mirror of
https://github.com/ace-step/ACE-Step-1.5.git
synced 2026-07-02 16:37:04 +00:00
8abe5a172d
- Build FFmpeg 7.1 from source for torchcodec compatibility (Ubuntu 22.04 ships FFmpeg 4.4, torchcodec 0.10.0 needs FFmpeg 7) - Build torchcodec v0.10.0 from source with ENABLE_CUDA=0 (Jetson AI Lab wheel has ABI mismatch with desktop NVDEC) - Add bitsandbytes>=0.49.0 for INT8 quantization support - Make nano-vllm and torchcodec hard build requirements - Update torchao comment to reflect current status - Add lora_output and acestep_output volume mounts to compose
106 lines
3.5 KiB
YAML
106 lines
3.5 KiB
YAML
# =============================================================================
|
|
# 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
|
|
# LoRA training output — persists trained adapters across rebuilds
|
|
- ./lora_output:/app/lora_output
|
|
# ACE-Step generation output
|
|
- ./acestep_output:/app/acestep_output
|
|
|
|
# ---- 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:
|