env.example

# VibeVoice API Configuration
# Copy this file to .env and adjust values as needed

# ============================================================
# Model Configuration
# ============================================================

# Path to VibeVoice model (HuggingFace model ID or local path).
# Recommended choices based on hardware:
#   - ncoder-ai/VibeVoice-Large-AWQ     : RECOMMENDED. Drop-in AWQ-INT4 quantized large model.
#                                         ~8.4 GB VRAM, RTF ~0.70 on RTX 3090 + i5-12600K.
#                                         Audio quality indistinguishable from FP16 at
#                                         INFERENCE_STEPS=7. Just point at this — no extra env vars.
#   - rsxdalv/VibeVoice-Large           : FP16, ~18 GB VRAM. Fastest single-model RTF (~0.54)
#                                         on a 24 GB Ampere card. Use when VRAM is plentiful.
#   - FabioSarracino/VibeVoice-Large-Q8 : INT8 (bitsandbytes). ~11 GB VRAM. Quality matches FP16
#                                         at INFERENCE_STEPS=7+ (clean) — at lower step counts
#                                         the quantization noise becomes audible. Slower per-op
#                                         dispatch on Ampere — AWQ-INT4 is faster + smaller.
#   - microsoft/VibeVoice-1.5B          : Smaller model, faster, single-speaker only (no multi-
#                                         speaker dialogue, no voice cloning).
#   - microsoft/VibeVoice-Realtime-0.5B : Real-time variant. Single-speaker only.
VIBEVOICE_MODEL_PATH=ncoder-ai/VibeVoice-Large-AWQ
#VIBEVOICE_MODEL_PATH=rsxdalv/VibeVoice-Large
#VIBEVOICE_MODEL_PATH=microsoft/VibeVoice-1.5B
#VIBEVOICE_MODEL_PATH=FabioSarracino/VibeVoice-Large-Q8
# Local path example:
# VIBEVOICE_MODEL_PATH=/path/to/your/model

# Device for inference: cuda, cpu, or mps
VIBEVOICE_DEVICE=cuda

# Optional: Specify which GPU to use
# CUDA_VISIBLE_DEVICES=0

# Number of diffusion inference steps (5-50, higher = better quality but slower).
# 7 is the recommended sweet spot for both FP16 and the Q8 model — the Q8 model
# produces audible quantization noise below 7 steps but is clean at 7+.
# 5 works fine for FP16; for Q8 use 7+.
VIBEVOICE_INFERENCE_STEPS=7

# Model dtype: bfloat16, float16, or float32 (auto-detected if not set)
# VIBEVOICE_DTYPE=bfloat16

# Attention implementation: flash_attention_2, sdpa, or eager (auto-detected if not set)
# VIBEVOICE_ATTN_IMPLEMENTATION=flash_attention_2

# Enable torch.compile for optimized inference (~40% speedup on Ampere).
# Note: First request triggers compilation (~10-30 s); subsequent requests are fast.
# Persists across container restarts via /app/torch_cache mount (see docker-compose.yml).
TORCH_COMPILE=true

# torch.compile mode: default, reduce-overhead, or max-autotune (slower compile, faster inference)
# TORCH_COMPILE_MODE=max-autotune

# Runtime torchao quantization (used only with FP16 base models — pre-quantized
# models like ncoder-ai/VibeVoice-Large-AWQ or FabioSarracino/VibeVoice-Large-Q8
# do NOT need this; transformers auto-detects their embedded quantization_config
# and wires the kernels automatically).
# Options:
#   - int8_torchao         : Weight-only INT8. ~40% VRAM cut. SLOW on Ampere (RTF ~1.1)
#                            because dequantize-to-FP16-before-matmul adds overhead. Useful
#                            on Hopper/Ada.
#   - int8_dynamic_torchao : W8A8 dynamic. Very slow on Ampere (RTF ~4.5+) — torch.compile
#                            can't fuse the dynamic quant ops. Don't use on RTX 30/40 series.
#   - int4_torchao         : INT4 weights. Smallest VRAM, even slower than INT8 on Ampere.
#   - (unset)              : No runtime quantization. Recommended — use a pre-quantized
#                            model path like ncoder-ai/VibeVoice-Large-AWQ instead.
# VIBEVOICE_QUANTIZATION=int8_torchao

# ============================================================
# Voice Configuration
# ============================================================

# Directory containing voice preset audio files
# The API automatically loads ALL audio files from this directory as voice presets
# Supported formats: .wav, .mp3, .flac, .ogg, .m4a, .aac
VOICES_DIR=./demo/voices

# OpenAI voice name to VibeVoice preset mapping (JSON format)
# Maps OpenAI-compatible voice names (alloy, echo, fable, onyx, nova, shimmer) to your voice preset files
# Format: {"openai_voice_name": "vibevoice_preset_name", ...}
OPENAI_VOICE_MAPPING={"alloy": "en-Alice_woman", "echo": "en-Carter_man", "fable": "en-Maya_woman", "onyx": "en-Frank_man", "nova": "en-Mary_woman_bgm", "shimmer": "en-Alice_woman"}

# ============================================================
# API Server Configuration
# ============================================================

API_HOST=0.0.0.0
API_PORT=8001
API_WORKERS=1
API_CORS_ORIGINS=*

# ============================================================
# Generation Defaults
# ============================================================

# Default CFG scale (1.0-3.0, higher = more faithful to prompt)
DEFAULT_CFG_SCALE=1.8

# Default audio response format
DEFAULT_RESPONSE_FORMAT=mp3

# Maximum generation length in seconds
MAX_GENERATION_LENGTH=5400

# Text generation sampling (False = greedy decoding, True = sampling)
DEFAULT_DO_SAMPLE=False

# Sampling parameters (only used if DO_SAMPLE=True)
DEFAULT_TEMPERATURE=1.0
DEFAULT_TOP_P=1.0
DEFAULT_TOP_K=50
DEFAULT_REPETITION_PENALTY=1.0

# ============================================================
# Logging
# ============================================================

LOG_LEVEL=INFO