cd /news/artificial-intelligence/native-vllm-and-rocm-7-15-for-rx-600… · home topics artificial-intelligence article
[ARTICLE · art-101130] src=github.com ↗ pub= topic=artificial-intelligence verified=true sentiment=↑ positive

Native vLLM and ROCm 7.15 for RX 6000 (RDNA2) on Windows 11 – 26 Tflops FP16

A new native installer brings vLLM and ROCm 7.15 (TheRock) to AMD Radeon RX 6000 (RDNA2) GPUs on Windows 11, achieving 26 TFLOPS FP16 and 58-62 tok/s on Qwen3.5-4B, verified on an RX 6750 XT 12 GB. The one-click installer, developed by sebastianmechno-sys, eliminates WSL2, NVIDIA, and compiler requirements, providing an OpenAI-compatible chat server and prebuilt binaries.

read6 min views1 publishedAug 18, 2026
Native vLLM and ROCm 7.15 for RX 6000 (RDNA2) on Windows 11 – 26 Tflops FP16
Image: Michielbdejong (auto-discovered)

Native vLLM + ROCm 7.15 (TheRock) for the whole AMD Radeon RDNA2 family on Windows — no WSL2, no NVIDIA, no compiler. One-click installer, everything prebuilt, and an OpenAI-compatible chat server that looks and works like the NVIDIA stack.

Verified on AMD Radeon RX 6750 XT 12 GB (gfx1031) — Windows 11 native — August 2026

Result Number
rocBLAS FP16 GEMM (native bench) 25 674 Gflops ≈ 26 TFLOPS
vLLM decode, Qwen3.5-4B 4-bit ~58-62 tok/s (8.3 → 62.5 = 7.5× optimized)
Item Requirement
OS Windows 10/11 (Windows 11 recommended; tar must support zstd — automatic on Win11)
GPU AMD RDNA2 — RX 6400 / 6500 / 6600 / 6650 / 6700 / 6750 / 6800 / 6900 / 6950 (all XT/M variants), 8+ GB VRAM for the 4B model
Driver AMD Software: Adrenalin Edition (the normal gaming driver)
Disk ~25 GB free on C:
Internet only during install (~6 GB: stack ~2.3 GB + model ~3.8 GB)
Admin one UAC click (installer auto-elevates)

No compiler, no ROCm installer, no manual setup — everything ships prebuilt.

Download the repository (ZIP or

git clone

). You donot need to download the release archives (*.tar.zst

) manually — the installer fetches them automatically from theReleasestab. - Double-click— by default it downloads from this repo's releases; pass a GitHub username to use your own fork instead. It checks GPU + disk, then installs everything it does not already have (re-run is always safe and fast):INSTALL.bat

Step Action 1/6 GPU detection (warns if not RDNA2) + Python 3.11.9 2/6 4 archives from GitHub Releases → C:\Python311

,C:\TheRock

,C:\vw_*_build

3/6 venv fix + torch self-check 4/6 Qwen3.5-4B 4-bit model (skipped if already in HuggingFace cache) 5/6 writes config.bat

(the file to edit to change model later)6/6 verification benchmark - → opens the web chat in your browser (starts the model server automatically the first time). While the model thinks you see a smallCHAT.bat

*Thinking…*spinner; when it's done, only the final answer streams in — the internal reasoning stays hidden. Tokens stream live with a tok/s counter, everything local on your AMD GPU. - → starts the model server on its own (OpenAI-compatible API onSERVE.bat

http://127.0.0.1:8000/v1

, likevllm serve

on NVIDIA). Use it with any OpenAI client, or just runCHAT.bat

. - → all 3 verification checks in one run: ROCm GPU detection, native rocBLAS FP16 power (VERIFY.bat

~26 TFLOPS) and the full 512-token vLLM benchmark (~58-62 tok/s).

Test Config Result
rocBLAS FP16 GEMM 4096×4096×4096, rocblas-bench.exe 25 674 Gflops (≈26 TFLOPS)
vLLM Qwen3.5-4B decode 512 tok, greedy, CUDA graphs 59.4 tok/s (up to 62.5)
Optimization progression eager fp16 baseline 8.3 → 62.5 tok/s (7.5×)

Full optimization history:

# configuration tok/s
1 fp16 eager (baseline) 8.3
2 + CUDA graphs + skinny GEMV 24.4
3 + AWQ 4-bit quantization 29.9
4 + native HIP W4 GEMV kernel 35.9
5 + M=1 GEMV for lm_head 58.1
6 + direct-store kernel path 59.1
7 + weight-cast caching 62.5

