YOUR MODELS. YOUR MACHINE.
Local AI, without the
constant troubleshooting.
Run video, images, music, voice and language models from one workspace. Rocsy coordinates the GPU so you can line up your next idea while the current one renders.
A practical home for
AMD-powered AI.
Work stays local
Keep model files, datasets and outputs on your host.
One front door
A shared login and workspace for your AI services.
A GPU that takes turns
Coordinate workloads instead of racing for VRAM.
THE CORE OF ROCSY
Know what your GPU is doing.
See the running job, inspect what comes next, and handle a failed job with useful diagnostic information.
A clean handoff
A busy GPU returns a lock response. The next studio waits instead of starting over the active job.
Recover after a crash
Locks have a time limit. An abandoned lock can expire instead of leaving the workspace stuck.
Less babysitting
Track running and waiting jobs in one place, with push notifications when work finishes.
CREATE / RUN / MANAGE
A workspace for the whole workflow.
Focused interfaces over shared infrastructure. Search a capability or model, then open a studio’s details.
No studios match that search. Try “video”, “LoRA” or “voice”.
UNDER THE HOOD
Shared infrastructure. Fewer moving parts.
Containers coordinate. The host computes.
GPU-heavy backends run natively on Linux. The services around them handle jobs, media operations and the shared gateway.
Your files remain accessible.
Outputs, logs and datasets are bind-mounted on the host. Downloads feed a shared model library instead of separate copies for every studio.
Built around real AMD hardware.
Developed on a 24 GB RX 7900 XTX, with an RDNA-tuned llama.cpp build and per-model inference controls. Other hardware needs its own compatibility check.
SELF-HOSTING
Bring your GPU. Keep control.
Rocsy is a self-hosted workspace. It grew out of getting local generation running reliably on a Radeon card.
- Linux + a compatible AMD GPUROCm support and VRAM requirements depend on your card and models.
- Docker Compose + host backendsComfyUI and llama.cpp run alongside coordinating services.
- Storage for models and outputsVideo models and training datasets can need substantial disk space.
How the stack starts
With the service source and host dependencies installed, configure the environment and start the services:
cp .env.example .env
docker compose up -d --buildSet your gateway password and service secrets first. This is an overview, not a complete installation guide.
Optional LoRA training
The Trainer uses a separate Compose profile. Its ROCm image is large, and training shares the GPU’s memory budget.
docker compose --profile trainer up -d --build trainerNetwork access and local data
Model downloads, remote access and optional integrations may use external services. Local inference and host storage stay under your control.
BUILT BY BITTYBUNNYBYTES
One developer. One very busy Radeon.
Rocsy started with a month of ROCm setup and compatibility fixes. The goal: more time creating, less time managing backends.