# The same trader on an AMD GPU (ROCm). Layer it over the base file: # # docker compose -f docker-compose.yml -f docker-compose.rocm.yml up -d --build # # Host prerequisites: the amdgpu kernel driver, nothing else. The ROCm 7.2 runtime ships inside the torch wheel (built # for gfx1030, gfx1100-1103, gfx1150/1151 -- Ryzen AI Max -- gfx1200/1201 and the Instinct cards), so neither ROCm on # the host nor a ROCm base image is needed. The container reaches the GPU through two device nodes: /dev/kfd (compute) # and /dev/dri (the render nodes). It runs as root (the base file), which opens them without the host's render group. # DEVICE=auto picks the GPU when the container sees one. An integrated GPU draws on system memory: DEVICE_MEMORY_GB caps # the batch sizes derived from what the driver reports. services: fly: build: args: TORCH_INDEX: rocm72 environment: DEVICE: auto devices: - /dev/kfd - /dev/dri