# Core — required for every inference path
huggingface_hub>=0.24
safetensors>=0.4
# Audio file I/O (saving, voice-reference loading) — soundfile only, so the
# pack never touches torchaudio's file-backend machinery (v1's torchcodec
# failure mode). torchaudio is still needed for tensor ops in the vendored
# audio VAE (mel-spectrogram + resample), not for file I/O.
soundfile>=0.12
torchaudio
numpy
einops>=0.7
# Gemma-3-12B text encoder (vendored LTX-2 code loads it via transformers).
# accelerate is required by transformers to place the model with device_map
# (the only way to load the 4-bit/bnb text encoder, which can't be .to()'d).
transformers>=4.53
accelerate>=0.25.0
sentencepiece>=0.1.99
# Perth watermark apply/detect nodes.
resemble-perth

# GGUF support (calcuis/gguf) — quantized audio-DiT convert + load. Optional:
# only needed if you use .gguf DiT files or the convert tool.
gguf-connector>=3.6

# Optional — 4-bit Gemma text encoder (the recommended ~8 GB path).
# Without it, choose the bf16 text encoder (~24 GB, HF-gated) instead.
bitsandbytes>=0.45.0; platform_system != "Darwin"

# Optional — RE-USE voice-reference denoising (NSCLv1, non-commercial):
# needs mamba-ssm + causal-conv1d (no prebuilt wheels on Windows/macOS);
# unavailable installs degrade gracefully with denoise skipped.
