Fedora first-run setup

Two tracks

Zero to first token

1. 1 · Cited, not verified on our rig

   Supported release and kernel

   AMD’s 10.1 matrix does not list Fedora. The Fedora HC SIG says the amdgpu driver is built with the kernel, so you do not pin an extra AMD kernel module. Prefer a current Fedora release the SIG documents (42 / 43 / Rawhide gfx lists include gfx1201). Confirm `uname -r` is the stock Fedora kernel. A custom kernel is not verified on our rig. Dual-boot Ubuntu 24.04.5 if you want the measured local desktop path.

   • Ubuntu setup (measured)
   • R9700 measured page
   • GPU directory

2. 2 · Cited, not verified on our rig

   render and video groups

   AMD Linux prerequisites and the Fedora HC SIG both add the login to render and video: sudo usermod -a -G render,video $LOGNAME (SIG) or $USER. Log out, or reboot, before you trust rocminfo. Root “works” by skipping the check — that is not the fix. Not verified on our rig.

   • Verify ROCm (Ubuntu)
   • What is ROCm

3. 3 · Cited, not verified on our rig

   Mesa RADV Vulkan first

   Consumer first token on AMDai is llama.cpp + Vulkan. On Fedora that is Mesa RADV: mesa-vulkan-drivers and vulkan-tools, then vulkaninfo. Look for the discrete Radeon and a Mesa RADV ICD — not AMDVLK. If vulkaninfo is empty, stop; do not guess a driver. BIOS notes (Above 4G, ReBAR/SAM, CSM off) apply on the same board. Not verified on our rig.

   • llama.cpp
   • Vulkan vs HIP
   • BIOS

4. 4 · Cited, not verified on our rig

   ROCm / HIP from Fedora packages

   After Vulkan tokens are the goal, Fedora’s HIP stack is sudo dnf install rocm (optional rocm-devel for headers). That is the SIG’s command. It is Fedora’s package set, not AMD Core SDK 10.1. Confirm rocminfo prints the gfx the card actually is (gfx1201 on an R9700 / 9070-class card). Do not invent an HSA override. Moving that Fedora set later is a distro upgrade, not AMD 10.1 — see /rocm/upgrade. Not verified on our rig.

   • Install ROCm (Ubuntu)
   • Upgrade or reinstall ROCm
   • ROCm problems

5. 5 · Cited, not verified on our rig

   Verify: vulkaninfo, rocminfo, then a monitor

   Order: vulkaninfo (RADV + discrete Radeon), then rocminfo (ROCk module + Agent gfx) if you installed Fedora rocm. The SIG’s rocminfo example shows a gfx1201 agent. A userspace monitor such as amdgpu_top can show the card busy after tokens start — it is not required for first token and is not verified on our rig. We do not invent a Fedora tok/s.

   • Verify ROCm (Ubuntu)
   • Troubleshoot

6. 6 · Cited, not verified on our rig

   llama.cpp — Vulkan first, then HIP

   Build from ggml-org with the flags build.md documents: cmake -B build-vk -DGGML_VULKAN=ON, then cmake --build build-vk -j. HIP is second: GGML_HIP=ON only after rocminfo lists this gfx. Ollama, Strata, and vLLM are not a Fedora first-run on AMDai. Not verified on our rig.

   • llama.cpp
   • Backends
   • Speed up llama.cpp

7. 7 · Cited, not verified on our rig

   Pick Q6, Q5, or Q4 — Q3 is a labeled estimate

   Download the same official-instruct GGUF the Ubuntu recipes name. Start at Q6, Q5, or Q4 and check VRAM + KV on the calculator. Q3 stays a labeled estimate unless a Fedora run card exists — none does in this repo. Q1/Q2 stay off. Tok/s on /benchmarks and /results are Ubuntu llama-bench on the local desktop. Treat them as comparison, not a Fedora promise.

   • Calculator
   • 35B recipe
   • Quants
   • Qwen compare

8. 8 · Cited, not verified on our rig

   Fit VRAM and the two RAM lines

   The calculator prints 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 Ubuntu peaks are 37.32 GB at 4K and 42.9 GB at 256K — measured on the local desktop, not a Fedora kit. Keep the GGUF on NVMe. Not verified on our rig.

   • Calculator
   • Flash RAM offload
   • Benchmarks

9. 9 · Cited, not verified on our rig

   First-run command (Ubuntu-measured shape)

   Use the house daily-driver launch from the 35B recipe / tuning guide: ./build-vk/bin/llama-server -m ~/models/Qwen3.6-35B-A3B-UD-Q5_K_XL.gguf -ngl 99 -c 262144 --cache-type-k f16 --cache-type-v f16 --host 127.0.0.1 --port 8080. That command is the Ubuntu local desktop shape. On Fedora it is not verified on our rig and not measured on the local desktop — confirm vulkaninfo, then tokens, then a loopback /v1. Do not copy Ubuntu tok/s onto this box. Read the 4K run card to see what “good” looked like on the measured path.

   • 35B recipe
   • 4K run card
   • Run a local Qwen
   • Tuning flags

10. 10 · Cited, not verified on our rig

    When it breaks

    Start from a symptom on /troubleshoot. Paste a log in the error matcher. Cited rows live on /rocm-problems. Firmware: /bios. Jargon: /glossary. Short answers: /faq. Stack Lab remains the Ubuntu companion — packages are not published yet and there is no Fedora package. Not verified on our rig.

    • Troubleshoot
    • Error paste
    • ROCm problems
    • FAQ
    • Glossary
    • Stack Lab

RAM lines (Ubuntu-measured)

Use the calculator before you download a second quant. The two labels stay Minimum system RAM needed and RAM needed at full context. The house 35B peaks below are Ubuntu, not Fedora.

Open in Stack Lab

Stack Lab remains Ubuntu-first. The button is the allowlisted scheme. There is no Fedora package and no public download.

Sources

• AMD — ROCm 10.1.0 compatibility matrix (official OS rows)
• AMD — Install ROCm 10.1.0 (apt / dnf / zypper on listed OS rows)
• AMD — Linux prerequisites (render and video groups)
• Fedora Wiki — Heterogeneous Computing SIG (ROCm on Fedora)
• Fedora Packages — rocm metapackage
• Fedora Packages — mesa-vulkan-drivers (RADV)
• llama.cpp — build.md (Vulkan / HIP)