ODS V3 Pre-Release: Public testing and refinement ahead of the official V3 launch. Turn your PC, Mac, or Linux box into a private AI server.
Python ai-agentsamdcomfyuidockerllama-cppllmlocal-ain8nnvidiaopen-webuiragself-hostedspeech-to-textstrix-halotext-to-speechworkflow-automation
# ODS — AMD ROCm overlay (upstream llama.cpp, HIP)
# Layered after docker-compose.amd.yml when AMD_INFERENCE_BACKEND=rocm. It
# swaps in the official ggml-org ROCm image of the same b9014 release and adds
# the ROCm compute device; /dev/dri and the video and render groups come from
# docker-compose.amd.yml.
#
# The image is built for gfx908, gfx90a, gfx942, gfx1030, gfx1100, gfx1101,
# gfx1102, gfx1150, gfx1151, gfx1200 and gfx1201. Instinct (CDNA) cards need
# this overlay because they have no Vulkan; on every other AMD GPU it is
# opt-in. The image is about 7 GB compressed, against about 0.2 GB for Vulkan.
#
# Devices, groups and ROCBLAS_USE_HIPBLASLT=0 match the ROCm setup ODS ran on
# Strix Halo before Lemonade. That setup also forced HSA_OVERRIDE_GFX_VERSION
# to 11.5.1; this image is built for gfx1151, so the installer sets the
# override only for GPUs outside the list above.
#
# Use with: docker compose -f docker-compose.base.yml -f docker-compose.amd.yml \
# -f docker-compose.amd-rocm.yml up -d
services:
llama-server:
image: ghcr.io/ggml-org/llama.cpp:server-rocm-b9014@sha256:68403f82fe496302bb1c681bab2ee569a04cc13da8fcf14456376b485562236d
devices:
- /dev/kfd:/dev/kfd
environment:
# Passed through only when .env sets it. The installer sets it only for
# GPUs the image was not built for; ROCm rejects an empty value, and a
# wrong one can hang the GPU, so there is no default.
- HSA_OVERRIDE_GFX_VERSION
- ROCBLAS_USE_HIPBLASLT=${ROCBLAS_USE_HIPBLASLT:-0}
dashboard-api:
environment:
- AMD_INFERENCE_BACKEND=${AMD_INFERENCE_BACKEND:-rocm}