plotarmordev/drama-tensorfold is a repository publishing a recipe and engine patches for running Fish Audio Drama 3 speech synthesis on DGX Spark (GB10) with the TensorFold native Zig/CUDA engine. There is no public release version. It builds a local speech server that provides an OpenAI-style `/v1/audio/speech` endpoint, supporting both full and streaming responses. It is written in Python and licensed under Apache-2.0.
The repository includes benchmark and accuracy check materials, and also publishes the results of a comparison with the sglang-omni reference implementation. According to the repository description, 14 speed-improvement patches kept 46 responses byte-for-byte identical. Time to first audio is 0.141 seconds, lower than the reference implementation's 0.521 seconds, and the real-time factor is given as 0.634 versus 0.939. This material is worth consulting when looking at a native CUDA-based TTS path on GB10.
However, the current state is 'testing, not qualified' and it does not meet the reference accuracy criteria. The presented figures are results from after patch 0040, which trades roughly 5% of speed for accuracy. The comparison is also limited to a single host and a single model, so it does not back performance or output quality in other environments. It is therefore more reasonable to view this repository as an optimization recipe under evaluation under limited conditions, rather than a validated deployment path.