First-run setup

Pick a path

Ubuntu 24.04+ (recommended)

Ubuntu 24.04.5 is on the ROCm 10.1 Radeon OS row. This is the path that produced the official 35B llama.cpp cards. Dual-boot or a spare drive beats forcing Windows on day one.

1. Step 1 · measured

   Hardware check

   Confirm the SKU. The measured house box is one AMD Radeon AI PRO R9700 (32 GB, gfx1201). There is no consumer RX 9700. A 16 GB RX 9070 XT is the same generation with half the VRAM. Seat power, then read BIOS notes (Above 4G, ReBAR/SAM, CSM off, matching lane widths) before you blame ROCm. The per-board BIOS table is empty until submissions.

   • GPU directory
   • R9700
   • Hardware
   • BIOS

2. Step 2 · measured

   Drivers and ROCm

   Do not fork a second install guide. Follow Install ROCm, then Verify GPU and ROCm. Pin one AMD Core SDK (10.1 line as of the weekly card). Do not apt install rocm from Ubuntu’s archive. 24.04.5 is an official 10.1.0 Radeon row. Vulkan llama.cpp still works if HIP configure is a mess.

   • Install ROCm
   • Verify ROCm
   • What is ROCm
   • Weekly card

3. Step 3 · measured

   Pick a backend — llama.cpp first

   Consumer default: llama.cpp Vulkan for the daily-driver 35B. Strata v0.1.42 is second, and only for Qwen 3.8 Flash. vLLM is a multi-client try, not a first token. HIP is the ROCm compile when you need it. Engine notes stay on the backends pages.

   • Backends
   • llama.cpp
   • Vulkan vs ROCm
   • Strata
   • AI software

4. Step 4 · measured

   Pick a measured recipe

   Daily driver: Qwen3.6-35B-A3B Unsloth UD-Q5_K_XL on one R9700, llama.cpp Vulkan, 262144 context, f16 KV. Fast agent: Qwen 3.8 Flash UD-Q6_K_XL on Strata. Two cards: read Dual / multi-GPU, then 35B on two cards before you split. Q3 is a labeled estimate unless a card is measured. Q1/Q2 stay off. Open the recipe, then /qwen and /flash-daily for the rest of the table.

   • 35B recipe
   • Flash recipe
   • Qwen compare
   • Flash daily
   • Dual / multi-GPU
   • Multi-GPU how-to
   • Two R9700s
   • Stacks

5. Step 5 · measured

   First run

   Run a local Qwen walks llama.cpp Vulkan on one R9700. Overnight cards recorded llama-bench, not llama-server: 102.39 tok/s decode at 4K, 90.52 at 32K, 66.43 at 128K, 44.4 at 256K. Peak VRAM total 25.97 GB at 4K and 29.34 GB at 256K; peak system RAM 37.32 GB and 42.9 GB (Linux GB). Then Serve Qwen if you want a loopback /v1. Keep the GGUF on NVMe.

   • Run a local Qwen
   • Serve Qwen
   • 4K run card
   • Connect a harness

Windows

Cited Windows guide at /setup/windows. Vulkan llama.cpp is the honest first token. Full ROCm userspace is a Linux product. ROCm 10.1 lists WSL2 as a technical preview — cited walk at /setup/wsl. local desktop cards are Ubuntu — we do not have Windows or WSL tok/s.

1. Step 1 · estimate

   Hardware check

   Same SKU check as Ubuntu. The measured speeds on AMDai are still the Ubuntu local desktop cards. BIOS notes apply on the same board. We do not invent a Windows-only motherboard list.

   • GPU directory
   • BIOS
   • Hardware

2. Step 2 · estimate

   Vulkan first — ROCm is not the first-run path

   Build or install llama.cpp with Vulkan. That is the labeled estimate for a first Windows token. Native Windows ROCm and PyTorch-on-Windows exist as a narrower product; do not expect rocm-smi culture. WSL2 ROCm 10.1 is a tech preview at /setup/wsl, not a measured daily driver. Prefer a spare-drive Ubuntu install if you want the cards on this site to match your box.

