AMEVA-Vulkan-Runtime

Unified Cross-Modal Vulkan GPU Acceleration Runtime & HAL for Mobile Android

PyPI Version npm Version License Tests Platform
1-Line Quick Installation

Install the official package directly into your runtime:

pip install ameva-vulkan-runtime
# or: npm install ameva-vulkan-runtime

The Engineering Challenge

Running multi-modal AI on mobile Android is plagued by fragmented GPU drivers, loader crashes between Bionic and Mesa, tensor alignment buffer overflows, and redundant binary bloat across individual packages.

The Architectural Breakthrough

Provides a single, zero-hardcoded C++20 Vulkan Hardware Abstraction Layer (HAL) and universal runtime for STT, Vision, LLM, Diffusion, and Training with a granular 12-stage validation hierarchy (V0-V11) and zero-data-loss auto-recovery.

Key Capabilities & Built-in Hardening

Single Loader Chain Pinning

Eliminates Termux Mesa vs Android Bionic symbol collisions via dynamic dladdr provenance and sys_gpdf2 isolation.

12-Stage Probing & Fallback

Validates GPU capability from dlopen (V0) to E2E model inference (V11) with transparent CPU NEON recovery.

Multi-Modal Cross-Acceleration

Powers Whisper STT, LLaVA Vision, LLaMA/BitNet LLM, and Stable Diffusion from a single 58MB shared core.

Supported Compute Kernels & Operations

Subsystem Category Operations & Kernels Status
Supported GPUs Qualcomm Adreno 830/700/600, ARM Mali-G78/G68/G77, Samsung Xclipse 950/920 Production
Validation Hierarchy V0 (Loader) ~ V11 (E2E Diffusion / Whisper / LLaMA Inference) Production
Zero-Drift Quirks Fixes Adreno Subgroup Control Bypass, Mali 128-byte Strict Memory Alignment Production

Canonical Usage Example

import ameva_vulkan_runtime as avr

# 1. Probe & Validate Hardware (V0-V11)
doctor = avr.Doctor()
report = doctor.run_self_test()
print(f"GPU Backend Status: {report.status}, Device: {report.device_name}")

# 2. Acquire High-Performance Vulkan Context for STT / LLM / Diffusion
ctx = avr.create_context(device="auto", memory_limit_mb=1024)
print(f"Context initialized via: {ctx.loader_path} (API {ctx.vulkan_version})")

Getting Started & Deep Guides