docker-compose.amd.yml

Rama main · réplica de lectura del repositorio original.

services:
  ollama:
    build:
      args:
        OLLAMA_BASE: ollama/ollama:0.40.10-rocm
    devices:
      - /dev/kfd
      - /dev/dri
    group_add:
      - "${VIDEO_GID:-video}"
      - "${RENDER_GID:-render}"
    environment:
      - HSA_OVERRIDE_GFX_VERSION
      - HIP_VISIBLE_DEVICES
      - ROCR_VISIBLE_DEVICES
      - OLLAMA_SCHED_SPREAD

  llamacpp:
    image: ghcr.io/ggml-org/llama.cpp:server-vulkan@sha256:a644fb651975d9dea49140302577a3b945993fd5efd3400bf89c7d525d03c103
    devices:
      - /dev/kfd
      - /dev/dri
    group_add:
      - "${VIDEO_GID:-video}"
      - "${RENDER_GID:-render}"
    environment:
      - GGML_VK_VISIBLE_DEVICES
      - LLAMA_ARG_TENSOR_SPLIT
      - LLAMA_ARG_MAIN_GPU

  # only the karaoke sidecar gets a GPU. the voice sidecar stays
  # CPU-only by design, see docker/tts.Dockerfile
  karaoke:
    build:
      args:
        TORCH_INDEX: "${KARAOKE_TORCH_INDEX:-https://download.pytorch.org/whl/rocm6.2.4}"
    devices:
      - /dev/kfd
      - /dev/dri
    group_add:
      - "${VIDEO_GID:-video}"
      - "${RENDER_GID:-render}"
    environment:
      - HSA_OVERRIDE_GFX_VERSION