Proxmox 9.x / Strix Halo / GPU Passthrough

dlasher · Mar 3, 2026

Having spent the last few days fighting all the subtle parts of getting this working, I put together a quick guide on how to get Proxmox, running on a Strix Halo, machine, with working GPU passthrough into LXC containers. To be clear, this is a "works right now" recipe, subject to kernel changes, ROCM changes, etc.

---------------------------------

Guide: The Strix Halo AI Powerhouse (Proxmox 9.1 + ROCm 7.2)

Target Hardware: AMD STRIX HALO box (Minisforum S1-MAX etc)
Goal: 128GB total RAM, 64GB ram for CPU/applications + 64GB VRAM local AI server with full hardware acceleration

Introduction

The AMD Strix Halo (RDNA 3.5 / gfx1151) is a game-changer for local AI. By leveraging a high-speed unified memory architecture, this APU can address massive amounts of system RAM as video memory. This guide details how to configure Proxmox 9.1 to carve out a 64GB VRAM pool and pass it through to a high-performance LXC container.

Phase 1: Host BIOS & Kernel Tuning

Unlocking the memory gates to allow the GPU to access 64GB of RAM.

1. BIOS Settings

IOMMU: Enabled.
UMA Framebuffer: Auto. (The kernel parameters below will override and expand this).
Resizable BAR: Enabled.

2. Host Kernel Parameters

Edit /etc/default/grub (or /etc/kernel/cmdline) on your Proxmox Host to enable IOMMU pass-through and define the Graphics Translation Table (GTT) size.

Code:

# Edit this line in /etc/default/grub
GRUB_CMDLINE_LINUX_DEFAULT="quiet splash amd_iommu=on iommu=pt amdgpu.gttsize=65536 ttm.pages_limit=16777216 video=1024x768@60"

amdgpu.gttsize=65536: Maps 64GB of system RAM for GPU use.
ttm.pages_limit=16777216: Sets the page limit to exactly 64GB (16777216 times 4096 byte pages).
video=1024x768@60: Because console graphics mode autosense is always wrong.

Apply and Reboot:

Code:

update-grub && reboot

Phase 2: Host Driver & Firmware Installation

Strix Halo requires the latest firmware blobs and the ROCm 7.2 userspace stack.

1. Update Firmware (Critical for gfx1151)

Run this on the Proxmox Host:

Code:

apt update && apt install -y git
git clone --depth 1 https://git.kernel.org/pub/scm/linux/kernel/git/firmware/linux-firmware.git
cp linux-firmware/amdgpu/gc_11_5_1* /lib/firmware/amdgpu/
cp linux-firmware/amdgpu/sdma_6_1_1* /lib/firmware/amdgpu/
update-initramfs -u

2. Install ROCm 7.2 Userspace (Host)

Code:

apt update && apt install -y wget gpg curl
wget -q https://repo.radeon.com/amdgpu-install/latest/ubuntu/noble/amdgpu-install_7.2.70200-1_all.deb -O /tmp/amdgpu.deb
apt install -y /tmp/amdgpu.deb
amdgpu-install -y --usecase=graphics,rocm
usermod -aG render,video root

3. Host Verification

Confirm the host sees the hardware:

Code:

/usr/bin/rocminfo | grep "gfx1151"

Phase 3: LXC Creation & Hardware Mapping

1. Create the Container

Privileged: Yes (Required for the KFD driver handshake).
Template: Ubuntu 24 LTS
RAM: 16GB (The GPU will pull from the 64GB GTT pool independently).

2. Map Devices (On Host)

Run these commands on the Proxmox Host to find your GIDs and map the hardware to your LXC (replace 1201 with your actual Container ID):

Code:

# Find GIDs
RENDER_GID=$(getent group render | cut -d: -f3)
VIDEO_GID=$(getent group video | cut -d: -f3)

# Native Proxmox device passthrough
pct set 1201 -dev0 /dev/kfd,gid=$RENDER_GID
pct set 1201 -dev1 /dev/dri/renderD128,gid=$RENDER_GID
pct set 1201 -dev2 /dev/dri/card0,gid=$VIDEO_GID

Phase 4: Container Internal Setup

Inside the Ubuntu 24 LTS LXC, install the ROCm stack without kernel modules (no-dkms) and configure Ollama.

