gfx906

llama.cpp-gfx906

Open-source local AI server configs, GFX906 runtime maintenance, reproducible benchmarks, and QC methods for affordable AI research infrastructure.

FlashAttention-style custom attention backend for vLLM on AMD MI50/MI60/Radeon VII (gfx906). Downstream fork of mixa3607/ML-gfx906 with replacement HIP kernels and a vllm.general_plugins entry point.

AMD Instinct MI50/MI60 (gfx906, HIP/ROCm) port of NInfer - the specialized single-GPU Qwen inference engine. First AMD target in the ecosystem. Audit complete, port in progress.

Deploy Qwen3.6-27B GGUF on an AMD MI50 (gfx906) eGPU via vLLM + ROCm — docker-compose bundle with patches for Qwen3.5/3.6 hybrid Gated-DeltaNet support.

ROCm/Unsloth/bitsandbytes 4-bit lab and VRAM benchmarks for AMD MI50/gfx906 LLM fine-tuning

openPangu Flash92 — C++/CUDA & C++/HIP ROCm Engines. Fully resident, Blackwell tensor cores, no Python. Powered by openPangu.

Restore videos with pixelated/mosaic regions rocm_gfx906 (mi50 gpu)

Run ComfyUI on AMD Radeon VII (gfx906) via Docker. ROCm 5.7 + PyTorch 2.3.1, SDXL 1024×1024 in ~28s/image. Pinned to ComfyUI v0.3.60 — the last gfx906-compatible build.

Benchmarks and runnable setup for Qwen LLMs on AMD Radeon VII (gfx906) via llama.cpp + ROCm in Docker

Qwen 3.ocho (C++): stochastic trajectory rendering with a decision/commit loop on a native CUDA/HIP/ROCm inference engine (NVIDIA GB10 + 4x AMD MI50)

LLM inference in C/C++, gfx906 qwen3.8 optimisations

Custom C++/HIP inference engine for Qwen3.8-27B on 4x AMD MI50 (gfx906/ROCm). Written from scratch - not llama.cpp, not a wrapper, no Python in the execution path. 103.7 tok/s with chained MTP speculative decoding.

Run llama.cpp on Vega 64 / gfx900 and Radeon VII / MI50 / gfx906 with ROCm 7.2 — including mixed-GPU inference.