# Unleash the Power of AMD: Official Support for Unsloth is Here!

<div data-with-frame="true"><figure><img src="/files/lJerDyxpdxkq1UpVvaos" alt=""><figcaption></figcaption></figure></div>

## 1. Introduction of Unsloth

Unsloth is a library for fine-tuning LLMs, built on custom Triton kernels and optimized PyTorch. It makes high-performance fine-tuning accessible without top-tier hardware, and now that includes AMD GPUs.

## 2. Install Unsloth on AMD

Getting started with Unsloth on your AMD hardware is straightforward. We recommend using a modern Linux distribution for the best experience.

The easiest way to install is via the one-line installer, which sets up Unsloth Studio, a full graphical interface for managing fine-tuning runs, running inference, generating datasets with recipes, and comparing models, as well as all necessary dependencies including ROCm-optimized kernels:

```
curl -fsSL https://unsloth.ai/install.sh | sh
```

This will automatically detect your AMD GPU, install the correct ROCm-compatible PyTorch wheels, and launch the Studio UI so you can get started immediately.

### 2.1 Install via pip (Python library only)

If you prefer to use Unsloth as a Python library without the Studio UI, you can install it directly with pip:

```
pip install unsloth[amd]
```

Note: this does not install PyTorch, you'll need a ROCm-compatible PyTorch installed separately first. See the [Unsloth AMD docs](https://unsloth.ai/docs/get-started/install/amd) for the full setup.

## 3. Using Unsloth Studio on AMD

Once installed, Unsloth Studio launches automatically in your browser, no cloud, no NVIDIA hardware needed. Below is a quick workflow to get started on AMD; check out the [official Unsloth Studio docs](https://unsloth.ai/docs/new/studio/start) for the full usage guide.

After running the installer, your AMD GPU is detected via ROCm™ software and the UI is ready to go. Select your model, dataset, and parameters, then hit Start Training.

<div data-with-frame="true"><figure><img src="/files/c7hpcfg8ZDfgCb9YXwBA" alt=""><figcaption></figcaption></figure></div>

*Unsloth Studio's fine-tuning configuration screen, fully customizable to get your model training exactly how you want it.*

<div data-with-frame="true"><figure><img src="/files/lHQxdw0LgktAbrkGg5yM" alt=""><figcaption></figcaption></figure></div>

*Unsloth Studio launching on AMD hardware, the terminal confirms ROCm detection while the UI loads your first model.*

The Studio tracks everything in real time, training progress, loss, VRAM usage, and GPU temperature, so you always know exactly what your hardware is doing.

<div data-with-frame="true"><figure><img src="/files/HWKjivUHxGE6UXszXhoc" alt=""><figcaption></figcaption></figure></div>

*A live fine-tuning run on AMD, Unsloth Studio tracks progress in real time.*

Once training is complete, export to GGUF, Safetensors, or LoRA for deployment anywhere, including Hugging Face, llama.cpp, Ollama, vLLM, and more. You can also use this feature to convert existing HF models to your preferred format.<br>

<div data-with-frame="true"><figure><img src="/files/3RGuSiSgnH7XzJthO8OO" alt=""><figcaption></figcaption></figure></div>

*Export your model to your desired format and deploy it anywhere.*

<figure><img src="/files/WL8WCC2Ug1uYrxbsKRvO" alt=""><figcaption></figcaption></figure>

*Chat with your newly fine-tuned model directly in the Studio playground.*

That's the full AMD workflow. To check if your GPU is supported, see the hardware table below.

## 4. Supported AMD Hardware

The initial release of Unsloth's official AMD support targets the most popular and powerful cards. The following table outlines the currently supported AMD hardware:

| Architecture | Series                                                               | gfx code                  | Status          |
| ------------ | -------------------------------------------------------------------- | ------------------------- | --------------- |
| RDNA 4       | Radeon™ RX 9000 Series graphics (Desktop)                            | gfx1200, gfx1201          | Fully Supported |
| RDNA 3.5     | Ryzen AI 300 Series (Strix Point) / Ryzen AI MAX Series (Strix Halo) | gfx1150, gfx1151, gfx1152 | Fully Supported |
| RDNA 3       | Radeon™ RX 7000 Series graphics (Desktop)                            | gfx1100, gfx1101, gfx1102 | Fully Supported |
| RDNA 2       | Radeon™ RX 6000 Series                                               | gfx1030–gfx1036           | Fully Supported |
| CDNA 4       | Instinct™ MI350 Series GPUs (Data Center)                            | gfx950                    | Fully Supported |
| CDNA 3       | Instinct™ MI300 Series GPUs (Data Center)                            | gfx940, gfx941, gfx942    | Fully Supported |
| CDNA 2       | Instinct™ MI200 Series GPUs                                          | gfx90a                    | Supported       |
| CDNA 1       | Instinct™ MI100 GPUs                                                 | gfx906, gfx908            | Not supported   |

*Fully Supported - hardware-specific kernel tuning is active (optimal performance).*&#x20;

*Supported - runs via the general HIP/ROCm path; functional but without architecture-specific optimizations.*

### 4.1 CI/CD for AMD

We didn't just ship AMD support, we built 95 automated tests to make sure it stays working ([PR #4824](https://github.com/unslothai/unsloth/pull/4824)). The suite covers the full ROCm stack and runs entirely in CI without physical AMD hardware, so every PR gets validated before it ships.

