AMD Strix Halo / Ryzen AI Halo local LLM setup and benchmark guide for Ryzen AI MAX+ 395 and Radeon 8060S: Ollama, llama.cpp Vulkan/RADV, ROCm, 101 t/s Qwen3-Coder, CHADROCK MTP, 120B GGUF, and raw evidence.

vLLM Qwen 3.6-27B (AWQ-INT4) + DFlash speculative decoding on AMD Strix Halo (gfx1151 iGPU, 128 GB UMA, ROCm 7.13). 24.8 t/s single-stream, vision, tool calling, 256K context, OpenAI-compatible, Docker. Matches DGX Spark FP8+DFlash+MTP at a third of the cost. No CUDA.

SGLang Docker image for AMD Strix Halo (gfx1151 / Radeon 8060S)

Docker Compose for llama.cpp GGUF servers on AMD Strix Halo: Qwen, Gemma, and Laguna packages (abliterated and quantized), stock Vulkan plus ROCmFP4/MTP and ROCmFPX, parallel slots, with prefill/decode and quality metrics measured on this rig.

ComfyUI with Flash Attention for AI+ MAX 395 (gfx1151)

vLLM + Qwen3.6-27B (BF16) OpenAI-compatible inference server on AMD Strix Halo (Ryzen AI Max+ 395, gfx1151). Vision input, 256K context, /v1/responses with separated reasoning, via TheRock ROCm.

llama.cpp + Qwen3.6-27B (Q8_0 GGUF) OpenAI-compatible inference server on AMD Strix Halo (Ryzen AI Max+ 395, gfx1151). 256K context, ~7.5 t/s decode via TheRock ROCm Docker.

Claude Code skill for AMD Strix Halo (Ryzen AI MAX+ 395) ML setup. Handles PyTorch installation (official wheels don't work with gfx1151), GTT memory config, and environment setup. Enables 30B parameter models.

A turnkey, fully-local AI workstation engineered for the AMD Ryzen AI Max+ 395. LLM inference, voice, document parsing, browser automation, agents — all on-device.

Run the official Kimi K3 MoE checkpoint on one 128 GB AMD Strix Halo box. ROCm-resident static weights, MXFP4 experts streamed from NVMe via io_uring. C engine, chat client, OpenAI-compatible server. Very experimental.

ComfyUI on AMD Strix Halo (RDNA 3.5 / gfx1151) via Docker. Ubuntu 26.04 LTS + uv-managed Python 3.12 + pinned TheRock ROCm 7.13 wheels. Fixes the silent CPU fallback Debian / Python 3.13 images hit on gfx1151.

Drop-in amdsmi for AMD GPUs on WSL2 (DirectX/dxgkrnl passthrough) where the native amdsmi is unavailable — unblocks vLLM, PyTorch ROCm, GPU monitoring/tooling, etc.

DeepSeek V4 Flash 284B on AMD Strix Halo (gfx1151) — up to 32 tok/s decode & ~250 tok/s prefill via ROCmFPX, DSpark & ROCm 7.2

Optimized dual AMD Strix Halo (gfx1151) vLLM MoE inference: TP=2 over USB4, tuned int4 MoE kernel, ~8us interconnect, serialized serving

Unlock fast, local LLM inference on AMD-powered mini PCs delivering 65-87 t/s for large models without cloud or subscription costs

llama.cpp HIP vs Vulkan benchmark on AMD Strix Halo (gfx1151) with ROCm 7.2.2 — Ubuntu 26.04

Docker infrastructure for AMD Strix Halo (RDNA 3.5 / gfx1151): PyTorch + ROCm base container and a separate Ollama LLM service. Two folders, two Compose files, one Strix Halo box.

Local text to textured GLB on an AMD Strix Halo iGPU (gfx1151): FLUX.2 klein, a Vulkan-only TRELLIS.2 engine, and humanoid auto-rigging with SkinTokens on ROCm. No Blender, no CUDA.

ROCmFPX llama.cpp fork for Windows 🏆 — native build, headless OpenAI-compatible server & benchmarks. Tested on AMD Strix Halo (gfx1151), runs on other GPUs too.

ComfyUI-Crystools for ROCM RDNA3.5 STRIXHALO