Asynchronous Self-Healing KV Cache for Silicon-Native LLMs
Project description
ASH-KV: The Self-Healing Middleware for LLMs
ASH-KV is a high-performance, hardware-aware middleware layer designed to provide Runtime Integrity for Large Language Models. By surgically intercepting and correcting the KV cache at the silicon level, it prevents logical drift and clinical hallucinations with zero detectable latency.
🏛️ Core Value Pillars
⚡ Zero-Latency Integrity
Surgical KV cache mutation at Metal (Apple Silicon) and CUDA (NVIDIA) speeds. Our Fused Kernels ensure that the "Immune System" adds virtually 0% overhead to inference throughput.
🔌 Hardware Agnostic (Universal HAL)
The Hardware Abstraction Layer (HAL) automatically detects your silicon and hot-swaps between MLX and PyTorch backends. The same code runs on an M4 MacBook or an NVIDIA H100 server.
🛡️ Adaptive Shielding & Real-Time Healing
Autonomous sensitivity scaling via the AdaptiveSensitivity Agent. Integrated with a Deterministic Clinical Rules Engine (DCRE), ASH-KV monitors token generation in real-time and prunes attention heads the microsecond a contraindication is detected.
♾️ Infinite Horizon (NVMe Paging)
Break the VRAM ceiling. ASH-KV dynamically offloads "Cold" context chunks to NVMe storage, allowing for 100k+ token windows on consumer-grade hardware without OOM crashes.
🚀 Quick Start
1. Installation
pip install .
2. Corporate Integration (3 Lines of Code)
Integrate ASH-KV into any production pipeline to add an immediate safety layer.
from mlx_ash_kv.api import protect
# Wrap your existing model with the ASH-KV shield
protected_model, cache, shield, proxies = protect(model, sensitivity=0.85)
# Inference continues normally, but with real-time surgical healing
🛠️ Command Center (CLI)
ASH-KV comes with a professional CLI for systems verification and benchmarking.
ash-kv install: Verify hardware drivers, silicon backend, and NVMe Paging Stress Test.ash-kv benchmark: Run the 100-case "Hard Truth" evaluation suite.ash-kv monitor: Launch the Live Diagnostic TUI to see layer-wise health and [HOT/WARM] memory distribution.ash-kv demo: Launch the Gradio B2B Reliability Playground.
🔬 Scientific Foundation
ASH-KV implements Asynchronous Self-Healing protocols that offload hallucination detection to secondary silicon (like the ANE or secondary GPU cores), ensuring the main generation loop remains unobstructed.
⚠️ DISCLAIMER
ASH-KV is a hardware-level reliability layer designed to assist professionals. It is NOT a substitute for professional medical or legal judgment. All AI-generated outputs, even those "healed" by ASH-KV, must be verified by qualified human professionals before making clinical or legal decisions.
Built for the future of mission-critical Agentic Reasoning.
Project details
Release history Release notifications | RSS feed
Download files
Download the file for your platform. If you're not sure which to choose, learn more about installing packages.
Source Distribution
Built Distribution
Filter files by name, interpreter, ABI, and platform.
If you're not sure about the file name format, learn more about wheel file names.
Copy a direct link to the current filters
File details
Details for the file mlx_ash_kv-8.2.0.tar.gz.
File metadata
- Download URL: mlx_ash_kv-8.2.0.tar.gz
- Upload date:
- Size: 14.5 kB
- Tags: Source
- Uploaded using Trusted Publishing? No
- Uploaded via: twine/6.2.0 CPython/3.11.14
File hashes
| Algorithm | Hash digest | |
|---|---|---|
| SHA256 |
dae8b3f492e9a2defa9f9b8ecd07b12bb9ad4386f45ef07e98ade5db24d245c2
|
|
| MD5 |
f56b3c245b5fd5a3621a442746fd1f81
|
|
| BLAKE2b-256 |
76b4edae68f4f0a200a3bb397cdb90ece9afa106274718faf3b7e3c8d68aba8d
|
File details
Details for the file mlx_ash_kv-8.2.0-py3-none-any.whl.
File metadata
- Download URL: mlx_ash_kv-8.2.0-py3-none-any.whl
- Upload date:
- Size: 16.2 kB
- Tags: Python 3
- Uploaded using Trusted Publishing? No
- Uploaded via: twine/6.2.0 CPython/3.11.14
File hashes
| Algorithm | Hash digest | |
|---|---|---|
| SHA256 |
0fdfc44fbcec2ba713394099a16d62227ad2839a0c9c9b5be50262fea135bd22
|
|
| MD5 |
8bd22930b419b34022e80355ac08955e
|
|
| BLAKE2b-256 |
d44fe0dc0ae94459adf782d7451a0c00319516f20244f39a6b944f11fef5979a
|