diff --git a/.gitignore b/.gitignore
new file mode 100644
index 0000000..85e7c1d
--- /dev/null
+++ b/.gitignore
@@ -0,0 +1 @@
+/.idea/
diff --git a/Dockerfile b/Dockerfile
new file mode 100644
index 0000000..8e40cc9
--- /dev/null
+++ b/Dockerfile
@@ -0,0 +1,44 @@
+# ROCm-enabled PyTorch base image from AMD
+# Pick a tag that matches your installed ROCm driver best; this one is a good default.
+FROM rocm/pytorch:rocm7.1.1_ubuntu22.04_py3.10_pytorch_release_2.9.1
+
+# OCI metadata for discoverability
+LABEL org.opencontainers.image.source="https://github.com/zuzupebbles/unsloth-rocm"
+LABEL org.opencontainers.image.url="https://hub.docker.com/r/zuzupebbles/unsloth-rocm"
+LABEL org.opencontainers.image.title="Unsloth ROCm image"
+LABEL org.opencontainers.image.description="Unsloth + ROCm PyTorch for AMD GPUs (e.g. 7900 XTX)"
+LABEL org.opencontainers.image.licenses="Apache-2.0"
+
+# Basic utilities
+RUN apt-get update && \
+    DEBIAN_FRONTEND=noninteractive apt-get install -y \
+        git wget curl vim nano && \
+    rm -rf /var/lib/apt/lists/*
+
+# Upgrade pip and install Unsloth (AMD is supported via regular pip install)
+# Unsloth itself is Apache-2.0 licensed.
+RUN pip install --upgrade pip && \
+    pip install --no-cache-dir unsloth
+
+# Optional: tools you’ll almost certainly want when experimenting
+RUN pip install --no-cache-dir \
+      "transformers>=4.45.0" \
+      datasets \
+      accelerate \
+      peft \
+      safetensors \
+      bitsandbytes  || true
+# ^ bitsandbytes may be CUDA-only; failing here is fine, Unsloth on ROCm does not strictly need it.
+
+# Create a non-root user that matches your host UID/GID at runtime (set via docker-compose)
+ARG USERNAME=unsloth
+ARG UID=1000
+ARG GID=1000
+RUN groupadd -g ${GID} ${USERNAME} && \
+    useradd -m -u ${UID} -g ${GID} -s /bin/bash ${USERNAME}
+
+USER ${USERNAME}
+WORKDIR /workspace
+
+# Keep the container alive by default; you’ll exec into it
+CMD ["bash"]
diff --git a/README.md b/README.md
index e69de29..1cf295c 100644
--- a/README.md
+++ b/README.md
@@ -0,0 +1,100 @@
+# unsloth-rocm 🦥🔥
+
+Unsloth + PyTorch-on-ROCm in a ready-to-use Docker image for AMD GPUs  
+(Navi / RDNA2 / RDNA3 – e.g. 7900 XTX).
+
+- Base image: [`rocm/pytorch`](https://hub.docker.com/r/rocm/pytorch) (PyTorch with ROCm backend)   
+- Fine-tuning framework: [Unsloth](https://github.com/unslothai/unsloth) (Apache-2.0). 
+
+This repo contains:
+
+- `Dockerfile` – how the image is built
+- `docker-compose.build.yml` – a sample compose file for local build + dev use
+
+The idea is to give ROCm users a “batteries-included” Unsloth container and a clean example of how to wire up devices / groups / HF cache.
+
+---
+
+## Requirements
+
+On the host (e.g. Arch, Ubuntu, etc.):
+
+- AMD GPU supported by ROCm / `rocm/pytorch` (MI* or newer Navi, see AMD’s support matrix)   
+- ROCm stack installed and working (`rocminfo` sees your GPU)
+- Docker or compatible container runtime
+- User is in the appropriate groups, e.g.:
+
+```bash
+  sudo usermod -aG video,docker $USER
+  # log out & back in afterwards
+```
+
+You should also see `/dev/kfd` and `/dev/dri` on the host:
+
+```bash
+ls /dev/kfd
+ls /dev/dri
+```
+
+---
+
+## Build locally
+
+Clone the repo (whichever remote you like – GitHub or your local Forgejo):
+
+```bash
+git clone https://github.com/zuzupebbles/unsloth-rocm.git
+cd unsloth-rocm
+```
+
+Build the image:
+
+```bash
+# ensure UID/GID match your host user (especially on Arch)
+export UID
+export GID
+
+docker compose -f docker-compose.build.yml build
+```
+
+This will produce a local image named:
+
+```text
+zuzupebbles/unsloth-rocm:local
+```
+
+You can confirm with:
+
+```bash
+docker images | grep unsloth-rocm
+```
+
+---
+
+## Run for local dev
+
+Bring up a dev container:
+
+```bash
+docker compose -f docker-compose.build.yml up -d
+docker compose -f docker-compose.build.yml exec unsloth-rocm bash
+```
+
+Inside the container you’ll have:
+
+* user: `unsloth`
+* working dir: `/workspace`
+* volumes mounted:
+
+    * `./workspace` → `/workspace`
+    * `~/.cache/huggingface` → `/data/hf-cache`
+    * `./repos-clone` → `/repos-clone`
+
+Quick sanity check in the container:
+
+```bash
+python -c "import torch; print(torch.cuda.is_available(), torch.version.cuda, torch.version.hip, torch.version.git_version)"
+python -c "import unsloth; print(unsloth.__version__)"
+```
+
+(ROCm builds often report HIP instead of CUDA, that’s expected.)
diff --git a/docker-compose.build.yml b/docker-compose.build.yml
new file mode 100644
index 0000000..374de58
--- /dev/null
+++ b/docker-compose.build.yml
@@ -0,0 +1,52 @@
+services:
+  unsloth-rocm:
+    build:
+      context: .
+      dockerfile: Dockerfile
+      args:
+        UID: ${UID:-1000}
+        GID: ${GID:-1000}
+
+    # Image name for local use; you can retag this for Docker Hub.
+    image: zuzupebbles/unsloth-rocm:local
+
+    container_name: unsloth-rocm
+    restart: unless-stopped
+
+    # ROCm device plumbing – basically AMD's recommended set.
+    devices:
+      - /dev/kfd
+      - /dev/dri
+    group_add:
+      - "video"   # host group that has access to /dev/kfd and /dev/dri
+      # - "render"  # optional, if your distro uses it
+
+    cap_add:
+      - SYS_PTRACE
+    security_opt:
+      - seccomp=unconfined
+    ipc: host
+    shm_size: "16g"
+
+    environment:
+      # If you only want GPU 0, set explicitly:
+      - ROCM_VISIBLE_DEVICES=0
+
+      # HF cache so you don't redownload models all the time
+      - HF_HOME=/data/hf-cache
+      - HUGGINGFACE_HUB_CACHE=/data/hf-cache
+      - TRANSFORMERS_CACHE=/data/hf-cache
+
+    volumes:
+      # Workspace for notebooks / scripts
+      - ./workspace:/workspace
+
+      # Hugging Face cache on host
+      - ~/.cache/huggingface:/data/hf-cache
+
+      # Optional: local repo clones or model sources
+      - ./repos-clone:/repos-clone
+
+    working_dir: /workspace
+    stdin_open: true
+    tty: true