1. Install ROCm 7.2 (LXC)

Code:

apt update && apt install -y wget gpg curl zstd
wget -q 'https://repo.radeon.com/amdgpu-install/latest/ubuntu/noble/amdgpu-install_7.2.70200-1_all.deb' -O /tmp/amdgpu.deb
apt install -y /tmp/amdgpu.deb
amdgpu-install -y --usecase=rocm --no-dkms
usermod -aG render,video root

2. Install & Reconfigure Ollama

Code:

curl -fsSL https://ollama.com/install.sh | sh
systemctl edit ollama.service

Paste the following into the override file:

Code:

[Service]
# Force RDNA 3.5 recognition
Environment="HSA_OVERRIDE_GFX_VERSION=11.5.0"

# Stability Fix: Disable bugged SDMA for unified memory
Environment="HSA_ENABLE_SDMA=0"
Environment="OLLAMA_VULKAN=1"

# Connectivity & Performance
Environment="OLLAMA_HOST=0.0.0.0"
Environment="OLLAMA_ORIGINS=*"
Environment="OLLAMA_NUM_PARALLEL=2"
Environment="OLLAMA_KV_CACHE_TYPE=q8_0"
Environment="OLLAMA_KEEP_ALIVE=24h"

Phase 5: Functional Validation

Perform these checks inside the LXC to ensure the stack is operational.

1. Driver Check

Code:

/usr/bin/rocminfo | grep "gfx1151"

Expected Output: Name: gfx1151 and Name: amdgcn-amd-amdhsa--gfx1151.

2. Functional LLM Test

Code:

systemctl daemon-reload
systemctl restart ollama
ollama pull qwen2.5:0.5b
ollama run qwen2.5:0.5b "Why is the sky blue?"

If the reply is instant and doesn't crash, your 64G/64G Strix Halo workstation is live.

dlasher · Mar 3, 2026

OPTIONAL:

If you don't want to blind copy the firmware over the top, or you might actually want to check if you're ALREADY on the latest firmware, use this script instead.

Code:

#!/bin/bash
# pve-firmware-sync.sh

# 1. Pull latest firmware blobs from upstream
apt update && apt install -y git
git clone --depth 1 https://git.kernel.org/pub/scm/linux/kernel/git/firmware/linux-firmware.git /tmp/linux-firmware

# Track if ANY file across all patterns was updated
ANY_UPDATED=0

sync_firmware() {
    local src_pattern=$1
    local dest_dir="/lib/firmware/amdgpu/"
    local changed_in_this_run=0

    # Expand the pattern into a list of files
    for src_file in $src_pattern; do
        filename=$(basename "$src_file")
        dest_file="${dest_dir}${filename}"

        # If dest doesn't exist OR hashes differ, copy it
        if [[ ! -f "$dest_file" ]] || [[ "$(md5sum < "$src_file")" != "$(md5sum < "$dest_file")" ]]; then
            echo "Updating: $filename"
            cp "$src_file" "$dest_file"
            changed_in_this_run=1
            ANY_UPDATED=1
        else
            echo "Skipping: $filename (already up to date)"
        fi
    done
}

# 2. Perform the sync
sync_firmware "/tmp/linux-firmware/amdgpu/gc_11_5_1*"
sync_firmware "/tmp/linux-firmware/amdgpu/sdma_6_1_1*"

# 3. Only rebuild if ANY_UPDATED was set to 1
if [ "$ANY_UPDATED" -eq 1 ]; then
    echo "------------------------------------------------"
    echo "Changes detected. Rebuilding initramfs..."
    update-initramfs -u
    echo "Reboot recommended to apply new firmware."
else
    echo "------------------------------------------------"
    echo "Firmware is already synchronized. No rebuild required."
fi

