ACAE-1B-Foundation

This is the 1 Billion Parameter Foundational Model trained natively in C++ using LibTorch on an AMD MI300X cluster.

Architecture

• Size: 1 Billion Parameters (dim=1024, layers=24, heads=16)
• Engine: Custom C++ ATLLM Engine
• Hardware: AMD Instinct MI300X (ROCm/HIP)
• Training Data: 20 Billion Tokens (Physics Wikipedia, MMLU, ArXiv)

Usage

This model is serialized in raw PyTorch C++ .pt format. It must be loaded using the C++ torch::load mechanism inside the ACAE inference engine.