@claude-flow/cli 3.10.36 → 3.10.37

package/dist/src/memory/rabitq-index.js

@@ -81,11 +81,11 @@ export async function buildRabitqIndex(options) {
             const SQL = await initSqlJs();
             const fileBuffer = fs.readFileSync(dbPath);
             const db = new SQL.Database(fileBuffer);
-            const result = db.exec(`
-        SELECT id, key, namespace, embedding
-        FROM memory_entries
-        WHERE status = 'active' AND embedding IS NOT NULL
-        LIMIT 50000
+            const result = db.exec(`
+        SELECT id, key, namespace, embedding
+        FROM memory_entries
+        WHERE status = 'active' AND embedding IS NOT NULL
+        LIMIT 50000
       `);
             if (result[0]?.values) {
                 for (const row of result[0].values) {

package/dist/src/runtime/headless.js

@@ -62,34 +62,34 @@ function parseArgs() {
  * Show help message
  */
 function showHelp() {
-    console.log(`
-Headless Runtime for RuFlo V3
-
-Usage:
-  headless --worker <type>    Run a specific worker
-  headless --daemon           Start background daemon
-  headless --benchmark        Run performance benchmarks
-  headless --status           Show system status
-
-Workers: ${HEADLESS_WORKER_TYPES.join(', ')}
-
-Options:
-  -w, --worker <type>   Worker type to run
-  -d, --daemon          Run as daemon
-  -b, --benchmark       Run benchmarks
-  -s, --status          Show status
-  -t, --timeout <ms>    Execution timeout (default: 60000)
-  -v, --verbose         Verbose output
-  -h, --help            Show help
-
-Environment:
-  CLAUDE_FLOW_HEADLESS=true   Enable headless mode
-  CLAUDE_CODE_HEADLESS=true   Enable Claude Code headless
-
-Examples:
-  headless --worker audit --timeout 120000
-  headless --daemon
-  headless --benchmark
+    console.log(`
+Headless Runtime for RuFlo V3
+
+Usage:
+  headless --worker <type>    Run a specific worker
+  headless --daemon           Start background daemon
+  headless --benchmark        Run performance benchmarks
+  headless --status           Show system status
+
+Workers: ${HEADLESS_WORKER_TYPES.join(', ')}
+
+Options:
+  -w, --worker <type>   Worker type to run
+  -d, --daemon          Run as daemon
+  -b, --benchmark       Run benchmarks
+  -s, --status          Show status
+  -t, --timeout <ms>    Execution timeout (default: 60000)
+  -v, --verbose         Verbose output
+  -h, --help            Show help
+
+Environment:
+  CLAUDE_FLOW_HEADLESS=true   Enable headless mode
+  CLAUDE_CODE_HEADLESS=true   Enable Claude Code headless
+
+Examples:
+  headless --worker audit --timeout 120000
+  headless --daemon
+  headless --benchmark
 `);
 }
 /**

package/dist/src/ruvector/flash-attention.d.ts

@@ -0,0 +1,195 @@
+/**
+ * Flash Attention Implementation for RuVector Intelligence System
+ *
+ * Implements block-wise attention computation for faster similarity calculations.
+ * Achieves O(N) memory instead of O(N^2) through tiling strategy.
+ *
+ * Key optimizations:
+ * - Block-wise computation to fit in L1 cache
+ * - Fused softmax-matmul operations
+ * - Float32Array for all operations
+ * - Online softmax for numerical stability
+ *
+ * Target: 2-5x speedup on CPU vs naive attention
+ *
+ * Created with love by ruv.io
+ */
+export interface FlashAttentionConfig {
+    /** Block size for tiling (32-64 optimal for CPU L1 cache) */
+    blockSize: number;
+    /** Number of dimensions in embedding vectors */
+    dimensions: number;
+    /** Temperature for softmax scaling */
+    temperature: number;
+    /** Enable numerical stability optimizations */
+    useStableMode: boolean;
+    /** Use optimized CPU path (default: true) */
+    useCPUOptimizations: boolean;
+}
+export interface AttentionResult {
+    /** Output vectors after attention */
+    output: Float32Array[];
+    /** Attention weights (optional, for debugging) */
+    weights?: Float32Array[];
+    /** Computation time in milliseconds */
+    computeTimeMs: number;
+}
+export interface BenchmarkResult {
+    /** Naive attention time in milliseconds */
+    naiveTimeMs: number;
+    /** Flash attention time in milliseconds */
+    flashTimeMs: number;
+    /** Speedup factor (naive / flash) */
+    speedup: number;
+    /** Number of vectors benchmarked */
+    numVectors: number;
+    /** Dimensions of vectors */
+    dimensions: number;
+    /** Memory usage estimate for naive (bytes) */
+    naiveMemoryBytes: number;
+    /** Memory usage estimate for flash (bytes) */
+    flashMemoryBytes: number;
+    /** Memory reduction factor */
+    memoryReduction: number;
+}
+export declare class FlashAttention {
+    private config;
+    private lastSpeedup;
+    private benchmarkHistory;
+    private scoreBuffer;
+    private expBuffer;
+    private accumBuffer;
+    constructor(config?: Partial<FlashAttentionConfig>);
+    /**
+     * Main attention computation using Flash Attention algorithm
+     *
+     * @param queries - Query vectors [N x D]
+     * @param keys - Key vectors [M x D]
+     * @param values - Value vectors [M x D]
+     * @returns Attention output [N x D]
+     */
+    attention(queries: Float32Array[], keys: Float32Array[], values: Float32Array[]): AttentionResult;
+    /**
+     * CPU-optimized attention with aggressive optimizations
+     *
+     * Key optimizations:
+     * - Blocked score computation (better cache utilization)
+     * - Top-K sparse attention (only use most relevant keys)
+     * - Pre-allocated buffers to avoid GC pressure
+     * - 8x loop unrolling for dot products
+     * - Fused max-finding during score computation
+     */
+    private cpuOptimizedAttention;
+    /**
+     * Partial dot product using only first N dimensions (for screening)
+     */
+    private partialDotProduct;
+    /**
+     * Partial sort to get top-K elements (QuickSelect-like)
+     * Only ensures first K elements are the largest, not sorted
+     */
+    private partialSort;
+    /**
+     * Swap two indices in array
+     */
+    private swapIndices;
+    /**
+     * Fast dot product with 8x unrolling
+     */
+    private fastDotProduct;
+    /**
+     * Block-wise attention computation (Flash Attention core algorithm)
+     *
+     * Algorithm:
+     * For each block of queries Q_b:
+     *   For each block of keys K_b:
+     *     S_b = Q_b @ K_b.T / sqrt(d)  // Block scores
+     *     P_b = softmax(S_b)           // Block attention
+     *     O_b += P_b @ V_b             // Accumulate output
+     *
+     * @param Q - Query vectors
+     * @param K - Key vectors
+     * @param V - Value vectors
+     * @param blockSize - Block size for tiling
+     */
+    blockAttention(Q: Float32Array[], K: Float32Array[], V: Float32Array[], blockSize: number): Float32Array[];
+    /**
+     * Get the speedup factor from the last benchmark
+     */
+    getSpeedup(): number;
+    /**
+     * Run benchmark comparing naive vs CPU-optimized attention
+     *
+     * @param numVectors - Number of vectors to test
+     * @param dimensions - Dimensions per vector
+     * @param iterations - Number of iterations for averaging
+     */
+    benchmark(numVectors?: number, dimensions?: number, iterations?: number): BenchmarkResult;
+    /**
+     * Get benchmark history
+     */
+    getBenchmarkHistory(): BenchmarkResult[];
+    /**
+     * Get configuration
+     */
+    getConfig(): FlashAttentionConfig;
+    /**
+     * Update configuration
+     */
+    setConfig(config: Partial<FlashAttentionConfig>): void;
+    /**
+     * Naive O(N^2) attention implementation for comparison
+     */
+    private naiveAttention;
+    /**
+     * Compute block of attention scores
+     */
+    private computeBlockScores;
+    /**
+     * Online softmax with output accumulation (key to Flash Attention)
+     *
+     * Uses the online softmax trick to maintain numerical stability
+     * while processing blocks incrementally.
+     */
+    private onlineSoftmaxAccumulate;
+    /**
+     * Compute dot product of two vectors
+     */
+    private dotProduct;
+    /**
+     * Stable softmax implementation
+     */
+    private softmax;
+    /**
+     * Generate random vectors for benchmarking
+     */
+    private generateRandomVectors;
+    /**
+     * Validate input arrays
+     */
+    private validateInputs;
+}
+/**
+ * Get singleton FlashAttention instance
+ *
+ * @param config - Optional configuration (only used on first call)
+ * @returns FlashAttention instance
+ */
+export declare function getFlashAttention(config?: Partial<FlashAttentionConfig>): FlashAttention;
+/**
+ * Reset singleton (for testing)
+ */
+export declare function resetFlashAttention(): void;
+/**
+ * Compute attention using Flash Attention
+ */
+export declare function computeAttention(queries: Float32Array[], keys: Float32Array[], values: Float32Array[], config?: Partial<FlashAttentionConfig>): AttentionResult;
+/**
+ * Run Flash Attention benchmark
+ */
+export declare function benchmarkFlashAttention(numVectors?: number, dimensions?: number, iterations?: number): BenchmarkResult;
+/**
+ * Get current speedup from last benchmark
+ */
+export declare function getFlashAttentionSpeedup(): number;
+//# sourceMappingURL=flash-attention.d.ts.map

package/dist/src/ruvector/flash-attention.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"flash-attention.d.ts","sourceRoot":"","sources":["../../../src/ruvector/flash-attention.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;GAeG;AAMH,MAAM,WAAW,oBAAoB;IACnC,6DAA6D;IAC7D,SAAS,EAAE,MAAM,CAAC;IAClB,gDAAgD;IAChD,UAAU,EAAE,MAAM,CAAC;IACnB,sCAAsC;IACtC,WAAW,EAAE,MAAM,CAAC;IACpB,+CAA+C;IAC/C,aAAa,EAAE,OAAO,CAAC;IACvB,6CAA6C;IAC7C,mBAAmB,EAAE,OAAO,CAAC;CAC9B;AAED,MAAM,WAAW,eAAe;IAC9B,qCAAqC;IACrC,MAAM,EAAE,YAAY,EAAE,CAAC;IACvB,kDAAkD;IAClD,OAAO,CAAC,EAAE,YAAY,EAAE,CAAC;IACzB,uCAAuC;IACvC,aAAa,EAAE,MAAM,CAAC;CACvB;AAED,MAAM,WAAW,eAAe;IAC9B,2CAA2C;IAC3C,WAAW,EAAE,MAAM,CAAC;IACpB,2CAA2C;IAC3C,WAAW,EAAE,MAAM,CAAC;IACpB,qCAAqC;IACrC,OAAO,EAAE,MAAM,CAAC;IAChB,oCAAoC;IACpC,UAAU,EAAE,MAAM,CAAC;IACnB,4BAA4B;IAC5B,UAAU,EAAE,MAAM,CAAC;IACnB,8CAA8C;IAC9C,gBAAgB,EAAE,MAAM,CAAC;IACzB,8CAA8C;IAC9C,gBAAgB,EAAE,MAAM,CAAC;IACzB,8BAA8B;IAC9B,eAAe,EAAE,MAAM,CAAC;CACzB;AAMD,qBAAa,cAAc;IACzB,OAAO,CAAC,MAAM,CAAuB;IACrC,OAAO,CAAC,WAAW,CAAa;IAChC,OAAO,CAAC,gBAAgB,CAAyB;IAGjD,OAAO,CAAC,WAAW,CAA6B;IAChD,OAAO,CAAC,SAAS,CAA6B;IAC9C,OAAO,CAAC,WAAW,CAA6B;gBAEpC,MAAM,GAAE,OAAO,CAAC,oBAAoB,CAAM;IActD;;;;;;;OAOG;IACH,SAAS,CACP,OAAO,EAAE,YAAY,EAAE,EACvB,IAAI,EAAE,YAAY,EAAE,EACpB,MAAM,EAAE,YAAY,EAAE,GACrB,eAAe;IA2BlB;;;;;;;;;OASG;IACH,OAAO,CAAC,qBAAqB;IAyI7B;;OAEG;IACH,OAAO,CAAC,iBAAiB;IAYzB;;;OAGG;IACH,OAAO,CAAC,WAAW;IAqCnB;;OAEG;IACH,OAAO,CAAC,WAAW;IAMnB;;OAEG;IACH,OAAO,CAAC,cAAc;IAwBtB;;;;;;;;;;;;;;OAcG;IACH,cAAc,CACZ,CAAC,EAAE,YAAY,EAAE,EACjB,CAAC,EAAE,YAAY,EAAE,EACjB,CAAC,EAAE,YAAY,EAAE,EACjB,SAAS,EAAE,MAAM,GAChB,YAAY,EAAE;IA0DjB;;OAEG;IACH,UAAU,IAAI,MAAM;IAIpB;;;;;;OAMG;IACH,SAAS,CACP,UAAU,GAAE,MAAY,EACxB,UAAU,GAAE,MAAY,EACxB,UAAU,GAAE,MAAU,GACrB,eAAe;IAsDlB;;OAEG;IACH,mBAAmB,IAAI,eAAe,EAAE;IAIxC;;OAEG;IACH,SAAS,IAAI,oBAAoB;IAIjC;;OAEG;IACH,SAAS,CAAC,MAAM,EAAE,OAAO,CAAC,oBAAoB,CAAC,GAAG,IAAI;IAQtD;;OAEG;IACH,OAAO,CAAC,cAAc;IAwCtB;;OAEG;IACH,OAAO,CAAC,kBAAkB;IAwB1B;;;;;OAKG;IACH,OAAO,CAAC,uBAAuB;IA2D/B;;OAEG;IACH,OAAO,CAAC,UAAU;IAqBlB;;OAEG;IACH,OAAO,CAAC,OAAO;IA4Bf;;OAEG;IACH,OAAO,CAAC,qBAAqB;IAyB7B;;OAEG;IACH,OAAO,CAAC,cAAc;CA+BvB;AAQD;;;;;GAKG;AACH,wBAAgB,iBAAiB,CAAC,MAAM,CAAC,EAAE,OAAO,CAAC,oBAAoB,CAAC,GAAG,cAAc,CAKxF;AAED;;GAEG;AACH,wBAAgB,mBAAmB,IAAI,IAAI,CAE1C;AAMD;;GAEG;AACH,wBAAgB,gBAAgB,CAC9B,OAAO,EAAE,YAAY,EAAE,EACvB,IAAI,EAAE,YAAY,EAAE,EACpB,MAAM,EAAE,YAAY,EAAE,EACtB,MAAM,CAAC,EAAE,OAAO,CAAC,oBAAoB,CAAC,GACrC,eAAe,CAGjB;AAED;;GAEG;AACH,wBAAgB,uBAAuB,CACrC,UAAU,CAAC,EAAE,MAAM,EACnB,UAAU,CAAC,EAAE,MAAM,EACnB,UAAU,CAAC,EAAE,MAAM,GAClB,eAAe,CAEjB;AAED;;GAEG;AACH,wBAAgB,wBAAwB,IAAI,MAAM,CAEjD"}