   • Windows guide
   • WSL2 preview
   • llama.cpp
   • Windows / WSL2 preview
   • Vulkan vs HIP
   • Vulkan vs ROCm compare
   • Install ROCm (Linux)

3. Step 3 · estimate

   Backend pick (estimate)

   llama.cpp Vulkan first. Strata and HIP/ROCm userspace are the Linux story we measured. We do not have a Windows Strata card. vLLM on Windows is not a first-run recommendation.

   • Backends
   • Vulkan vs ROCm
   • AI software

4. Step 4 · estimate

   Same files, Linux-measured speeds

   You can download the same GGUF the recipes name. The tok/s, VRAM, and RAM figures on /benchmarks and /results are Ubuntu llama-bench on the local desktop. Treat them as a comparison, not a Windows promise. Q3 stays a labeled estimate. Q1/Q2 stay off.

   • 35B recipe
   • Qwen compare
   • Flash daily
   • Dual / multi-GPU
   • Multi-GPU how-to
   • Two R9700s

5. Step 5 · estimate

   First token (estimate)

   Run official instruct in llama.cpp Vulkan and confirm Task Manager / GPU-Z shows the Radeon busy. We do not publish a Windows launch line or a Windows tok/s. After tokens flow, read the Linux cards so you know what “good” looked like on Ubuntu. A harness comes after loopback /v1 — that attach is also unmeasured on Windows here.

   • Run a local Qwen
   • Benchmarks
   • FAQ

Ubuntu-based distros

Linux Mint, Pop!_OS, and Kubuntu are Ubuntu-based. AMD’s ROCm 10.1.0 compatibility matrix lists Ubuntu 26.04.1, 24.04.5, and 22.04.5 by name — not Mint, Pop!_OS, or Kubuntu. Official docs do not put those derivatives on the 10.1 OS row, so we do not invent official AMD support. Follow the Ubuntu steps on this page only as unofficial; that path is not verified on our rig. Prefer stock Ubuntu 24.04.5 if you want the measured local desktop cards to match the box. Fedora, Arch, Docker / Podman, LM Studio, the coding-assistant guide, Ollama, Lemonade, Open WebUI, WSL2, ComfyUI, the multi-GPU llama.cpp how-to, and the always-on systemd server have their own cited pages.

• Fedora setup
• Arch setup
• Windows setup
• Docker / Podman
• LM Studio
• Coding assistant
• Ollama
• Lemonade
• Open WebUI
• WSL2
• ComfyUI
• Multi-GPU llama.cpp
• Always-on server

When to use Radeon Stack Lab

Open Stack Lab after the first token if you have the Origin / local repo. It is the Ubuntu companion that records amdai-bench/run@1 cards instead of hand-wiring every flag. There is no public download URL — packages are not published yet. Do not invent a release. Profiles follow the YAML: qwen-reliable is the 35B daily driver; qwen-fast-agent is Flash on Strata. Lab Notes ingests those cards onto /results.

• Stack Lab
• Run cards
• Contribute a recipe

How to read speed, VRAM, and RAM

llama-bench prints pp (prefill) and tg (decode). Decode is the chat tok/s. /benchmarks is the 33-card overnight matrix; each cell links to /results. The calculator prints two RAM lines: “Minimum system RAM needed” (the load at the context you picked) and “RAM needed at full context” (262144 for the house 35B). Overnight UD-Q5 one-GPU peaks are 37.32 GB at 4K and 42.9 GB at 256K. Those are measured peaks, pending re-verify after the RAM fix — not a kit you must buy.

• Benchmarks
• Run cards
• Calculator
• Compare 35B
• Flash RAM offload

When it breaks

Do not duplicate those pages here. Start from a symptom on /troubleshoot. Paste a log in the error matcher. Cited rows live on /rocm-problems (HIP no binary, archive shadow, MODE1 reset, unofficial HSA overrides). Half-upgraded ROCm: /rocm/upgrade. Firmware misses: /bios. Common failures: /troubleshooting.

• Troubleshoot
• ROCm problems
• Upgrade or reinstall ROCm
• BIOS
• Troubleshooting
• Error paste
• FAQ