HF_TOKEN=hf_your_huggingface_token_here
# Models will be auto-downloaded on first run.
# For TranslateGemma GGUF, place the file at:
#   ./models/translategemma-12b-q8_0.gguf
# Or set TRANSLATE_MODEL_PATH to the actual path.
#
# GPU selection:
# CUDA_VISIBLE_DEVICES=0  -> worker uses GPU 0
# ROCR_VISIBLE_DEVICES=1  -> SGLang TTS uses GPU 1