
+        const attention = await importWithInterop('@ruvector/attention');
+        if (config.useFlashAttention !== false) {
+            flashAttention = new attention.FlashAttention(dim, 64);
+            features.push('FlashAttention');
+        }
+        if (config.useMoE) {
+            moeAttention = attention.MoEAttention.simple(dim, 8, 2);
+            features.push('MoE (8 experts, top-2)');
+        }
+        if (config.useHyperbolic) {
+            hyperbolicAttention = new attention.HyperbolicAttention(dim, 1.0);
+            features.push('HyperbolicAttention');
+        }
+        optimizer = new attention.AdamWOptimizer(lr, 0.9, 0.999, 1e-8, 0.01);
+        features.push('AdamW Optimizer');
+        contrastiveLoss = new attention.InfoNceLoss(0.07);
+        features.push('InfoNCE Loss');
+        if (config.totalSteps) {
+            curriculum = new attention.CurriculumScheduler(config.totalSteps, config.warmupSteps || Math.floor(config.totalSteps * 0.1));
+            features.push('Curriculum Learning');
+        }
+        try {
+            hardMiner = new attention.HardNegativeMiner(5, 'semi_hard');
+            features.push('Hard Negative Mining');
+        }
+        catch {
+            // Mining not available, continue without it
+        }
+    }
+    catch (attentionError) {
+        // @ruvector/attention not available - attention features skipped
+        const reason = attentionError instanceof Error ? attentionError.message : String(attentionError);
+        console.warn(`[ruvector] @ruvector/attention unavailable (${reason}), attention features disabled`);
+    }
+    // --- SONA (optional, backward compatible) ---
+    if (config.useSona !== false) {
+        try {
+            const sona = await importWithInterop('@ruvector/sona');
+            const sonaRank = config.sonaRank || 4;
+            sonaEngine = new sona.SonaEngine(dim, sonaRank, alpha, lr);
+            sonaAvailable = true;
+            features.push(`SONA (${dim}-dim, rank-${sonaRank}, 624k learn/s)`);
+        }
+        catch (sonaError) {
+            sonaAvailable = false;
+            if (config.useSona === true) {
+                console.warn('SONA requested but not available:', sonaError);
+            }
+        }
+    }
+    initialized = true;
+    return { success: true, features, backend: activeBackend };
+}
+/**
+ * Operator types for scoped LoRA (0-16)
+ */
+export const OperatorType = {
+    GENERAL: 0,
+    ATTENTION: 1,
+    MLP: 2,
+    EMBEDDING: 3,
+    NORMALIZATION: 4,
+    PROJECTION: 5,
+    POOLING: 6,
+    CONVOLUTION: 7,
+    RECURRENT: 8,
+    ROUTING: 9,
+    MEMORY: 10,
+    REASONING: 11,
+    COORDINATION: 12,
+    OPTIMIZATION: 13,
+    SECURITY: 14,
+    TESTING: 15,
+    DEBUGGING: 16,
+};
+/**
+ * Train a pattern with MicroLoRA
+ */
+export async function trainPattern(embedding, gradient, operatorType) {
+    if (!initialized || !microLoRA) {
+        throw new Error('Training system not initialized');
+    }
+    // Use scoped LoRA if operator type specified
+    if (operatorType !== undefined && scopedLoRA) {
+        scopedLoRA.adapt_array(operatorType, gradient);
+        return {
+            deltaNorm: scopedLoRA.delta_norm(operatorType),
+            adaptCount: scopedLoRA.adapt_count(operatorType),
+        };
+    }
+    // Standard MicroLoRA adaptation
+    microLoRA.adapt_array(gradient);
+    totalAdaptations++;
+    return {
+        deltaNorm: microLoRA.delta_norm(),
+        adaptCount: microLoRA.adapt_count(),
+    };
+}
+/**
+ * Forward pass through LoRA
+ */
+export function forward(input, operatorType) {
+    if (!initialized || !microLoRA) {
+        throw new Error('Training system not initialized');
+    }
+    totalForwards++;
+    if (operatorType !== undefined && scopedLoRA) {
+        return scopedLoRA.forward_array(operatorType, input);
+    }
+    return microLoRA.forward_array(input);
+}
+/**
+ * Reward-based adaptation (reinforcement learning)
+ */
+export function adaptWithReward(improvement, operatorType) {
+    if (!initialized) {
+        throw new Error('Training system not initialized');
+    }
+    if (operatorType !== undefined && scopedLoRA) {
+        scopedLoRA.adapt_with_reward(operatorType, improvement);
+    }
+    else if (microLoRA) {
