# crane-serve containers, one service per GPU backend, selected by profile.
#
#   COMPOSE_PROFILES=rocm MODEL_DIR=~/models MODEL=Qwen3-4B docker compose up -d
#
# Or put the variables in a `.env` file next to this one and just run
# `docker compose up -d`. With no profile selected, nothing starts.
#
# MODEL is a model directory (or single .gguf file) relative to MODEL_DIR.
# CRANE_PORT overrides the published host port (default 8080).
#
# ROCm micro-benchmarks (profile `rocm-bench`, not started by `up` under `rocm`):
#
#   docker compose run --rm crane-bench-rocm                      # both, defaults
#   docker compose run --rm crane-bench-rocm gdn_bench 16 512 128 128 100
#   docker compose run --rm crane-bench-rocm topk_bench 248320 40 200
#
# `<<:` merges keys, not lists or nested maps: a service that sets its own
# `volumes` or `environment` must repeat the anchor's entries.

# crane-serve settings shared by every backend.
x-crane-serve: &crane-serve
  # `[::]` listens on IPv4 and IPv6. The default 0.0.0.0 is IPv4-only, and
  # rootless podman (pasta) forwards host `localhost` (::1) as IPv6, which
  # would then be refused inside the container.
  command: ["--model-path", "/models/${MODEL:?set MODEL to a path under MODEL_DIR}", "--host", "[::]", "--port", "8080"]
  ports:
    - "${CRANE_PORT:-8080}:8080"
  volumes:
    - ${MODEL_DIR:?set MODEL_DIR to the host model directory}:/models:ro
  restart: unless-stopped

# AMD ROCm GPU access, Strix Halo (gfx1151). See docker/rocm/Dockerfile.
# Do NOT set HSA_OVERRIDE_GFX_VERSION: gfx1151 is supported natively.
x-rocm: &rocm
  build:
    context: .
    dockerfile: docker/rocm/Dockerfile
  devices:
    - /dev/dri:/dev/dri
    - /dev/kfd:/dev/kfd
  # Host GIDs (render=105, video=39 on this host); names would be resolved
  # against the container's /etc/group instead. Check with `getent group`.
  group_add:
    - "105"
    - "39"
  security_opt:
    - seccomp:unconfined
    - apparmor:unconfined
    - label=disable

services:
  crane-serve-rocm:
    <<: [*crane-serve, *rocm]
    profiles: [rocm]
    build:
      context: .
      dockerfile: docker/rocm/Dockerfile
      target: runtime
    image: localhost/crane-serve:rocm10
    environment:
      HSA_ENABLE_SDMA: "0"
      # rocBLAS profile logging (dumps a GEMM-shape summary to stderr on
      # exit). Diagnostic only; drop it for quieter logs.
      ROCBLAS_LAYER: "4"
      # Passed through from the shell / .env when set:
      CRANE_PROF:          # forward-pass submission vs GPU time, to stderr
      CRANE_TOPK_HOST:     # 1 = host sort instead of the GPU top-k kernel
      CRANE_GDN_PORTABLE:  # 1 = op-by-op GDN recurrence instead of the fused kernel
    volumes:
      - ${MODEL_DIR:?set MODEL_DIR to the host model directory}:/models:ro
      # hipcc-compiled kernel cache; the first start compiles for minutes.
      - candle-rocm-cache:/var/cache/candle-rocm

  # gdn_bench [BH S K V iters]: fused GDN recurrence kernel in isolation
  #   (defaults 16 512 128 128 100, Qwen3.5-0.8B single-sequence prefill).
  # topk_bench [N K iters]: GPU top-k kernel vs host sort
  #   (defaults 248320 40 200, Qwen3.5 vocab sampling).
  crane-bench-rocm:
    <<: *rocm
    profiles: [rocm-bench]
    build:
      context: .
      dockerfile: docker/rocm/Dockerfile
      target: bench
    image: localhost/crane-bench:rocm10
    environment:
      HSA_ENABLE_SDMA: "0"
      # 1 makes topk_bench's "gpu" arm fall back to the host sort (A/B check).
      # gdn_bench calls the fused kernel directly, so CRANE_GDN_PORTABLE and
      # CRANE_PROF only affect crane-serve.
      CRANE_TOPK_HOST:
    volumes:
      - candle-rocm-cache:/var/cache/candle-rocm

volumes:
  candle-rocm-cache: