# Copy to .env and fill in. Loaded by docker-compose; for local runs, `export`
# these (or `set -a; source .env; set +a`) before `uv run python server.py`.

# --- Required ---

# Postgres voice store. Use a session-mode connection (Supabase "Session pooler"
# or a direct connection) and include ?sslmode=require.
DATABASE_URL=postgresql://USER:PASSWORD@HOST:5432/postgres?sslmode=require

# Hugging Face token with access to the gated model (needed on first run to
# download MODEL_PATH from config.py).
HF_TOKEN=hf_xxx

# --- Optional runtime tunables (defaults in config.py) ---
# Every value below is read at startup; change it here and restart the container
# (no image rebuild). Anything NOT listed here is a locked, checkpoint-intrinsic
# constant and must not be changed without retraining.

# Model to serve: any HF repo id or local path. Set this to serve a finetune
# (default is the base checkpoint).
# TTS_MODEL_PATH=nineninesix/kani3-dpo-cfg-round-2
# TTS_CODEC_MODEL=nvidia/nemo-nano-codec-22khz-1.89kbps-21.5fps   # must match training

# GPU / serving (tune per machine + load)
# TTS_GPU_MEMORY_UTILIZATION=0.82   # vLLM VRAM fraction; lower on smaller GPUs
# TTS_MAX_NUM_SEQS=256              # max concurrent streams
# TTS_CODEC_MAX_DECODE_BATCH=32     # codec decode batch cap (memory vs throughput)

# Streaming granularity / latency
# TTS_CHUNK_SIZE=86                 # frames per steady-state codec decode
# TTS_STREAM_FETCH_EVERY=86         # frames per worker poll (keep aligned with CHUNK_SIZE)
# TTS_FIRST_CHUNK_FRAMES=10         # smaller -> lower TTFB
# TTS_LOOKBACK_FRAMES=8             # frames of prior context for continuity

# Generation
# TTS_TEMPERATURE=0.3              # applied to all 32 heads uniformly
# TTS_TOP_P=0.95
# TTS_MAX_TOKENS=1000              # max frames (~ minutes of audio)
# TTS_STOP_THRESHOLD=0.5

# Output tail shaping (anti-click)
# TTS_END_OF_SPEECH_FADE_MS=0.0
# TTS_END_OF_SPEECH_SILENCE_MS=0.0