Useful Resources
Curated links for gfx1030 / RDNA2 work. Grouped by type — not everything here is “official” or wiki-maintained; see Verification status for claim audits on guides that reference these projects.
Community Discord
Primary hangout for V620 / gfx1030 inference work. Wiki pages often cite channel names
(#vllm-rdna, #llamacpp, #general, #benchmarks, #lmcache) — that is where the claim came from.
| Server | gfx1030 club |
| Invite | https://discord.gg/mESex2aBp |
Official documentation
- ROCm documentation
- ROCm install on Linux
- System requirements & supported GPUs
- GPU architecture specs (Navi 21 = gfx1030)
- ROCm Device Support Wishlist — community-tracked support matrix on the official ROCm repo
Docker images
Prebuilt images for RDNA (gfx1030 through RDNA4). See Running vLLM for run commands and Building images for how they are produced.
| Image | Purpose |
|---|---|
blivioniag/rocm-rdna | ROCm + PyTorch base for RDNA cards |
blivioniag/vllm-rdna | vLLM serving images (upstream and -extras / rdna_extras lineage) |
Tags are listed on Docker Hub and in Running (Docker). Re-pull before debugging — tags are refreshed in place when fixes land.
Upstream projects
Stock projects this wiki builds on or tracks. Use these when you want upstream behavior or to compare against forks.
- vLLM — upstream serving engine
- llama.cpp — GGUF inference (HIP/ROCm and Vulkan)
- vLLM PR #52391 — RDNA gfx1030 platform detection (upstream CI enablement for consumer Radeon)
- TheRock: RDNA2 (gfx103X) support — ROCm/TheRock gfx1030 enablement work
Community projects
Forks, recipes, and tooling maintained outside (or alongside) upstream. Often where gfx1030-specific performance work happens first.
| Project | What it is |
|---|---|
blivioniag/v620_toolbox | V620 power cap + PCIe P2P — powertuning/ (Fedora), ubuntu_powertuning/ (Ubuntu 26.04); see Power tuning |
Tamalero/amd-v620-soft-unlock | V620 OverDrive soft unlock in passthrough VMs (QEMU romfile= pptable patch; no flash, keeps 72 CUs) — Power tuning |
opengfx1030/vllm-rdna-docker | Docker bake for rocm-rdna / vllm-rdna — moved here #vllm-rdna Sep 14 2026; Hub tags still blivioniag/* — Building images |
BlivionIaG/vllm-rdna-docker | Historical bake repo (pre-move) |
opengfx1030/vllm-rdna @ rdna_extras | Official vLLM extras fork — RDNA HIP kernels + Flash-Next cherry-picks (Sep 14–15); PRs/issues here — fork landscape |
BlivionIaG/hippihx | HIP kernel op zoo (gfx1030 / gfx1100 / gfx900 tile contracts) — #hippihx Sep 16: intended to keep rdna_extras thin; not a published vLLM image yet |
ashhart/TensorFold #389 | Draft ROCm backend (--backend rocm) for MLX affine checkpoints (Qwen3.5 / 3.8 dense, Qwen3.6-35B-A3B). RDNA2 is FP16 v_dot2, not WMMA. Earlier community work: BlivionIaG/TensorFold#1. #tensorfold (Oct 4): not faster than vLLM yet. #tensorfold (Oct 6): the Python engine was frozen; the ROCm path is being rewritten in Zig. No wiki serve recipe. Author PR has a V620 / W7800 matrix — Needs verify. |
BlivionIaG/vllm-rdna-qa | Maintainer QA / landing playbook for opengfx1030/vllm-rdna (#vllm-rdna Sep 17) — not a second kernel wiki |
alanoo81/flashnext-v620-pp3 | Temporary 3× V620 PP3 overlay (leapdragon image + org MoE HIP) — #vllm-rdna Sep 17–18; #general Sep 23 unmaintained (KV-pool dump loops at high concurrency) — recipe |
yiminyuan/vllm @ gfx1030/v0.28.0 | Community DeepSeek-V4 Flash gfx1030 tree — #vllm-rdna Sep 17; weights; Needs verify |
