torchquery 2.2.0

High-performance SDC detection and neural healing for billion-scale tensors.

TorchQuery 🛡️

---

TorchQuery is a high-performance reliability engine for PyTorch. It provides a "Neural Shield" against Silent Data Corruption (SDC), hardware bit-flips, and numerical instability in massive Deep Learning models.

🚀 Key Features

• Billion-Scale Protection: Optimized streaming logic designed to handle tensors with $10^9$ elements without crashing.
• Neural Healing: Automatically detects and repairs corrupted weights or neurons using statistical outlier detection ($\sigma$-clamping).
• Distributed SyncBatch: Cluster-aware protection using All-Reduce to ensure safety across multi-GPU and multi-server environments.
• Zero-Invasive: Simply wrap your existing tensors or model parameters; no architecture changes required.

---

Visualizing Silent Data Corruption (SDC)

Hardware glitches—like cosmic rays or VRAM overclocks—can cause random bit-flips. These create massive statistical outliers or NaNs in your tensor data.

[Image Link to Image_5.png]

TorchQuery acts as a Neural Shield that sweeps your multidimensional arrays. It identifies values that can lead to exploding gradients (3e38) or numerical instability (NaN), "healing" them before they propagate.

Pre-Sweep State:

• NaN (Not a Number): Corrupts entire model during backpropagation.
• 3e38: Causes exploding gradients, destroying training stability.

Post-Sweep State:

• Invalid data is removed, leaving behind validated tensor values.