Hierarchos-Native v0.2.0: Rust+Vulkan LLM training adds RDNA3 WMMA and native multimodal training
PhysicsDisastrous462 · reddit · 2026-11-17
A solo developer released Hierarchos-Native v0.2.0, a Rust + Vulkan framework for training, fine-tuning, and running LLMs without Python, PyTorch, CUDA, ROCm, or oneAPI.
Key updates
- RDNA3 WMMA acceleration: opt-in Vulkan cooperative-matrix path for AMD RDNA3 with FP16 operands and FP32 accumulation; master weights, gradients, AdamW moments, and normalization/softmax/loss all stay FP32. Isolated GEMM speedups on Ryzen Z1 Extreme: QKV 2.17×, FFN up 2.12×, FFN down 2.31×, vocab projection 1.83×
- Honest perf caveats: small matrices run slower due to conversion overhead, so a profitability policy keeps small ops on FP32 kernels; end-to-end training was actually slower in the 512-row fixture (14.04 ms vs 12.29 ms), so Mixed WMMA is opt-in and disabled by default — the author prioritizes real speedups over misleading kernel numbers
- Native multimodal training: RGB preprocessing, CLIP/LLaVA projector integration, SigLIP/Gemma 3 vision training, Gemma 4 vision, a dedicated DeepSeek V4.1 image-training path, with backprop through the image tower and projector — actual training, not frozen-model embedding passthrough
Open source at github.com/necat101/Hierarchos-Native with Windows portable builds and a WMMA guide.
More from Infra
- How rollout staleness hurts on-policy distillation, and how to build a faster pipeline — burkov · 2026-11-17
- SK hynix's Nvidia revenue doubled to $24B, but US customer concentration is the real risk — tengyanAI · 2026-11-17
- macOS TCP Stack Breaks After Exactly 49.7 Days of Uptime Due to 32-bit Timestamp Overflow — DanielLockyer · 2026-11-17
- Running a 2M-param embedding model on an ESP32 with no_std Rust at 830 embeddings/s — qdrant_engine · 2026-11-17
- Voice agent failed 0/5 turns in background noise; adaptive client-side VAD gate fixed it at lower latency — shrapknife · 2026-11-17
- Strix Halo int4×int4 WMMA: Gemma 4 31B at 2.43× llama.cpp with matched quality — No-Dragonfly6246 · 2026-11-17