.env.example

# ACE-Step Environment Configuration
# Copy this file to .env and modify as needed
#
# This file is used by:
# - Python scripts (acestep_v15_pipeline.py, api_server.py, etc.)
# - Windows launcher (start_gradio_ui.bat)
# - Linux/macOS launchers (start_gradio_ui.sh, start_gradio_ui_macos.sh)
#
# Settings in .env will survive repository updates, unlike hardcoded values
# in launcher scripts which get overwritten on each update.

# ==================== Model Settings ====================
# DiT model path
ACESTEP_CONFIG_PATH=acestep-v15-turbo

# LM model path (used when LLM is enabled)
# Available: acestep-5Hz-lm-0.6B, acestep-5Hz-lm-1.7B, acestep-5Hz-lm-4B
ACESTEP_LM_MODEL_PATH=acestep-5Hz-lm-1.7B

# Device selection: auto, cuda, cpu, xpu
ACESTEP_DEVICE=auto

# LM backend: vllm (faster) or pt (PyTorch native)
ACESTEP_LM_BACKEND=vllm

# ==================== LLM Initialization ====================
# Controls whether to initialize the Language Model (LLM/5Hz LM)
#
# Flow: GPU detection (full) → ACESTEP_INIT_LLM override → Model loading
# GPU optimizations (offload, quantization, batch limits) are ALWAYS applied.
# ACESTEP_INIT_LLM only overrides the "should we try to load LLM" decision.
#
# Values:
#   auto (or empty)    = Use GPU auto-detection result (recommended)
#   true/1/yes         = Force enable LLM after GPU detection (may cause OOM)
#   false/0/no         = Force disable LLM (pure DiT mode, faster)
#
# Examples:
#   ACESTEP_INIT_LLM=auto       # Let GPU detection decide (recommended)
#   ACESTEP_INIT_LLM=           # Same as auto
#   ACESTEP_INIT_LLM=true       # Force enable even on low VRAM GPU
#   ACESTEP_INIT_LLM=false      # Force disable for pure DiT mode
#
# When LLM is disabled, these features are unavailable:
#   - Thinking mode (thinking=true)
#   - Chain-of-Thought caption/language detection
#   - Sample mode (generate from description)
#   - Format mode (LLM-enhanced input)
#
# Default: auto (based on GPU VRAM detection)
ACESTEP_INIT_LLM=auto

# ==================== Download Settings ====================
# Preferred download source: auto, huggingface, modelscope
# ACESTEP_DOWNLOAD_SOURCE=auto

# ==================== API Server Settings ====================
# API key for authentication (optional)
# ACESTEP_API_KEY=sk-your-secret-key

# ==================== Gradio UI Settings ====================
# Server port (default: 7860)
# PORT=7860

# Server name/host (default: 127.0.0.1 for local only, 0.0.0.0 for network access)
# SERVER_NAME=127.0.0.1

# UI language: en, zh, he, ja (default: en)
# LANGUAGE=en

# Default batch size for generation (1 to GPU-dependent max)
# When not specified, defaults to min(2, GPU_max)
# ACESTEP_BATCH_SIZE=2

# ==================== Startup Settings ====================
# Skip model loading at startup (models will be lazy-loaded on first request)
# Set to true to start server quickly without loading models
# ACESTEP_NO_INIT=false