# Cleanup
rm -rf /tmp/linux-firmware

# bash ./sync.firmware.sh
Hit:1 http://ftp.us.debian.org/debian trixie InRelease
Hit:2 http://security.debian.org trixie-security InRelease
Hit:3 http://ftp.us.debian.org/debian trixie-updates InRelease
Hit:4 https://repo.radeon.com/amdgpu/30.30/ubuntu noble InRelease
Hit:5 http://download.proxmox.com/debian/pve trixie InRelease
Hit:6 https://repo.radeon.com/rocm/apt/7.2 noble InRelease
Hit:7 https://repo.radeon.com/graphics/7.2/ubuntu noble InRelease
All packages are up to date.
git is already the newest version (1:2.47.3-0+deb13u1).
Summary:
Upgrading: 0, Installing: 0, Removing: 0, Not Upgrading: 0
Cloning into '/tmp/linux-firmware'...
remote: Enumerating objects: 4286, done.
remote: Counting objects: 100% (4286/4286), done.
remote: Compressing objects: 100% (2969/2969), done.
remote: Total 4286 (delta 1672), reused 3272 (delta 1190), pack-reused 0 (from 0)
Receiving objects: 100% (4286/4286), 721.90 MiB | 24.74 MiB/s, done.
Resolving deltas: 100% (1672/1672), done.
Updating files: 100% (4462/4462), done.
Skipping: gc_11_5_1_imu.bin (already up to date)
Skipping: gc_11_5_1_me.bin (already up to date)
Skipping: gc_11_5_1_mec.bin (already up to date)
Skipping: gc_11_5_1_mes1.bin (already up to date)
Skipping: gc_11_5_1_mes_2.bin (already up to date)
Skipping: gc_11_5_1_pfp.bin (already up to date)
Skipping: gc_11_5_1_rlc.bin (already up to date)
Skipping: sdma_6_1_1.bin (already up to date)
------------------------------------------------
Firmware is already synchronized. No rebuild required.

kriz · Monday at 12:50

Just in case anyone else stumbles across this.
The instructions worked for me. My notes are attached below:

Phase 1:

1 BIOS Settings
-
2. Host Kernel Parameters

I read that the amdgpu.gttsize parameter was deprecated on the Strix Halo Wiki (referenced depracation message)
Just ommitting it worked:

Code:

# Edit this line in /etc/default/grub
GRUB_CMDLINE_LINUX_DEFAULT="quiet splash amd_iommu=on iommu=pt  ttm.pages_limit=25165824 video=1024x768@60"

This was the output on my Minisforum MS-S1 Max:

Code:

root@pve:~# dmesg | grep -i amdgpu | grep memory
[    3.639795] amdgpu 0000:bd:00.0:  512M of VRAM memory ready
[    3.639797] amdgpu 0000:bd:00.0:  98304M of GTT memory ready.

Phase 2:

1. Update Firmware
My kernel is/was: Linux pve 7.0.0-3-pve #1 SMP PREEMPT_DYNAMIC PMX 7.0.0-3 (2026-04-21T22:56Z) x86_64 GNU/Linux
I did not have to update the firmware mentioned in phase 2

2 Install ROCm 7.2
Rocm latest version is 7.2.2 so a small url update is needed. There was a sed cmd to downgrade a part to 7.2.1 (from what i understand, I do not know what I am doing) on the AMD website so i went with those commands:

Code:

apt update &amp;&amp; apt install -y wget gpg curl
wget https://repo.radeon.com/amdgpu-install/7.2.2/ubuntu/noble/amdgpu-install_7.2.2.70202-1_all.deb
 apt install ./amdgpu-install_7.2.2.70202-1_all.deb
sed -i "s|graphics/7.2.2|graphics/7.2.1|" /etc/apt/sources.list.d/rocm.list
apt update
amdgpu-install -y --usecase=graphics,rocm
usermod -aG render,video root

I got an error with amdgpu-install -y --usecase=graphics,rocm :

Code:

root@pve:~# amdgpu-install  --usecase=graphics,rocm
INFO: i386 architecture has not been enabled with dpkg.
Installation of 32-bit run time has been excluded.
Hit:1 http://deb.debian.org/debian trixie InRelease
Hit:2 http://deb.debian.org/debian trixie-updates InRelease
Hit:3 https://repo.radeon.com/amdgpu/30.30.2/ubuntu noble InRelease
Hit:4 http://security.debian.org/debian-security trixie-security InRelease
Hit:5 https://repo.radeon.com/rocm/apt/7.2.2 noble InRelease
Hit:6 https://repo.radeon.com/graphics/7.2.1/ubuntu noble InRelease
Hit:7 http://download.proxmox.com/debian/pve trixie InRelease
Reading package lists... Done
Reading package lists... Done
Building dependency tree... Done
Reading state information... Done
amdgpu-lib is already the newest version (1:7.2.70201-2303469.24.04).
rocm is already the newest version (7.2.2.70202-86~24.04).
amdgpu-dkms is already the newest version (1:6.16.13.30300200-2317211.24.04).
0 upgraded, 0 newly installed, 0 to remove and 1 not upgraded.
WARNING: amdgpu dkms failed for running kernel

According to claude/chatgpt the Kernel ist too new but the amdgpu DKMS module is apparently not needed, the driver that comes with the kernel is enough.

