amd-strix-halo-toolboxes/README.md

# AMD Strix Halo Llama.cpp Toolboxes

This project provides pre-built containers (“toolboxes”) for running LLMs on **AMD Ryzen AI Max “Strix Halo”** integrated GPUs. Toolbx is the standard developer container system in Fedora (and now works on Ubuntu, openSUSE, Arch, etc).

## Watch the YouTube Video

[![Watch the YouTube Video](https://img.youtube.com/vi/wCBLMXgk3No/maxresdefault.jpg)](https://youtu.be/wCBLMXgk3No)

## Why Toolbx?

* Reproducible: never pollute your host system
* Seamless: shares your home and GPU devices, works like a native shell
* Flexible: easy to switch between Vulkan (open/closed drivers) and ROCm

## Table of Contents

1. [Llama.cpp Compiled for Every Backend](#1-llamacpp-compiled-for-every-backend)  
    1.1 [Supported Container Images](#11-supported-container-images)
2. [Quickest Usage Example](#2-quickest-usage-example)  
    2.1 [Creating the toolboxes with GPU access](#21-creating-the-toolboxes-with-gpu-access)  
    2.2 [Running models inside the toolboxes](#22-running-models-inside-the-toolboxes)  
    2.3 [Downloading GGUF Models from HuggingFace](#23-downloading-gguf-models-from-huggingface)  
3. [Performance Benchmarks (Key Results)](#3-performance-benchmarks-key-results)  
4. [Memory Planning & VRAM Estimator](#4-memory-planning--vram-estimator)  
5. [Building Containers Locally](#5-building-containers-locally)  
6. [Host Configuration](#6-host-configuration)  
    6.1 [Test Configuration](#61-test-configuration)  
    6.2 [Kernel Parameters (tested on Fedora 42)](#62-kernel-parameters-tested-on-fedora-42)  
    6.3 [Ubuntu 24.04](#63-ubuntu-2404)
7. [More Documentation](#7-more-documentation)  
8. [References](#8-references)



## 1. Llama.cpp Compiled for Every Backend

This project uses [Llama.cpp](https://github.com/ggerganov/llama.cpp), a high-performance inference engine for running local LLMs (large language models) on CPUs and GPUs. Llama.cpp is open source, extremely fast, and is the only engine supporting all key backends for AMD Strix Halo: Vulkan (RADV, AMDVLK) and ROCm/HIP

* **Vulkan** is a cross-platform, low-level graphics and compute API. Llama.cpp can use Vulkan for GPU inference with either the open Mesa RADV driver or AMD's "official" open AMDVLK driver. This is the most stable and supported option for AMD CPUs at the moment.
* **ROCm** is AMD's open-source answer to CUDA: a GPU compute stack for machine learning and HPC. With ROCm, you can run Llama.cpp on AMD GPUs in a way similar to how CUDA works on NVIDIA - this is not the most stable/mature, but recently it's been getting better.

### 1.1 Supported Container Images

You can check the containers on DockerHub: https://hub.docker.com/r/kyuz0/amd-strix-halo-toolboxes/tags.  

| Container Tag        | Backend/Stack            | Purpose / Notes |
| -------------------- | ------------------------ | --------------- |
| `vulkan-amdvlk`      | Vulkan (AMDVLK)           | Fastest backend—AMD open-source driver. ≤2 GiB single buffer allocation limit, some large models won't load. |
| `vulkan-radv`        | Vulkan (Mesa RADV)        | Most stable and compatible. Recommended for most users and all models. |
| `rocm-6.4.2`         | ROCm 6.4.2 (HIP)          | Latest stable ROCm. Great for BF16 models. Occasional crashes possible. |
| `rocm-6.4.2-rocwmma` | ROCm 6.4.2 (HIP) + ROCWMMA | ROCm with ROCWMMA enabled for improved flash attention on RDNA3+/CDNA. |
| `rocm-6.4.3`         | ROCm 6.4.3 (HIP) + hipBLASLt*          | Latest stable ROCm. Great for BF16 models. Occasional crashes possible. |
| `rocm-6.4.3-rocwmma` | ROCm 6.4.3 (HIP) + ROCWMMA + hipBLASLt*  | ROCm with ROCWMMA enabled for improved flash attention on RDNA3+/CDNA. |
| `rocm-7rc`           | ROCm 7.0 RC (HIP) + hipBLASLt*         | Release candidate for ROCm 7.0. Same behavior as beta. |
| `rocm-7rc-rocwmma`   | ROCm 7.0 RC (HIP) + ROCWMMA + hipBLASLt*       | Release candidate for ROCm 7.0, with hipBLASLt and ROCWMMA for improved flash attention on RDNA3+/CDNA |

\* All these toolboxes now export `ROCBLAS_USE_HIPBLASLT=1` as this currently results in better perfromance and stability in *MOST* cases.