+        microLoRA.adapt_with_reward(improvement);
+    }
+    totalAdaptations++;
+}
+/**
+ * Record a learning trajectory
+ */
+export function recordTrajectory(embedding, operatorType, attentionType, executionMs, baselineMs) {
+    if (!trajectoryBuffer) {
+        throw new Error('Trajectory buffer not initialized');
+    }
+    trajectoryBuffer.record(embedding, operatorType, attentionType, executionMs, baselineMs);
+}
+/**
+ * Get trajectory statistics
+ */
+export function getTrajectoryStats() {
+    if (!trajectoryBuffer || trajectoryBuffer.is_empty()) {
+        return null;
+    }
+    return {
+        successRate: trajectoryBuffer.success_rate(),
+        meanImprovement: trajectoryBuffer.mean_improvement(),
+        bestImprovement: trajectoryBuffer.best_improvement(),
+        totalCount: trajectoryBuffer.total_count(),
+        highQualityCount: trajectoryBuffer.high_quality_count(0.1),
+        variance: trajectoryBuffer.variance(),
+    };
+}
+/**
+ * Compute attention with Flash Attention (2.49x-7.47x faster)
+ */
+export function computeFlashAttention(query, keys, values) {
+    if (!flashAttention) {
+        throw new Error('Flash attention not initialized');
+    }
+    return flashAttention.computeRaw(query, keys, values);
+}
+/**
+ * Compute MoE routing
+ */
+export function computeMoEAttention(query, keys, values) {
+    if (!moeAttention) {
+        throw new Error('MoE attention not initialized');
+    }
+    return moeAttention.computeRaw(query, keys, values);
+}
+/**
+ * Compute hyperbolic attention (for hierarchical patterns)
+ */
+export function computeHyperbolicAttention(query, keys, values) {
+    if (!hyperbolicAttention) {
+        throw new Error('Hyperbolic attention not initialized');
+    }
+    return hyperbolicAttention.computeRaw(query, keys, values);
+}
+/**
+ * Compute contrastive loss for training
+ */
+export function computeContrastiveLoss(anchor, positives, negatives) {
+    if (!contrastiveLoss) {
+        throw new Error('Contrastive loss not initialized');
+    }
+    const loss = contrastiveLoss.compute(anchor, positives, negatives);
+    const gradient = contrastiveLoss.backward(anchor, positives, negatives);
+    return { loss, gradient };
+}
+/**
+ * Optimizer step
+ */
+export function optimizerStep(params, gradients) {
+    if (!optimizer) {
+        throw new Error('Optimizer not initialized');
+    }
+    return optimizer.step(params, gradients);
+}
+/**
+ * Get curriculum difficulty for current step
+ */
+export function getCurriculumDifficulty(step) {
+    if (!curriculum) {
+        return 1.0; // Full difficulty if no curriculum
+    }
+    return curriculum.getDifficulty(step);
+}
+/**
+ * Mine hard negatives for better training
+ */
+export function mineHardNegatives(anchor, candidates) {
+    if (!hardMiner) {
+        throw new Error('Hard negative miner not initialized');
+    }
+    return hardMiner.mine(anchor, candidates);
+}
+/**
+ * Benchmark the training system
+ */
+export async function benchmarkTraining(dim, iterations) {
+    const attention = await importWithInterop('@ruvector/attention');
+    lastBenchmark = attention.benchmarkAttention(dim || 256, 100, iterations || 1000);
+    return lastBenchmark ?? [];
+}
+// ============================================
+// SONA Functions (v2 enhancement, optional)
+// ============================================
+/**
+ * Check if SONA is available
+ */
+export function isSonaAvailable() {
+    return sonaAvailable && sonaEngine !== null;
+}
+/**
+ * Force-learn a pattern with SONA (1.6μs, 624k ops/s)
+ * This is a one-shot learning mechanism for immediate pattern storage
+ */
+export function sonaForceLearn(embedding, reward) {
+    if (!sonaEngine) {
+        throw new Error('SONA not initialized. Call initializeTraining with useSona: true');
+    }
+    sonaEngine.forceLearn(embedding, reward);
+    totalSonaLearns++;
+}
+/**
+ * Search for similar patterns with SONA (16.7μs, 60k searches/s)
+ * Returns the k most similar patterns from the pattern bank
+ */
+export function sonaFindPatterns(embedding, k = 5) {
+    if (!sonaEngine) {
+        throw new Error('SONA not initialized. Call initializeTraining with useSona: true');
+    }
+    // SONA requires Array, not Float32Array
+    const embeddingArray = Array.from(embedding);
