V2.2 — Chat UI fixed + one-click launcher

V2.2 — Chat UI fixed + one-click launcher + license compliance

Fixed

• chat.html boot crash (the big one): el.themeBtn was missing from the element map since V2.1 → Uncaught TypeError in boot() → the chat UI never started for anyone: no welcome screen, dead send button, stuck on "connecting..." forever. One line fix, found via Chrome headless console + static element-map analysis. The engine itself was always fine (26 TFLOPS / 58 tok/s benchmarks are real).
• CHAT.bat one-click, for real: the chat page is now served on http://127.0.0.1:808x by scripts/chat_server.py (picks a free port, reuses a running instance, opens the browser by itself) instead of file:// — browsers block API calls from file:// pages, which kept the UI dead even with the server up.
• SERVE.bat: guard against double engine start (two engines fighting for 12 GB of VRAM).
• serve.py: reads config.bat directly, so it works even when launched standalone; explicit CORS args (match vLLM defaults).

License compliance

• Added THIRD_PARTY_NOTICES.md (vLLM, PyTorch, ROCm/TheRock, Python, fonts, trademarks).
• Attribution headers on the MrPie (ThePie88) kernels (Apache 2.0) + credit in EN and zh-CN READMEs.

Install

Code-only release — the engine archives are unchanged. INSTALL.bat keeps fetching the V2.1 release assets (Tag=V2.1 in INSTALL.ps1). Just re-download this repository and run INSTALL.bat (or CHAT.bat if already installed).

Full Changelog: V2.1...V2.2

V2.1 - One-click Native vLLM + ROCm for AMD RDNA2 on Windows

Code + chat overhaul - archives are unchanged since V2.0 (the installer still fetches them from the V2.0 release).

Installer & integrity

• SHA256 verification of every downloaded part against SHA256SUMS.txt (published on the V2.0 release) - corrupted/truncated downloads are rejected and retried, no partial files ever left behind
• -SkipChecksum for forks/mirrors without the checksum file
• -Variant auto|rdna2|rdna3|rdna4 with automatic GPU-family detection; VRAM-aware default model (4 GB cards get Qwen2.5-1.5B-AWQ automatically)
• CI pipeline validates the manifest against the release API + PSScriptAnalyzer lint on every push; release/checksum tooling (make_checksums.ps1, validate_release.ps1)

Server (serve.py / SERVE.bat)

• Explicit enable_thinking (default THINKING=0 = direct answers, no "Thinking Process" preamble in responses)
• SERVE.bat now auto-detects the installed model from the HuggingFace cache when config.bat is missing - fresh ZIP downloads work out of the box

Chat (chat.html)

• Total redesign: animated aurora background, glassmorphism, gradient accents, entrance animations
• Stop button: abort generation mid-stream and keep the partial answer
• Light/dark theme toggle (persisted)
• Export conversation to Markdown (.md)
• Fixes: send button no longer stuck after the first message; model badge/select refresh when the server comes online late; streaming no longer pollutes other chats when switching mid-generation; watchdog resets the UI if the server goes silent

Docs & trust

• SECURITY.md, CONTRIBUTING.md, GPU verification + bug report issue templates, RELEASING.md, MULTIARCH.md

v2.0 - One-click Native vLLM + ROCm for AMD RDNA2 on Windows

v2.0 - One-click Native vLLM + ROCm for AMD RDNA2 on Windows

v2.0 — One-click native vLLM + ROCm 7.15 for AMD RX 6000 (RDNA2) on Windows

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, plus an OpenAI-compatible chat server that works like the NVIDIA stack.

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

---

Verification

Test	Result
ROCm GPU detection (native process)	HIP 7.15.26290, 12.0 GB VRAM, 20 CUs
rocBLAS FP16 GEMM 4096×4096×4096	25 674 Gflops (≈26 TFLOPS)
vLLM decode, Qwen3.5-4B AWQ 4-bit, 512 tok	59.4 tok/s (peak 62.5)
Optimization progression	8.3 → 62.5 tok/s (7.5×)

---

Installation

You do not need to extract or download these archives manually.
Just clone/download the repository and run INSTALL.bat — the installer
auto-elevates, checks GPU and disk, downloads the four archives below from
this release, installs Python 3.11.9, fetches the Qwen3.5-4B 4-bit model from
Hugging Face and runs a verification benchmark.

