vLLM on RDNA — Overview
WIP: This section is actively expanded from Discord and fork release notes. See Verification status before treating benchmarks as gospel.
The quickest way to serve LLMs on gfx1030 / RDNA is the prebuilt
blivioniag/vllm-rdna images on a
blivioniag/rocm-rdna ROCm + PyTorch base. A single
image targets seven RDNA architectures (gfx1030 through RDNA4).
Not sure whether to use vLLM or llama.cpp? Start with llama.cpp or vLLM.
Where to start
| Goal | Page |
|---|---|
| First serve on a V620 (Docker or fork + launcher) | Onboard on V620 |
| Pull an image and run your first model | Running (Docker) |
| Which image / model / card count (27B, presets) | Recipes |
| GLM-5.3-Flash on 8× (local 0.30 fork) | Recipes — GLM-5.3-Flash |
| Qwen3.8 Flash-Next (benchmarks and serve lines) | Flash-Next |
Flash-Next MonoKernel (--monokernel, HEAD) | MonoKernel |
Flash-Next DFlash2 (RECIPE=flashnext-dflash2, HEAD) | DFlash2 |
Flash-Next 2× expert offload (serve_offload.sh, HEAD #146) | 2× MoE offload |
| Env vars, CUDA graphs, Docker Compose | Configuration |
| GPTQ vs AWQ, KV cache, MTP, INT4 | Quantization |
Custom RDNA2 HIP kernels (-extras) | vLLM forks |
| What shipped vs HEAD vs backlog | Fork progress |
| Which fork is official vs historical | Fork landscape |
Host uv venv (7.2.0 or 7.14.0) | Host venv |
One launcher for Flash-Next / 27B / EXL3 (serve_rdna.sh) | Serve scripts |
| Rebuild or extend Docker images | Building images |
| Something broke | vLLM troubleshooting |
Image variants
| Variant | When to use |
|---|---|
v0.31.0-extras / v0.31.0-extras-rocm7.14.0 | Official extras on vLLM 0.31.0 — rdna_extras. #announcements (11 Oct 2026): recommended day-to-day Docker path (ROCm 7.14). Flash-Next MTP-2 (BF16 + INT4-disk) served from this tag. Same digest under both names. Start at Onboard. |
v0.28.0-extras / v0.27.1-extras* | Historical extras bakes. Source for 0.28 is rdna_extra/v0.28.0. Keep for A/B only. |
v0.27.1 / v0.27.1-rocm7.14.0 | Stock upstream vLLM — baseline or comparison. #vllm-rdna Sep 24: plain v0.27.1-rocm7.14.0 is stale / removable; prefer -extras or host venv. |
Image tags are refreshed in place when fixes land — always docker pull before debugging. Confirm your
-extras image includes the latest extras commits (AWQ dispatch, GDN HIP, TP graph fix).
#vllm-rdna (Aug 31 2026): blivioniag/vllm-rdna:v0.27.1-extras was refreshed again — re-pull even if
you already had that tag.
Official source: Kernel work lives in
opengfx1030/vllm-rdna(rdna_extras= v0.31.0 as of 11 Oct 2026). Day-to-day serving is Hubblivioniag/vllm-rdna:v0.31.0-extras. First-time path: Onboard on V620. Olderv0.27.1-extras*/v0.28.0-extras*tags stay on Hub for A/B.#vllm-rdnaSep 14–15: Flash-Next commits are inrdna_extras; the standaloneleapdragon/vllm-rdna2-qwenbranch is no longer the moving target. Prefer the 0.31 Hub tag or a host venv over the old leapdragon image. Details: Fork landscape.Upstream vLLM: Official GPU docs still omit Navi 21 / gfx1030. Keep the community fork until upstream documents it.
Related
- Choose a stack — llama.cpp vs vLLM, and what fits on a V620.
- Flash-Next — card counts, AutoRound, and serve lines.
- Multi-GPU PCIe P2P — important for tensor parallel.
- Environment variables — cheat-sheet.
- Community recipe book (mirror
opengfx1030/vllm-rdna2-recipe) — presets, patches, concurrent MTP PRs.