Version Archive & Changelog
Changelog history and immutable releases
v2.0.1 - Official AMEVA Cluster Virtual RAM Integration & Standalone Manual
Production Stable Release
2026-10-04
- Enforced explicit AMEVA Cluster (pip install ameva-cluster) control plane requirement for multi-device distributed memory pooling.
- Published complete step-by-step AMEVA Cluster deployment and 44GB virtual RAM execution manual.
- Integrated MasterTunnel cryptographic loopback security for inter-device tensor pipelines.
- Added cluster-virtual-ram documentation portal page with full Zero-Drift compliance.
- Bumped production package version to 2.0.1 across PyPI and npm registries.
v2.0.0 - Major Multimodal Engine & 44GB Disaggregated Cluster Virtual RAM Release
Production Major Release
2026-10-03
- Added 6-Modality Target-Aware LoRA Training Hub (LLM, Diffusion, Vision VLM, Whisper STT, TTS, BitNet 1.58-bit).
- Introduced Single-Device GPU Slicing Architecture (--vocab-slice, --chunk-layers, --stream-layers) preventing mobile watchdog timeouts.
- Implemented 44GB Disaggregated Cluster Virtual RAM Pooling across heterogeneous mobile devices with 300MB Guard-Band protection.
- Integrated client-side On-Device Reinforcement Learning (GRPO, DPO, PPO).
- Delivered 100% Dual-Engine Parity across Python CLI, Node.js CLI, Python SDK, and Node.js SDK.
- Zero-conversion SafeTensors adapter export directly consumable by ComfyUI, llama.cpp, termux-vision, termux-stt, termux-tts, and bitnet.cpp.
v1.1.6 - Vulkan Autograd & OpenCL Kernel Hardening
Stable Release
2026-09-28
- Implemented Vulkan compute shader forward-backward pass kernels for ARM Mali and Adreno.
- Added SHA-256 persistent disk caching for intermediate autograd tensors.
- Fixed Bionic ARM64 thread synchronization race condition.
v1.0.0 - Initial Open Source Foundation Release
Initial Release
2026-09-01
- Dependency-free C-vectorized DAG Autograd runtime.
- Basic TinyTransformerLM and LoRA adapter fine-tuning.
- SafeTensors file format writer.