Here's my working ROCm setup on a GMKtec NucBox K8 Plus (Ryzen 7 8845HS, Radeon 780M with 16GB shared VRAM), running CachyOS.

**Hardware**: Radeon 780M (HawkPoint1, gfx1100.3 — reported as gfx_target_version 110003)

**Kernel parameters** (in `/etc/default/grub` → `GRUB_CMDLINE_LINUX_DEFAULT`):

```

amd_iommu=on iommu=pt amdgpu.cwsr_enable=0

```

- `amd_iommu=on iommu=pt` — passthrough mode for IOMMU, needed for GPU device access in containers

- `amdgpu.cwsr_enable=0` — disables Compute Static Wave Restore, fixes hangs/crashes with some ROCm workloads on iGPUs

**Environment variable** (in `\~/.zshrc` or `\~/.bashrc` or `/etc/environment`):

```bash

export HSA_OVERRIDE_GFX_VERSION=11.0.2

```

This is the key one — the 780M reports as gfx1100.3 but ROCm doesn't officially support it. Setting `HSA_OVERRIDE_GFX_VERSION=11.0.2` makes it behave as gfx1100.2 (Radeon RX 7900 series) which is fully supported. Without this, most ROCm apps will refuse to run or crash.

**Device permissions**: Make sure your user is in the `render` and `video` groups:

```bash

sudo usermod -aG render,video $USER

```

Devices needed: `/dev/kfd` (kernel fusion driver) and `/dev/dri/renderD128` (render node).

**For Docker/containers** (e.g. Immich ML, Ollama):

```yaml

devices:

- /dev/kfd:/dev/kfd

- /dev/dri:/dev/dri

environment:

- HSA_OVERRIDE_GFX_VERSION=11.0.2

group_add:

- video

- render

```

**ROCm version**: 7.2.1 (via CachyOS repos — includes pytorch-opt-rocm, hipblas, rocblas, miopen-hip, etc.)

**VRAM**: 16GB shared (configurable in BIOS — some NucBox models default to 4GB, you want to set it to max). Check with:

```bash

cat /sys/class/drm/card*/device/mem_info_vram_total

# Should show 17179869184 (16 GB)

```

This setup runs PyTorch with ROCm, Immich ML (photo recognition), and various LLM inference workloads without issues. The 780M is surprisingly capable for an iGPU — roughly equivalent to a discrete GTX 1650 in ML benchmarks.