leapdragon/vllm-rdna2-qwen | Older Flash-Next container / docs/rdna2 — prefer org rdna_extras / Hub v0.31.0-extras for day-to-day; tree still receives occasional commits (last push 11 Oct 2026) — ROCR idle-CPU fix |
blivioniag/vllm @ rdna2_extras | Historical — predecessor of opengfx1030/vllm-rdna; public repo 404 as of Oct 2026. Current Hub extras bakes clone the org repo; only older v0.26 / v0.27.1-extras* bake targets still name this clone |
leapdragon/vllm-rdna2-recipe | Community recipe book + GHCR presets (27B/122B; parts pile; concurrent MTP PRs) — wiki recipes |
opengfx1030/vllm-rdna2-recipe | Org mirror of the recipe book (#vllm-rdna) |
edwinbrowwn/llama.cpp-rdna2 | RDNA2/V620 llama.cpp fork — see overview |
Niko1221/Strata | Separate HIP engine (not llama.cpp / vLLM). gfx1030 notes in docs/AMD_HIP.md. Opt-in STRATA_HIP_PROMPT_F16=1 (PR #835) runs 16-bit prompt GEMMs FP16 I/O on gfx103x. #general Oct 7–8: 2× V620 0.1.39 + #835/#849 ~doubled PP vs 0.1.38; 0.1.40.2 similar with the env on. 3-card: prefill still behind vLLM; decode collapsed with concurrency. #strata (25 Oct): 0.1.49.2 signed DeepSeek-engine feed has no gfx1030 HIP package (HIP build lists gfx1200/gfx1201; CUDA packages are a different arch). #vllm-rdna (28 Oct): easier onboarding than vLLM; more interesting below 4 cards / when experts spill. No wiki serve recipe. |
BlivionIaG/hi2 | Quantizer for rdna_hint_flex (variable INT2/INT4/INT8 rows). #vllm-rdna / #hi2 (27 Oct): intended middle ground vs EXL3 (quality) and AWQ (speed). Public repo only — no vLLM recipe / Hub bake. |
GeorgeMA-Strong/llm-context-bench | Reproducible long-context PP/TG benches (real prompts) — used by #benchmarks |
vllm-project/vmon | Upstream vLLM author monitor for prefill/decode viz — #vllm-rdna Oct 13; not wiki-validated on gfx1030 |
LMCache/LMCache | KV cache layer (RAM/SSD/remote) — #lmcache WIP; use lmcache/standalone (CPU) + connector. Official kernels CDNA-only; hybrid Qwen needs HMA the connector lacks — troubleshooting |
2kiss/flash-attention-rdna2 | Independent RDNA2 FlashAttention kernel — #vllm-rdna Sep 22–24: stale; org target remains in-tree fa_rdna2 (currently disabled after #17) — fork |
intentee/paddler | LLM load balancer / multi-instance router — discussed for multi-agent llama.cpp; not wiki-validated |
sebastianmechno-sys/vllm-rocm-windows-rdna2 | Unofficial Windows 11 + ROCm 7.x vLLM for RX 6000 — not wiki-validated |
skyne98/wiki-gfx906 | Sibling wiki for gfx906 (Vega 20 / MI50) |
Tools & write-ups
- Unsloth: Qwen3.8-Flash-Next / MTP llama.cpp —
public MTP build notes (
#llamacpp/#generalSep 16–17). froggeric/Qwen-Fixed-Chat-Templates— drop-in Jinja for official Qwen chat templates (#vllm-rdnaOct 8; also usable with llama.cpp--chat-template-file). Community: fewer empty-think / KV-invalidation stalls.#general(Oct 9): tool calls holding up (OG Qwen3.8 + Open WebUI). Needs verify.- Disabling ECC on Radeon Pro GPUs (lunnova.dev) —
amdgpu.ras_enable=0+ two reboots; pinned in#vllm-rdna. See Disabling ECC. rocminfo,rocm-smi,amd-smi,clinfo— GPU visibility, topology, and P2P checks. See Environment variables.- UEFITool / ReBARUEFI — locate hidden MMIO High / ReBAR / SR-IOV setup vars. See Host firmware.
- Precision 7920 MMIO GPU Unlock Guide — circulating T5820/T7820/T7920 write-up; dump your own UEFI, do not copy another host’s offsets.
- GpuMMIOFix — OS-side PCI BAR remap above 4 GB
(
#motherboardSep 15–16). Community: typically reload each boot; Needs verify on V620 hosts — Host firmware.
Know a gfx1030 resource that belongs here? Contribute a link and say which section it fits (official / Docker / upstream / community).