# ODS — AMD Multi-GPU Overlay
# Applied when GPU_COUNT > 1 and GPU_BACKEND = amd.
# llama-server gets all of /dev/dri + ROCR_VISIBLE_DEVICES for multi-GPU inference.
# Single-GPU services use ROCR_VISIBLE_DEVICES for GPU isolation.
# Dashboard-api gets all GPUs for monitoring.
#
# Lemonade passthrough: Lemonade launches its inner llama-server via fork()+execv()
# and constructs the CLI explicitly — LLAMA_ARG_* env vars are NOT guaranteed to be
# read by the inner process. We pass --split-mode explicitly via --llamacpp-args
# (Lemonade's official mechanism). --split-mode and --tensor-split are NOT in
# Lemonade's blocked flag list (-m, --port, --ctx-size, -ngl, --jinja, --mmproj,
# --embeddings, --reranking), so they pass through safely.
#
# Backend selection: Lemonade's llamacpp.backend field accepts only auto|vulkan|cpu
# (the "rocm" literal was tightened out of the schema upstream). We set "auto" and
# rely on LEMONADE_LLAMACPP_ROCM_BIN (inherited from docker-compose.amd.yml) to
# point Lemonade at the custom ROCm llama-server binary.
# Note: command fully replaces the base docker-compose.amd.yml command.

services:
  llama-server:
    # devices and group_add inherited from docker-compose.amd.yml
    command:
      - serve
      - --port
      - "8080"
      - --host
      - "0.0.0.0"
      - --no-tray
      - --llamacpp
      - "${LEMONADE_LLAMACPP:-auto}"
      - --llamacpp-args
      - "--metrics --host 0.0.0.0 --split-mode ${LLAMA_ARG_SPLIT_MODE:-layer}"
      - --extra-models-dir
      - /models
    environment:
      LEMONADE_LLAMACPP: "${LEMONADE_LLAMACPP:-auto}"
      ROCR_VISIBLE_DEVICES: "${LLAMA_SERVER_GPU_INDICES:-}"
      # Tensor split ratio (e.g. "0.5,0.5") — passed via env var because it may
      # be empty (auto-split). Inherited by inner llama-server via fork()+execv().
      LLAMA_ARG_TENSOR_SPLIT: "${LLAMA_ARG_TENSOR_SPLIT:-}"
      HSA_OVERRIDE_GFX_VERSION: "${HSA_OVERRIDE_GFX_VERSION:-}"

  dashboard-api:
    devices:
      - /dev/kfd:/dev/kfd
      - /dev/dri:/dev/dri
    environment:
      GPU_COUNT: "${GPU_COUNT:-1}"
      GPU_ASSIGNMENT_JSON_B64: "${GPU_ASSIGNMENT_JSON_B64:-}"
    # volumes (sysfs mounts) inherited from docker-compose.amd.yml