Open-Source Rust+Vulkan Training Backend Supports 143 Modern Transformer Architectures Without CUDA

PhysicsDisastrous462 · reddit · 2026-09-17

An open-source project, Hierarchos Native, implements a training and inference backend in Rust + Vulkan compute, no CUDA or PyTorch required, covering 143 canonical Transformer architectures (224 model-type spellings with HF aliases). It natively implements forward/backward passes, AdamW, LoRA/PEFT, attention + KV-cache, dense and MoE, mixed precision, checkpointing, SafeTensors, and HF model/tokenizer/dataset loading, with architecture-specific Vulkan kernels for families like Llama, Qwen, DeepSeek, Mistral, and BERT. The author stresses it's not yet a CUDA replacement and is untested broadly; development ran on AMD RDNA 3 including a ROG Ally handheld. The bigger point: Vulkan runs across NVIDIA/AMD/Intel, so mature native stacks could loosen the single-vendor compute ecosystem lock-in.

Related event: Rust+Vulkan Training Backend Supports 143 Transformer Architectures(2 posts)→

Original post →

More from Infra

Infra channel →