---
title: Qwen3.8-27B UD-IQ4_XS on the Radeon RX 7900 XT 20GB (Windows, HIP)
url: "https://bitworks.io/amd-inference/recipes/rx-7900-xt-20gb-windows-hip-qwen3-8-27b-iq4xs/"
description: HIP engine setup for the RX 7900 XT 20GB on Windows, measured by Bitworks. FastLLM picks it on this card and falls back to Vulkan if it cannot use it.
---

# Qwen3.8-27B UD-IQ4_XS on the Radeon RX 7900 XT 20GB (Windows, HIP)

> HIP engine setup for the RX 7900 XT 20GB on Windows, measured by Bitworks. FastLLM picks it on this card and falls back to Vulkan if it cannot use it.

Human-readable page: https://bitworks.io/amd-inference/recipes/rx-7900-xt-20gb-windows-hip-qwen3-8-27b-iq4xs/

## Content

- **Status:** In lab testing (candidate)

- **GPU:** 1 × AMD Radeon RX 7900 XT 20GB

- **Operating system:** Windows

- **Backend:** hip

- **Model:** Qwen3.8-27B UD-IQ4_XS — Unsloth GGUF conversion of Qwen/Qwen3.8-27B (Apache-2.0)

- **Context:** 16384 tokens

- **Concurrency:** 1

- **Placement:** full-gpu

A Bitworks lab observation (not fully qualified) is published for this recipe.

Measured by Bitworks on our lab bench with the published methodology. Not fully qualified: residency, peak VRAM, soak and reference-quality gates are not all proven for this run.

## Overview

This is the HIP variant of the Vulkan recipe for the same card, model and quantization. It uses the same Qwen3.8-27B UD-IQ4_XS weights from Unsloth (a third-party GGUF conversion of the Qwen upstream weights), the same context size and the same concurrency. The Vulkan recipe sets no separate KV cache options, so neither does this one. Only the engine changes: llama.cpp build b11658 compiled for HIP, running on the AMD ROCm 10.0 runtime (rocBLAS and hipBLASLt for gfx1100).

Bitworks has measured this setup on the Radeon RX 7900 XT 20GB through FastLLM 2026.10.15.5, using llama.cpp b11658 (HIP) with the AMD ROCm 10.0 runtime. The measured results are shown with the observation on this page. The recipe itself stays a candidate.

FastLLM picks this HIP setup automatically on this card. The first time, it shows "Install engine components" with the license links for the AMD ROCm runtime and the llama.cpp HIP engine, and installs them once you accept. If you turned on automatic installation of engine components, it installs them without asking. The extra download is about 565 MB (about 2.33 GB on disk once unpacked), from github.com and stable.repo.amd.com. If the components are not installed, their license is declined, or Windows blocks or cannot load them, FastLLM uses the Vulkan setup for this card instead and says so.

## Requirements

- Windows with a current AMD graphics driver for the Radeon RX 7900 XT 20GB.

- FastLLM, which is still in development.

- The AMD ROCm 10.0 runtime. FastLLM fetches it from AMD onto your PC the first time it sets up this card.

- Your acceptance of each component's license. FastLLM asks before anything downloads, unless you turned on automatic installation of engine components.

## Use it

- In FastLLM, click Install & start. FastLLM picks this setup for the card.

- Accept each component's license when asked. Nothing downloads before you accept. FastLLM then fetches the engine and the AMD ROCm 10.0 runtime onto your PC.

- Accept the model license when asked. FastLLM downloads the model and starts the HIP setup.

If you do not accept the component licenses, FastLLM uses the Vulkan setup for this card instead and says so. You can install the HIP setup later from "Choose another setup".

For agents: how to buy here — https://bitworks.io/checkout.md