It's already proven its worth: [PR #4940](https://github.com/unslothai/unsloth/pull/4940) fixed a GPU target extraction bug where rocminfo emits both a generic family target (e.g. gfx9) and the specific architecture target (e.g. gfx90a) for the same GPU. The old code passed both to GPU\_TARGETS, which broke the HIP build. The fix deduplicates to keep only the specific target, dropping any generic prefix, verified on real MI300X hardware before merging.

5\. How AMD Support Was Built

Unsloth's custom Triton kernels were ported to AMD by working closely with the ROCm team, requiring a rework of the installer, hardware detection, and runtime. See [PR #4720](https://github.com/unslothai/unsloth/pull/4720) for full details. Key changes include:

* HIPification: Kernels ported to HIP/ROCm with profiling and memory optimization for AMD hardware.
* Multi-path ROCm detection: Detects AMD GPUs via fallbacks (rocminfo, amd-smi, hipconfig, etc.) across bare-metal, containers, and package-managed setups.
* HIP-native execution: Kernels route through HIP via Triton's is\_hip() runtime check, preserving NVIDIA precedence on mixed systems.
* AMD monitoring: New amd.py module for hardware monitoring and VRAM parsing.
* Radeon wheels ([PR #4770](https://github.com/unslothai/unsloth/pull/4770)): Auto-installs ROCm-compatible PyTorch wheels matched to the detected ROCm version.
* llama.cpp for ROCm: ROCm prebuilt binaries served to AMD hosts, with source compilation as fallback.

### 5.1. Future roadmap?

Our commitment to the AMD x Unsloth community doesn't stop here. Key items on the roadmap include:

* More robust CI/CD: AMD hardware added to CI/CD.
* Fix GPU utilization monitor support for AMD: fix coming soon.

## 6. Try Unsloth on AMD Developer Cloud

The easiest way to experience Unsloth on AMD today is through AMD's Developer Cloud, no setup required. It offers one-click notebook access powered by MI300X GPUs with 192GB VRAM, giving you instant access to some of the most powerful AI hardware available.

AMD offers $100 in free credits through the [AMD AI Developer Program](https://www.amd.com/en/developer/ai-dev-program.html), running any existing Unsloth notebook on AMD is as simple as swapping the domain in the Github URL:

<table data-header-hidden><thead><tr><th width="170"></th><th></th></tr></thead><tbody><tr><td><br></td><td>URL</td></tr><tr><td>Github URL</td><td><a href="https://github.com/unslothai/notebooks/blob/main/nb/AMD-gpt_oss_(20B)_Reinforcement_Learning_2048_Game_BF16.ipynb">https://github.com/unslothai/notebooks/blob/main/nb/AMD-gpt_oss_(20B)_Reinforcement_Learning_2048_Game_BF16.ipynb</a></td></tr><tr><td>AMD Dev Cloud</td><td><a href="https://amd-ai-academy.com/github/unslothai/notebooks/blob/main/nb/AMD-Gemma4_(E2B)_Reinforcement_Learning_Sudoku_Game.ipynb">https://amd-ai-academy.com/github/unslothai/notebooks/blob/main/nb/AMD-Gemma4_(E2B)_Reinforcement_Learning_Sudoku_Game.ipynb</a> </td></tr></tbody></table>

Questions or issues? Find us on [Discord](https://discord.com/invite/unsloth), browse the [docs](https://unsloth.ai/docs/get-started/install/amd), or file a bug on [GitHub](https://github.com/unslothai/unsloth/issues). We'd love to hear how Unsloth runs on your AMD hardware, and help you troubleshoot any issues.


---

# Agent Instructions: Querying This Documentation

If you need additional information that is not directly available in this page, you can query the documentation dynamically by asking a question.

Perform an HTTP GET request on the current page URL with the `ask` query parameter:

```
GET https://unsloth.ai/docs/blog/unleash-the-power-of-amd-official-support-for-unsloth-is-here.md?ask=<question>
```

The question should be specific, self-contained, and written in natural language.
The response will contain a direct answer to the question and relevant excerpts and sources from the documentation.

Use this mechanism when the answer is not explicitly present in the current page, you need clarification or additional context, or you want to retrieve related documentation sections.