ACE-Step-1.5/docker-compose.jetson.yml at main · ace-step/ACE-Step-1.5 · GitHub

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
# =============================================================================
# ACE-Step 1.5 — Docker Compose for NVIDIA Jetson
# =============================================================================
#
# Prerequisites:
#   - JetPack 6.x (L4T R36.x) with NVIDIA Container Runtime
#   - Docker Compose v2 (docker compose) or v1 (docker-compose)
#
# Usage:
#   # Start Gradio UI (default):
#   docker compose -f docker-compose.jetson.yml up
#
#   # Start REST API server instead:
#   ACESTEP_MODE=api docker compose -f docker-compose.jetson.yml up
#
#   # Build and start:
#   docker compose -f docker-compose.jetson.yml up --build
#
#   # Run in background:
#   docker compose -f docker-compose.jetson.yml up -d
#
#   # Stop:
#   docker compose -f docker-compose.jetson.yml down
#
#   # View logs:
#   docker compose -f docker-compose.jetson.yml logs -f
#
#   # Override JetPack version at build time:
#   L4T_VERSION=r36.3.0 docker compose -f docker-compose.jetson.yml up --build
#
# =============================================================================

services:
  acestep:
    build:
      context: .
      dockerfile: Dockerfile.jetson
      args:
        L4T_VERSION: ${L4T_VERSION:-r36.4.0}
    image: acestep-jetson:latest
    container_name: acestep-jetson

    # ---- GPU access ----
    runtime: nvidia

    # ---- Mode: "gradio" (web UI) or "api" (REST API) ----
    # Override from shell:  ACESTEP_MODE=api docker compose ... up
    env_file:
      - path: .env
        required: false
    environment:
      - NVIDIA_VISIBLE_DEVICES=all
      - ACESTEP_MODE=${ACESTEP_MODE:-gradio}
      # vllm with 4B LM = best quality (README recommendation for ≥24GB).
      # CUDA graph capture is auto-disabled on Jetson (enforce_eager).
      - ACESTEP_LLM_BACKEND=${ACESTEP_LLM_BACKEND:-vllm}
      - ACESTEP_INIT_SERVICE=${ACESTEP_INIT_SERVICE:-true}
      - ACESTEP_CONFIG_PATH=${ACESTEP_CONFIG_PATH:-acestep-v15-turbo}
      - ACESTEP_LM_MODEL_PATH=${ACESTEP_LM_MODEL_PATH:-acestep-5Hz-lm-4B}
      - ACESTEP_EXTRA_ARGS=${ACESTEP_EXTRA_ARGS:-}
      - TOKENIZERS_PARALLELISM=false

    # ---- Ports ----
    # Gradio UI on 7860, REST API on 8001
    ports:
      - "${GRADIO_PORT:-7860}:7860"
      - "${API_PORT:-8001}:8001"

    # ---- Persistent volumes ----
    volumes:
      # Model checkpoints — bind mount so models are visible on the host,
      # can be shared across containers, and survive image rebuilds.
      - ./checkpoints:/app/checkpoints
      # HuggingFace cache — avoids re-downloading models if the checkpoint
      # directory is empty and the downloader fetches from HF Hub.
      - hf_cache:/root/.cache/huggingface
      # Generated audio output
      - ./gradio_outputs:/app/gradio_outputs

    # ---- Resource management ----
    # Shared memory — needed for PyTorch DataLoader workers
    shm_size: "2gb"

    # ---- Restart policy ----
    restart: unless-stopped

    # ---- Health check ----
    # Inherited from Dockerfile; compose-level override for faster feedback
    healthcheck:
      test: >-
        curl -sf http://localhost:7860/ > /dev/null 2>&1 ||
        curl -sf http://localhost:8001/health > /dev/null 2>&1 ||
        exit 1
      interval: 30s
      timeout: 10s
      start_period: 300s
      retries: 3

volumes:
  # HuggingFace download cache — persists across container rebuilds so models
  # don't need to be re-downloaded from the Hub.
  hf_cache: