MCP Titan

import * as tf from '@tensorflow/tfjs-node'; export interface CompressedGradientPayload { deltas: Float32Array[]; shapes: number[][]; } export class DeltaCompressionHook { private previous: tf.Tensor[] = []; public compress(gradients: tf.Tensor[]): CompressedGradientPayload { const deltas: Float32Array[] = []; const shapes: number[][] = []; gradients.forEach((gradient, index) => { const previous = this.previous[index]; const previousValues = previous ? previous.dataSync() : new Float32Array(gradient.size).fill(0); const currentValues = gradient.dataSync(); const delta = new Float32Array(currentValues.length); for (let i = 0; i < currentValues.length; i += 1) { delta[i] = currentValues[i] - previousValues[i]; } deltas.push(delta); shapes.push(gradient.shape); }); this.previous = gradients.map(tensor => tensor.clone()); return { deltas, shapes }; } public decompress(payload: CompressedGradientPayload): tf.Tensor[] { return payload.deltas.map((delta, index) => tf.tensor(delta, payload.shapes[index]) ); } public reset(): void { this.previous.forEach(tensor => tensor.dispose()); this.previous = []; } } export interface LayerScheduleConfig { maxActiveLayers: number; } export class LayerScheduler { private readonly config: LayerScheduleConfig; constructor(config: LayerScheduleConfig) { this.config = config; } public selectActiveLayers(gradients: tf.Tensor[]): number[] { const scored = gradients.map((tensor, index) => ({ index, score: Math.abs(tensor.sum().arraySync() as number) })); scored.sort((a, b) => b.score - a.score); return scored.slice(0, this.config.maxActiveLayers).map(item => item.index); } } export class UpdateBuffer { private buffer = new Map<string, tf.Tensor>(); public push(layerName: string, update: tf.Tensor): void { const existing = this.buffer.get(layerName); if (existing) { const combined = tf.add(existing, update); existing.dispose(); update.dispose(); this.buffer.set(layerName, combined); } else { this.buffer.set(layerName, update.clone()); } } public flush(): Map<string, tf.Tensor> { const snapshot = new Map<string, tf.Tensor>(this.buffer); this.buffer.forEach(tensor => tensor.dispose()); this.buffer.clear(); return snapshot; } public clear(): void { this.buffer.forEach(tensor => tensor.dispose()); this.buffer.clear(); } }