Code:

# On the HOST — just verify the in-tree driver works:
/opt/rocm/bin/rocminfo | grep "gfx1151"
If your in-tree amdgpu driver (from kernel 7.0.0-3-pve) already exposes /dev/kfd and /dev/dri/renderD128, you're good. Check with:

bash
ls -la /dev/kfd /dev/dri/
If those devices exist, skip the DKMS step on the host and go straight to the LXC setup using --no-dkms:

bash
# Inside the LXC container only:
amdgpu-install -y --usecase=rocm --no-dkms
The container doesn't need DKMS at all — it just needs the ROCm userspace libraries.

Phases 3-5 just "worked" at least I thought they did. I then ran ollama ps which showed:

Code:

root@llmtest:~# ollama ps
NAME            ID              SIZE      PROCESSOR    CONTEXT    UNTIL
qwen2.5:0.5b    a8b0c5157701    492 MB    100% CPU     4096       23 hours from now

The group ids of the render group in the lxc container and on the host where not identical. I changed the groups in the lxc container, rebooted it, and got a quicker answer and:

Code:

root@llmtest:~# ollama ps
NAME            ID              SIZE      PROCESSOR    CONTEXT    UNTIL
qwen2.5:0.5b    a8b0c5157701    3.2 GB    100% GPU     32768      24 hours from now

Thank you very much for the instructions.

Search

Search

Proxmox 9.x / Strix Halo / GPU Passthrough

dlasher

Renowned Member

Guide: The Strix Halo AI Powerhouse (Proxmox 9.1 + ROCm 7.2)

Introduction

Phase 1: Host BIOS & Kernel Tuning

1. BIOS Settings

2. Host Kernel Parameters

Phase 2: Host Driver & Firmware Installation

1. Update Firmware (Critical for gfx1151)

2. Install ROCm 7.2 Userspace (Host)

3. Host Verification

Phase 3: LXC Creation & Hardware Mapping

1. Create the Container

2. Map Devices (On Host)

Phase 4: Container Internal Setup

1. Install ROCm 7.2 (LXC)

2. Install & Reconfigure Ollama

Phase 5: Functional Validation

1. Driver Check

2. Functional LLM Test

dlasher

Renowned Member

kriz

New Member

We value your privacy

Proxmox 9.x / Strix Halo / GPU Passthrough

dlasher

Renowned Member

Guide: The Strix Halo AI Powerhouse (Proxmox 9.1 + ROCm 7.2)​

Introduction​

Phase 1: Host BIOS & Kernel Tuning​

1. BIOS Settings​

2. Host Kernel Parameters​

Phase 2: Host Driver & Firmware Installation​

1. Update Firmware (Critical for gfx1151)​

2. Install ROCm 7.2 Userspace (Host)​

3. Host Verification​

Phase 3: LXC Creation & Hardware Mapping​

1. Create the Container​

2. Map Devices (On Host)​

Phase 4: Container Internal Setup​

1. Install ROCm 7.2 (LXC)​

2. Install & Reconfigure Ollama​

Phase 5: Functional Validation​

1. Driver Check​

2. Functional LLM Test​

​

dlasher

Renowned Member

kriz

New Member

We value your privacy

Guide: The Strix Halo AI Powerhouse (Proxmox 9.1 + ROCm 7.2)

Introduction

Phase 1: Host BIOS & Kernel Tuning

1. BIOS Settings

2. Host Kernel Parameters

Phase 2: Host Driver & Firmware Installation

1. Update Firmware (Critical for gfx1151)

2. Install ROCm 7.2 Userspace (Host)

3. Host Verification

Phase 3: LXC Creation & Hardware Mapping

1. Create the Container

2. Map Devices (On Host)

Phase 4: Container Internal Setup

1. Install ROCm 7.2 (LXC)

2. Install & Reconfigure Ollama

Phase 5: Functional Validation

1. Driver Check

2. Functional LLM Test