npu-cosyvoice prebuilt files

Precomputed artifacts for npu-cosyvoice, which runs CosyVoice3's flow-matching DiT on the AMD Ryzen AI (XDNA1) NPU. The voice bundle's setup step downloads these instead of exporting them locally, which saves about 15 minutes of CPU time.

path	what	made by
weights/	int8 DiT weights (W8A8, per-output-channel scales, SmoothQuant on q/k/v) and AdaLN weights	python -m npu_cosyvoice.export
cv_llm_q4_0.gguf	CosyVoice3's Qwen2 LM backbone (no output head), Q4_0, for llama.cpp	python -m npu_cosyvoice.llm (llama.cpp 9adc7f4)
aot/npu1/	the 512x1024x1024 int8 GEMM overlay for XDNA1	scripts/compile_xclbin.py --device npu1
SHA256SUMS	checksums; setup verifies them

All three are deterministic: rebuilding them with the repository's container reproduces these files bit for bit (see the repository README). You still need the original CosyVoice3 checkpoint for the parts that stay in PyTorch (vocoder, embeddings, voice encoder); setup downloads it from ModelScope.

License and attribution

Derived from Fun-CosyVoice3-0.5B-2512 by FunAudioLLM, released under Apache-2.0; these files are distributed under the same license (see LICENSE). Modifications: the DiT projection weights were quantized to int8 with SmoothQuant scaling on the q/k/v input, calibrated on four sentences; the LM backbone was converted to GGUF and quantized to Q4_0 with its unused output head removed. The NPU overlay is generated with MLIR-AIE (Apache-2.0 WITH LLVM-exception).