> These containers are **automatically** rebuilt whenever the Llama.cpp master branch is updated, ensuring you get the latest bug fixes and new model support. The easiest way to update to the newest versions is by running the `refresh-toolboxes.sh` [script below](#211-toolbox-refresh-script-automatic-updates).

> *Each container is based on Fedora Rawhide and is built for maximum compatibility and performance on Strix Halo.*

---

## 2. Quickest Usage Example

### 2.1 Creating the toolboxes with GPU access

To use Llama.cpp with hardware acceleration inside a toolbox container, you must expose the GPU devices from your host. The exact flags and devices depend on the backend:

* **For Vulkan (RADV/AMDVLK):** Only `/dev/dri` is required.
  *Add the user to the video group for access to GPU devices.*

  ```sh
  toolbox create llama-vulkan-radv \
    --image docker.io/kyuz0/amd-strix-halo-toolboxes:vulkan-radv \
    -- --device /dev/dri --group-add video --security-opt seccomp=unconfined
  ```

* **For ROCm:** You must expose both `/dev/dri` and `/dev/kfd`, and add the user to extra groups for compute access.

  ```sh
  toolbox create llama-rocm-6.4.2 \
    --image docker.io/kyuz0/amd-strix-halo-toolboxes:rocm-6.4.2 \
    -- --device /dev/dri --device /dev/kfd \
    --group-add video --group-add render --group-add sudo --security-opt seccomp=unconfined
  ```

*Swap in the image/tag for the backend you want to use.*

> **Note:**
>
> * `--device /dev/dri` provides graphics/video device nodes.
> * `--device /dev/kfd` is required for ROCm compute.
> * Extra groups (`video`, `render`, `sudo`) may be required for full access to GPU nodes and compute features, especially with ROCm.
> * Use `--security-opt seccomp=unconfined` to avoid seccomp sandbox issues (needed for some GPU syscalls).

### 2.1.1 Toolbox Refresh Script (Automatic Updates)

To pull the latest container images and recreate toolboxes cleanly, use the provided script:

#### 📦 `refresh-toolboxes.sh`

```bash
./refresh-toolboxes.sh all
```

This will:

1. Delete existing toolboxes (if any)
2. Pull the latest images from DockerHub
3. Recreate each toolbox with correct GPU access flags

You can also refresh just one or more toolboxes:

```bash
./refreshtoolboxes.sh llama-vulkan-amdvlk llama-rocm-6.4.2
```

### 2.2 Running models inside the toolboxes

Before running any commands, you must first enter your toolbox container shell using:

```sh
toolbox enter llama-vulkan-radv
```

*This will drop you into a shell inside the toolbox, using your regular user account. The container shares your host home directory—so anything in your home is directly accessible (take care: your files are exposed and writable inside the toolbox!).*

Once inside, the following commands show how to run local LLMs:

* `llama-cli --list-devices`
  *Lists available GPU devices for Llama.cpp.*
* `llama-cli --no-mmap -ngl 999 -fa -m <model>`
  *Runs inference on the specified model, with all layers on GPU and flash attention enabled (replace \*\* with your model path).*

## 2.3 Downloading GGUF Models from HuggingFace

Most Llama.cpp-compatible models are on [HuggingFace](https://huggingface.co/models?format=gguf). Filter for **GGUF** format, and try to pick Unsloth quantizations—they work great and are actively updated: https://huggingface.co/unsloth.

Download using the Hugging Face CLI. For example, to get the first shard of Qwen3 Coder 30B BF16 (https://huggingface.co/unsloth/Qwen3-Coder-30B-A3B-Instruct-GGUF):

```bash
HF_HUB_ENABLE_HF_TRANSFER=1 huggingface-cli download unsloth/Qwen3-Coder-30B-A3B-Instruct-GGUF \
  BF16/Qwen3-Coder-30B-A3B-Instruct-BF16-00001-of-00002.gguf \
  --local-dir models/qwen3-coder-30B-A3B/
```

`HF_HUB_ENABLE_HF_TRANSFER=1` uses a Rust-based package that enables faster download (install from [Pypi](https://pypi.org/project/hf-transfer/)).