Then:

• CHAT.bat — open the web chat (server starts automatically)
• SERVE.bat — OpenAI-compatible API at http://127.0.0.1:8000/v1
• VERIFY.bat — full 3-step verification

Installed layout:
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 C:\TheRock\ROCM_VLLM_RUNTIME vLLM + plugin + rocBLAS + rocblas-bench C:\vw_cext_build, C:\vw_hipgemv_build native HIP kernels

---

Assets (the installer downloads these automatically)

File	Size	Content
the-rock-venv.tar.zst	1.34 GB	torch ROCm venv
therock-rocm-dist.tar.zst	0.84 GB	ROCm runtime (HIP, rocBLAS, Tensile)
vllm-stack.tar.zst	0.14 GB	vLLM + plugin + rocBLAS + rocblas-bench.exe
native-kernels.tar.zst	~1 MB	HIP GEMV kernels (fat binary gfx1030/1031/1032)

---

Requirements

Item	Requirement
OS	Windows 10/11 (Windows 11 recommended)
GPU	AMD RDNA2 — RX 6400–6950, 8+ GB VRAM
Driver	AMD Adrenalin (standard gaming driver)
Disk	~25 GB free on C:
Internet	during installation only (~6 GB)

---

Highlights

• Native HIP W4 GEMV fat-binary kernel for the entire RDNA2 family
• Tuned decode: CUDA graphs + M=1 GEMV + weight-cast caching (7.5× throughput)
• OpenAI-compatible API + browser chat (Thinking… spinner, final answer only)
• Apache 2.0 — built on ROCm/TheRock, PyTorch ROCm and the vLLM project. Not affiliated with AMD

v1.1.0 - Native vLLM + ROCm 7.15 Runtime for AMD RX 6000 (RDNA2) on Windows 11

v1.1.0 - Native vLLM + ROCm 7.15 Runtime for AMD RX 6000 (RDNA2) on Windows 11
First working native implementation of vLLM + ROCm 7.x (TheRock) for AMD Radeon RX 6000 Series on Windows 11 — No WSL2

What this is
AMD officially lists RX 6750 XT, 6700 XT, 6600 XT as "Runtime only" on Windows, with HIP SDK excluded. This release closes that gap.

This project provides:

rocBLAS binaries built for gfx1031 via ROCm/TheRock
PyTorch 2.12 built against TheRock ROCm runtime (torch.cuda.is_available() True on RX 6750 XT)
vLLM 0.19.1 with custom Windows plugin windows_rocm using Triton Attention
No WSL2 wrapper — native Windows HIP and ROCm directly on RDNA2
Goal: make local LLM inference accessible to millions of RX 6000 users without NVIDIA hardware, and provide a reference for AMD engineers to re-enable RDNA2 in official ROCm 7 Windows builds.

Verified on RX 6750 XT 12GB (gfx1031) — August 2026
Environment:

torch 2.12.0+rocm7.15.0a20260728 | cuda_avail True | dev AMD Radeon RX 6750 XT
Platform plugin windows_rocm is activated
rocBLAS Benchmark — 25.9 TFLOPS FP16:

Device ID 0 : AMD Radeon RX 6750 XT gfx1031
rocBLAS version: 5.7.0.67811f1ee52
N,N,4096,4096,4096,1,4096,0,4096,4096, 2, 10, 25977.3, 5290.73
=> 25977.3 Gflops = 25.97 TFLOPS in 5.29ms
vLLM Inference — FIRST_TOKEN_OK:

Loading model: facebook/opt-125m
Available KV cache memory: 5.47 GiB / 159,264 tokens
PROMPT: 'Hello, my name is'
OUTPUT: ' J.C. and I am a student at the University of California, Berkeley...'
FIRST_TOKEN_OK
Input: 10.16 tok/s Output: 54.20 tok/s | Init 1.84s
Full logs in benchmarks/ folder on main page.

What's in this Release (708MB)
ROCm_VLLM_Runtime_RDNA2_Windows.zip contains minimal required runtime:

ROCM_VLLM_RUNTIME/ — vLLM 0.19.1 + vllm_windows_rocm plugin
rocm_binaries/ — HIP runtime + rocBLAS (hipInfo, rocblas-bench, dlls)
Does NOT include full 90GB TheRock build, only required binaries. Main repo on main branch is lightweight (<5MB) with README, setup.bat, run.bat, inference.py, assets and logs.

Quick Start
Download zip from this release and extract to C:\TheRock
Run setup.bat as Administrator (sets HSA_OVERRIDE_GFX_VERSION=10.3.1, HIP_VISIBLE_DEVICES=0, VLLM_TARGET_DEVICE=rocm, MASTER_ADDR=127.0.0.1)
Run run.bat — Expected: FIRST_TOKEN_OK
Custom model: python inference.py --model facebook/opt-125m --prompt "Hello, my name is"

Fixes in v1.1.0
Fixed zmq.error.ZMQError: Protocol not supported (ipc://...) on Windows by forcing VLLM_ENABLE_V1_MULTIPROCESSING=0 to use TCP instead of IPC
Fixed socket error 10049 by setting MASTER_ADDR/PORT
Improved performance: 54.20 tok/s output (was 38.75 tok/s)
Repo cleaned from 1GB to <5MB
Acknowledgements
Built on ROCm/TheRock, PyTorch ROCm and vLLM. Not affiliated with AMD.

License: Apache 2.0