0.1.3

Dots TTS ComfyUI v0.1.3

Added

• Added an opt-in compile toggle to Dots TTS Load Model.
• Uses native PyTorch Inductor and Triton compilation.
• Compatible with SDPA and Flash Attention.
• Added CUDA, Triton, Inductor, and compile-length validation.
• Added automatic compatibility handling for ComfyUI's cudaMallocAsync allocator.

Fixed

• Fixed the cudaMallocAsync does not yet support checkPoolLiveAllocations crash.
• Fixed streaming vocoder LSTM compilation on PyTorch 2.10.
• Fixed compiled graphs and static generation workspaces remaining after unload.
• Changing model, device, dtype, attention, or compile mode now fully unloads the previous bundle.

Changed

• CUDA Graph Trees are automatically disabled with cudaMallocAsync; Inductor and Triton compilation remain active.
• The terminal displays Preparing/compiling until the first audio patch is ready.
• Compilation is lazy. The first generation for each length bucket is slower, while later generations reuse the compiled graph.
• Compiled generation supports up to 1024 audio patches.

Notes

• Compilation is CUDA-only and requires a working Triton installation.
• For the fastest generation, use Dots TTS MF BF16 with steps=4.