## 3. Performance Benchmarks (Key Results)

Benchmarks were run on **AMD Ryzen AI Max “Strix Halo”** across all supported backends, testing both **prompt processing (PP)** and **token generation (TG)** throughput.
Reported values were analysed using error margins (mean ± σ). Backends whose ranges overlapped were treated as statistical ties rather than hard wins.

🌐 Interactive exploration of the latest benchmark runs: [Interactie Benchmark Viewer](https://kyuz0.github.io/amd-strix-halo-toolboxes/)


| Workload Focus                                    | 🏆 Recommended Backend/Config       | Win + Tie Count¹ | Typical Runner-Up                  | Stability Notes                                                                       |
| ------------------------------------------------- | ----------------------------------- | ---------------: | ---------------------------------- | ------------------------------------------------------------------------------------- |
| **Prompt processing** (pp512, Flash Attention ON) | **ROCm 7 RC + ROCWMMA + hipBLASLt** |               15 | Vulkan AMDVLK (4)                  | 0% errors in tests                                                                    |
| **Token generation** (tg128, Flash Attention ON)  | **Vulkan RADV**                     |               13 | Vulkan AMDVLK (1)                  | 0% errors in tests                                                                    |
| **Balanced workloads**                            | **Vulkan AMDVLK**                   |                — | RADV / ROCm 7 RC+ROCWMMA+hipBLASLt | Fast PP & decent TG; \~5.6 % load failure rate due to ≤ 2 GiB single-allocation limit |
| **BF16 models**                                   | **ROCm 7 RC + ROCWMMA + hipBLASLt** |                — | ROCm 6.4.2 + ROCWMMA               | Best PP & TG among ROCm backends; stable with Flash Attention ON                      |

¹ Counts show number of times the backend placed 1st (alone or tied) across tested models/quantisations.


### Key take-aways

* **ROCm 7 RC + ROCWMMA + hipBLASLt + Flash Attention ON**
  * Fastest prompt processing in the vast majority of tests (15/22 wins or ties).
  * Best ROCm option for BF16 models.
  * Zero recorded errors with Flash Attention ON.

* **Vulkan RADV**
  * Best token generation throughput (13/15 wins or ties).
  * Most stable and broadly compatible backend overall.

* **Vulkan AMDVLK**

  * Competitive in both PP and TG; benefits from margin-aware tie handling.
  * Limited by ≤ 2 GiB single buffer allocation, which can block some model architectures.
  * Other ROCm variants (beta, hblt0, 6.4.2 w/o ROCWMMA)
  * Inconsistent performance and/or higher error rates; best suited for experimental use.

📄 Full per-model analysis: [docs/benchmarks.md](docs/benchmarks.md)

## 4. Memory Planning & VRAM Estimator

Running large language models locally requires estimating **total VRAM required**—not just for the model weights, but also for the "context" (number of active tokens) and extra overhead.

Use `gguf-vram-estimator.py` to check exactly how much memory you need for a given `.gguf` model and target context length. Example output:

```
$ gguf-vram-estimator.py models/llama-4-scout-17b-16e/Q4_K_XL/Llama-4-Scout-17B-16E-Instruct-UD-Q4_K_XL-00001-of-00002.gguf --contexts 4096 32768 1048576

--- Model 'Llama-4-Scout-17B-16E-Instruct' ---
Max Context: 10,485,760 tokens
Model Size: 57.74 GiB
Incl. Overhead: 2.00 GiB

--- Memory Footprint Estimation ---
   Context Size |  Context Memory | Est. Total VRAM
---------------------------------------------------
         4,096 |       1.88 GiB  |      61.62 GiB
        32,768 |      15.06 GiB  |      74.80 GiB
     1,048,576 |      49.12 GiB  |     108.87 GiB
```

With Q4\_K quantization, **Llama-4-Scout 17B** can reach a 1M token context and still fit within a 128GB system, but... **it will be extremely slow to process such a long context**: see benchmarks (e.g. \~200 tokens/sec for prompt processing). Processing a 1M token context may take hours.

