Running AI training on unsupported AMD MI50s: patched ROCm and a Vulkan training stack
Savantskie1 · reddit · 2026-09-19
A user with dual AMD MI50 32GB cards (64GB HBM2 total) documents reviving deep learning workloads on gfx906, which AMD dropped after ROCm 5.7, and disproving the consensus that Vulkan training is impossible.
Highlights:
- The blocker in ROCm 6.4.3 is just the precompiled TensileLibrary lacking gfx906 kernels. By repacking the missing kernel files, both MI50s came up fully under ROCm 6.4.3, with PyTorch and llama.cpp working natively
- For vLLM: ROCm 7.x hit AMD bug #5653, the nlzy fork was blocked by a missing flash-attn V1 dependency, but a Docker image (ROCm 6.3.4 + PyTorch 2.11, precompiled flash-attn) got single-GPU inference working; dual-GPU tensor parallel is only blocked by a PCIe topology issue, fixable with a PLX switch
- The centerpiece: building a Vulkan training stack from scratch. Every AI assistant and forum claimed backward-pass infrastructure only exists in ROCm/CUDA, yet the author ran training over Vulkan (same path llama.cpp uses for inference) — while ChatGPT was insisting it was impossible
- The post doubles as a critique of AI assistants confidently repeating the same wrong 'not supported, upgrade your GPU' answer for months
More from Infra
- Seedance 2.5 video model now available on US-hosted infrastructure via fal — jfischoff · 2026-09-19
- SpaceX CFO says on track to deploy 100 GW of orbital compute per year by 2030 — NicoVerderosa · 2026-09-19
- YuE2 BF16 runs on 12GB VRAM and sounds better, plus the fix for pagefile thrashing — lazyspock · 2026-09-19
- Inco Splash hits 144 tok/s on Qwen3.8-27B M5 Max, 3x faster than Ollama — ResearchCrafty1804 · 2026-09-19
- macOS 27 ships with mlx_whisper built in, letting agents transcribe video locally — vista8 · 2026-09-19
- Ternary Bonsai 2 27B at 1.75bpw fits an 8GB GPU, hits 93.3% accuracy in audiobook speaker-attribution test — autonoma_2042 · 2026-09-19