TheRock builds ROCm (HIP runtime, rocBLAS, Tensile) as native Windows binaries — this is what makes ROCm exist on Windows at all.HSA_OVERRIDE_GFX_VERSION=10.3.0

presents any RDNA2 GPU as gfx1030; the HIP kernel ships as afat binary (gfx1030 + gfx1031 + gfx1032) so the whole RX 6000 series runs native code.- PyTorch 2.12 +rocm7.15

links against that runtime →torch.cuda.is_available() == True

on RDNA2 Windows. - vLLM plugin vllm_windows_rocm

registers the tuned kernels: native HIP W4 GEMV for quantized linears, M=1 skinny GEMV for dense ones (including the huge tied lm_head), CUDA-graph safe (registered as real torch ops). INSTALL.bat

(engine:INSTALL.ps1

, manifest:MANIFEST.json

) downloads the 4 prebuilt archives from GitHub Releases and the model from HuggingFace, installs base Python, fixes the venv, verifies with a benchmark. Idempotent: it only downloads what is missing.

C:\Python311                                 Python 3.11.9
C:\TheRock\.venv                             torch 2.12+rocm7.15 venv (vLLM 0.19.1)
C:\TheRock\build\dist\rocm                   ROCm runtime libraries
C:\TheRock\ROCM_VLLM_RUNTIME                 vLLM + plugin + rocBLAS + rocblas-bench
C:\vw_cext_build, C:\vw_hipgemv_build        native HIP kernels
%USERPROFILE%\.cache\huggingface             model weights

Release archives (this repo's Releases tab, tag v2.0

):

Archive Size Content
the-rock-venv.tar.zst
1.34 GB torch ROCm venv
therock-rocm-dist.tar.zst
0.85 GB ROCm runtime
vllm-stack.tar.zst
0.14 GB vLLM + plugin + rocBLAS + rocblas-bench.exe
native-kernels.tar.zst
~1 MB HIP GEMV kernels (fat binary)
vllm-rocm-windows-rdna2-oneclick/
├── INSTALL.bat               one-click installer (entry point)
├── INSTALL.ps1               installer engine (downloads, extracts, verifies)
├── CHAT.bat                  opens the web chat (auto-starts the server)
├── SERVE.bat                 starts the model server alone (OpenAI API)
├── VERIFY.bat                all 3 verification checks in one run
├── chat.html                 the browser chat UI (Thinking spinner + tok/s)
├── MANIFEST.json             release archive names + extract targets
├── plugin_overrides/         tuned plugin modules (awq_gemv, bf16_gemv)
├── kernels/                  native HIP W4 GEMV source + prebuilt fat binary
├── scripts/                  model server entry + benchmark + rebuild helper
├── assets/                   verification screenshots
└── results/                  raw benchmark logs + progression chart

Edit config.bat

(written by the installer): set SERVED_MODEL

to the model folder and MODEL_NAME

to the name shown in the chat / used by the API, then run CHAT.bat

again. The benchmark uses BENCH_MODEL

(same file):

C:\TheRock\.venv\Scripts\python.exe -c "from huggingface_hub import snapshot_download; print(snapshot_download('<owner>/<repo>'))"

prints the snapshot folder to put in config.bat

.

Symptom Fix
No module named 'vllm._C' warnings
expected — the Windows plugin loads native kernels instead
Python installer exit 1601 automatic NuGet fallback kicks in; nothing to do
extraction "Can't unlink" errors close stray Python processes, re-run INSTALL.bat (it resumes)
low tok/s close other GPU workloads; verify overall_tok_s ≥ 55 on a cold GPU
non-RDNA2 GPU installer warns; unvalidated outside RDNA2
tar says "Unrecognized archive format" on the .zst files
old Windows 10 without zstd — install Windows updates (bsdtar 3.5+ bundles libzstd; Windows 11 is fine)
chat page says "Server error" the model is still — wait for "Application startup complete" in the SERVE window (~1 minute the first time)
custom debugging set BENCH_MODE=eager , set VLLM_WIN_HIPGEMV=0 , set VLLM_WIN_BF16_GEMV=0

scripts\rebuild_kernel.py

recompiles kernels/src/gemv_w4.cu

(needs HIP SDK

  • MSVC). Only for non-RDNA2 targets or exotic torch ABIs; the shipped fat binary covers all RDNA2.

Built on ROCm/TheRock, PyTorch ROCm and the vLLM project. Not affiliated with AMD.

── more in #artificial-intelligence 4 stories · sorted by recency
── more on @amd 3 stories trending now
sponsored brought to you by zahid.host 4,200+ EU-deployed projects
reading about agents? ship yours in a single git push.

Run your AI side-project on zahid.host

EU-based hosting, git-push deploys, automatic HTTPS, no cold starts. Free tier with a custom domain — perfect for shipping the agent you just read about.

$git push zahid main
Live at https://your-agent.zahid.host
Get free account → Pricing
from €0/mo · no card required
LIVE [news/native-vllm-and-rocm…] indexed:0 read:6min 2026-08-18 ·