dots.tts Gradio Space for Hugging Face ZeroGPU with PyTorch AOTInductor startup compilation.

How it works

This Space runs an unmodified, pip-installed dots.tts package (pinned in requirements.txt) rather than a fork of the model source. All ZeroGPU-specific behavior lives in the local dots_tts_zerogpu/ patch layer, which:

• intercepts upstream's single compile seam (dots_tts.modules.backbone.inference_utils) and routes it to spaces.aoti_compile instead of torch.compile (ZeroGPU cannot run torch.compile);
• warms every generation bucket through upstream's own run_warmup inside a @spaces.GPU startup call, then installs the compiled artifacts into the serving process;
• forces the sdpa attention backend (the default flex-attention path relies on module-level torch.compile).

To upgrade to a newer upstream release: bump the dots.tts== pin in requirements.txt and re-verify that the compile seam symbols still exist (apply_zero_gpu_patches asserts this at startup).

Configuration

Set DOTS_TTS_MODEL_NAME_OR_PATH to a local model directory or Hugging Face model repo id. The app defaults to dots-studio/dots.tts-soar.

Other environment variables: DOTS_TTS_MODEL_REVISION, DOTS_TTS_PRECISION, DOTS_TTS_EXECUTION_MODE, DOTS_TTS_MAX_GENERATE_LENGTH, DOTS_TTS_DEFAULT_NUM_STEPS, DOTS_TTS_OPTIMIZE, DOTS_TTS_ENABLE_AOTI, DOTS_TTS_AOTI_COMPILE_ON_STARTUP, DOTS_TTS_ZERO_GPU_DURATION, DOTS_TTS_ZERO_GPU_COMPILE_DURATION.