+    totalSonaSearches++;
+    return sonaEngine.findPatterns(embeddingArray, k);
+}
+/**
+ * Process SONA background tasks (0.13μs, 7.5M ticks/s)
+ * Call periodically to process background learning and consolidation
+ */
+export function sonaTick() {
+    if (!sonaEngine) {
+        return; // Silent no-op if SONA not available
+    }
+    sonaEngine.tick();
+}
+/**
+ * Get SONA statistics
+ */
+export function getSonaStats() {
+    if (!sonaEngine) {
+        return {
+            available: false,
+            enabled: false,
+            stats: null,
+            totalLearns: totalSonaLearns,
+            totalSearches: totalSonaSearches,
+        };
+    }
+    try {
+        const statsJson = sonaEngine.getStats();
+        const stats = JSON.parse(statsJson);
+        return {
+            available: true,
+            enabled: sonaEngine.isEnabled(),
+            stats,
+            totalLearns: totalSonaLearns,
+            totalSearches: totalSonaSearches,
+        };
+    }
+    catch {
+        return {
+            available: true,
+            enabled: false,
+            stats: null,
+            totalLearns: totalSonaLearns,
+            totalSearches: totalSonaSearches,
+        };
+    }
+}
+/**
+ * Enable/disable SONA learning
+ */
+export function setSonaEnabled(enabled) {
+    if (!sonaEngine) {
+        return;
+    }
+    sonaEngine.setEnabled(enabled);
+}
+/**
+ * Flush SONA buffers (persist any pending patterns)
+ */
+export function sonaFlush() {
+    if (!sonaEngine) {
+        return;
+    }
+    sonaEngine.flush();
+}
+/**
+ * Get training statistics
+ */
+export function getTrainingStats() {
+    const stats = {
+        initialized,
+        backend: activeBackend,
+        totalAdaptations,
+        totalForwards,
+    };
+    if (microLoRA) {
+        stats.microLoraStats = {
+            paramCount: microLoRA.param_count(),
+            adaptCount: microLoRA.adapt_count(),
+            forwardCount: microLoRA.forward_count(),
+            deltaNorm: microLoRA.delta_norm(),
+        };
+    }
+    if (scopedLoRA) {
+        stats.scopedLoraStats = {
+            totalAdaptCount: scopedLoRA.total_adapt_count(),
+            totalForwardCount: scopedLoRA.total_forward_count(),
+        };
+    }
+    if (trajectoryBuffer && !trajectoryBuffer.is_empty()) {
+        stats.trajectoryStats = getTrajectoryStats();
+    }
+    // Include SONA stats if available
+    if (sonaAvailable) {
+        stats.sonaStats = getSonaStats();
+    }
+    if (lastBenchmark) {
+        stats.lastBenchmark = lastBenchmark;
+    }
+    return stats;
+}
+/**
+ * Reset the training system
+ */
+export function resetTraining() {
+    if (microLoRA)
+        microLoRA.reset();
+    if (scopedLoRA)
+        scopedLoRA.reset_all();
+    if (trajectoryBuffer)
+        trajectoryBuffer.reset();
+    // Reset SONA stats (engine doesn't have reset, just flush)
+    if (sonaEngine) {
+        sonaEngine.flush();
+    }
+    totalAdaptations = 0;
+    totalForwards = 0;
+    totalSonaLearns = 0;
+    totalSonaSearches = 0;
+    activeBackend = 'js-fallback';
+}
+/**
+ * Export trained weights
+ */
+export function exportWeights() {
+    if (!initialized || !microLoRA) {
+        return null;
+    }
+    return {
+        dim: microLoRA.dim(),
+        deltaNorm: microLoRA.delta_norm(),
+        adaptCount: microLoRA.adapt_count(),
+        trajectoryStats: getTrajectoryStats(),
+    };
+}
+/**
+ * Cleanup resources
+ */
+export function cleanup() {
+    if (microLoRA) {
+        microLoRA.free();
+        microLoRA = null;
+    }
+    if (scopedLoRA) {
+        scopedLoRA.free();
+        scopedLoRA = null;
+    }
+    if (trajectoryBuffer) {
+        trajectoryBuffer.free();
+        trajectoryBuffer = null;
+    }
+    // Cleanup SONA
+    if (sonaEngine) {
+        sonaEngine.flush();
+        sonaEngine = null;
+        sonaAvailable = false;
+    }
+    flashAttention = null;
+    moeAttention = null;
+    hyperbolicAttention = null;
+    optimizer = null;
+    contrastiveLoss = null;
+    curriculum = null;
+    hardMiner = null;
+    initialized = false;
+    totalAdaptations = 0;
+    totalForwards = 0;
+    totalSonaLearns = 0;
+    totalSonaSearches = 0;
+    lastBenchmark = null;
+}
+//# sourceMappingURL=ruvector-training.js.map