Contrast: Qwen3-235B Q3\_K (quantized, 97GiB model):

```
$ gguf-vram-estimator.py models/qwen3-235B-Q3_K-XL/UD-Q3_K_XL/Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003.gguf --contexts 65536 131072 262144

--- Memory Footprint Estimation ---
   Context Size |  Context Memory | Est. Total VRAM
---------------------------------------------------
        65,536 |     11.75 GiB |     110.75 GiB
       131,072 |     23.50 GiB |     122.50 GiB
       262,144 |     47.00 GiB |     146.00 GiB
```

For Qwen3-235B, **128GB RAM allows you to run with context up to \~130k tokens.**

* The estimator lets you plan ahead and avoid out-of-memory errors when loading or using models.
* For more examples and a breakdown of VRAM components, see [docs/vram-estimator.md](docs/vram-estimator.md).

---

## 5. Building Containers Locally

Pre-built toolbox container images are published on Docker Hub for immediate use. If you wish to build the containers yourself (for example, to customize packages or rebuild with a different llama.cpp version), see:

Full instructions: [docs/building.md](docs/building.md).

---

## 6. Host Configuration

This should work on any Strix Halo. For a complete list of available hardware, see: [Strix Halo Hardware Database](https://strixhalo-homelab.d7.wtf/Hardware)

### 6.1 Test Configuration

|                   |                                               |
| ----------------- | --------------------------------------------- |
| **Test Machine**  | HP Z2 Mini G1a                                |
| **CPU**           | Ryzen AI MAX+ 395 "Strix Halo"                |
| **System Memory** | 128 GB RAM                                    |
| **GPU Memory**    | 512 MB allocated in BIOS                      |
| **Host OS**       | Fedora 42, kernel 6.15.6-200.fc42.x86\_86\_64 |

### 6.2 Kernel Parameters (tested on Fedora 42)

Add these these boot parameters to enable unified memory and optimal performance:

```
amd_iommu=off amdgpu.gttsize=131072 ttm.pages_limit=33554432

```
| Parameter                   | Purpose                                                                                  |
| --------------------------- | ----------------------------------------------------------------------------------------- |
| `amd_iommu=off`             | Disables IOMMU for lower latency                                                         |
| `amdgpu.gttsize=131072`     | Enables unified GPU/system memory (up to 128 GiB); 131072 MiB ÷ 1024 = 128 GiB           |
| `ttm.pages_limit=33554432`  | Allows large pinned memory allocations; 33554432 × 4 KiB = 134217728 KiB ÷ 1024² = 128 GiB |

Source: https://www.reddit.com/r/LocalLLaMA/comments/1m9wcdc/comment/n5gf53d/?context=3&utm_source=share&utm_medium=web3x&utm_name=web3xcss&utm_term=1&utm_content=share_button.

**Apply the changes:**

```
# Edit /etc/default/grub to add parameters to GRUB_CMDLINE_LINUX
sudo grub2-mkconfig -o /boot/grub2/grub.cfg
sudo reboot
```

### 6.3 Ubuntu 24.04

Follow this guide by TechnigmaAI for a working configuration on Ubuntu 24.04:

[https://github.com/technigmaai/technigmaai-wiki/wiki/AMD-Ryzen-AI-Max--395:-GTT--Memory-Step%E2%80%90by%E2%80%90Step-Instructions-(Ubuntu-24.04)](https://github.com/technigmaai/technigmaai-wiki/wiki/AMD-Ryzen-AI-Max--395:-GTT--Memory-Step%E2%80%90by%E2%80%90Step-Instructions-%28Ubuntu-24.04%29)

## 7. More Documentation

* [docs/benchmarks.md](docs/benchmarks.md): Full benchmark logs, model list, parsed results  
* [docs/vram-estimator.md](docs/vram-estimator.md): Memory planning, practical example runs  
* [docs/building.md](docs/building.md): Local build, toolbox customization, advanced use  

## 8. References

* The main reference for AMD Ryzen AI MAX home labs, by deseven (there's also a Discord server): [https://strixhalo-homelab.d7.wtf/](https://strixhalo-homelab.d7.wtf/)
* Most comprehesive repostiry of test builds for Strix Halo by lhl -> [https://github.com/lhl/strix-halo-testing/tree/main](https://github.com/lhl/strix-halo-testing/tree/main)
* Ubuntu 24.04 configuration
  [https://github.com/technigmaai/technigmaai-wiki/wiki/AMD-Ryzen-AI-Max--395:-GTT--Memory-Step%E2%80%90by%E2%80%90Step-Instructions-(Ubuntu-24.04)](https://github.com/technigmaai/technigmaai-wiki/wiki/AMD-Ryzen-AI-Max--395:-GTT--Memory-Step%E2%80%90by%E2%80%90Step-Instructions-%28Ubuntu-24.04%29)