zvoice_temp

Inference artifacts for an official first-stage distilled ZipVoice checkpoint (step1250), eight Euler steps, t_shift0.5, guidance1, feat_scale0.1. Code and setup instructions: https://github.com/xiro2416/zvoice_temp .

Contents: original FP8 and INT8 SmoothQuant Q/DQ ONNX graphs with all external tensor files; packed16 FP8 rebuild source; joint text and real Vocos ONNX; the final FP8 B1024 optimized TensorRT plans and compatible historical timing cache; token/model/Vocos configuration; SHA256 artifact inventory.

The final optimized route is FP8 only. INT8 is an existing baseline and its optimization was suspended. There is no PyTorch fake-quant runtime or separate TensorRT-direct release. Q/DQ graphs describe quantization but do not guarantee that an arbitrary ONNX provider executes native low-precision kernels.

Validated hardware: NVIDIA RTX6000D, SM120/compute capability12.0,156SMs, 85,651MiB reported memory,600W power limit, driver595.71.05. TensorRT11.3.0.99; Python3.12; runtimeTorch2.11cu130/Triton3.6; builderTorch2.13cu132/Triton3.7/CUDA toolkit13.2.78. Plans are not universally portable; rebuilding requires the GitHub custom plugins.

Fixed workload: full B1024, T681=prompt375+target306, joint69+padL70, concurrency1. Preserve prompt conditioning and full evolving prompt+target state. The fixed engines cannot run arbitrary prompt/text lengths without rebuilding. Historical first/all PCM8.870/9.012s, board power mean/P95/max449/475/484W; not a remote-server guarantee. Historical70-case Chinese tone-pinyinCER15/1050; no claim of universal/English quality acceptance.

No training data, reference voices, private evaluation cases or credentials included. Raw source checkpoint SHA256: ac4dd40e24f44d8b87779aa181ecfde2048f66660df1c65efa006b7ad1ff6200. The raw PyTorch checkpoint is not distributed; ONNX files contain inference weights. ZipVoice source attribution and Apache2.0 notices are in the GitHub repository; no additional rights to third-party training data or voices are asserted.