Version Archive & Changelog

Changelog history and immutable releases

v1.3.13 - OpenCL Acceleration Engine & Unified 5-Backend Standard

Production Release (Latest) 2026-09-29
  • Unified 5-Backend Standard: Standardized all CLI subcommands (run, serve, benchmark) to enforce the unified backend choices: auto, gpu, vulkan, opencl, cpu.
  • Qualcomm Adreno OpenCL Acceleration: Added native opencl backend routing via optimized OpenCL 2.0 kernels, successfully resolving driver-level SPIR-V Vulkan defect on Adreno 650 (Snapdragon 865).
  • Installer Symlink Hardening: Purged conflicting symlinks to prevent unintentional overwrite of native binary assets.
  • Fleet Verification: Empirically verified on Galaxy S20 (OpenCL offload) and Galaxy S25 (Vulkan/Flash Attention).

v1.3.12 - Dynamic Linker Bionic Isolation & ChatML Repetition Mitigation

Stable Release 2026-09-29
  • Dynamic Linker Bionic Isolation: Purged regressive $PREFIX/lib injection from fallback runtime environment, guaranteeing Android 15/16 Bionic namespace isolation and zero libunwindstack symbol collisions.
  • ChatML Auto-Templating: Standardized <|im_start|> prompt encapsulation and automatic -r '<|im_end|>' reverse stop token injection, eliminating repetition loops on Qwen and Llama-3.
  • Cross-SoC Mobile Inference Stability: Retained verified -fa 0 defense on Adreno 600 while maintaining Flash Attention on Adreno 830 and ARM Mali GPUs.

v1.3.10 - Dynamic Network Resume & Qualcomm Adreno Flash Attention Defense

Stable Release 2026-09-28
  • Dynamic 5-Stage Network Resume: Exponential backoff with HTTP Range 206/416 self-healing for interrupted GGUF downloads.
  • Adreno Flash Attention Defense: Injected -fa 0 by default to prevent Qualcomm driver assertion crashes during mobile GPU inference.
  • Pure CPU Device Isolation: Set GGML_VK_VISIBLE_DEVICES='' under device='cpu' to prevent buggy vendor Vulkan driver crashes.
  • Chat Template Auto-Resolution: Automatic chat template expansion for Qwen and Llama-3 models with stop token cleanup.
  • Official ameva-runtime Adapter Integration: Deep binding with LlamaCppAdapter for Bionic HAL orchestration.

v1.2.0 - Unified AMEVA Vulkan HAL & Galaxy A35 Validation

Stable Release 2026-09-01
  • Unified AMEVA Vulkan HAL Integration: Deep binding with ameva-vulkan-runtime v1.1.0 utilizing Android Bionic Vulkan ICD (/system/lib64/libvulkan.so).
  • Strict 3-Tier Execution Mode: Added --device vulkan (Fail-Fast GPU compute), --device auto (transparent CPU recovery), and --device cpu (0ms direct NEON forward pass).
  • Dynamic Octa-Core Thread Tuning: Automatic detection of ARM big.LITTLE architectures (Exynos 1380, Snapdragon) injecting optimal big-core threads (-t 4).
  • Dual Package Distribution: Synchronously released to Python PyPI (termux-llamacpp v1.2.0) and Node.js npm (termux-llamacpp v1.2.0).
  • Real-Device Galaxy A35 Validation: End-to-end verification across CLI, Python SDK, Node.js npm, and background OpenAI REST server daemon.
  • Supply-Chain Receipt Gatekeeper: Added verified build receipt support (llama-server.build-receipt.json) for local native binaries.

v1.1.1 - Supply-Chain Cryptographic Verification & Multi-Channel Crawler

Stable Release 2026-08-30
  • Ed25519 Public Key Verification: Cryptographic manifest verification with SHA-256 binary validation.
  • Multi-Channel HuggingFace GGUF Crawler: Deep model discovery and automated quant downloader.
  • Process Supervisor Resilience: Guaranteed zero-zombie process termination with tracked PID ledger.

v1.0.2 - Bionic Dynamic Linker Stabilization & Dual-Ecosystem Publishing

Stable Release 2026-08-28
  • Bionic Shared Library Bundling: Packaged libgomp and libstdc++ to prevent missing symbol errors on raw Termux.
  • NPM Binary Wrapper: Added standalone ESM binary resolver for Node.js environments.

v1.0.1 - Initial Production GGUF LLM Runtime Release

Initial Release 2026-08-27
  • Prebuilt ARM64 llama.cpp Native Binaries: Direct extraction in under 3 seconds.
  • Loopback Isolated OpenAI Server: /v1/chat/completions REST and SSE streaming supervisor.