Cuda implementation broken

Notice: In order to resolve issues efficiently, please follow the template and include reproducible details.

🐛 Bug

No output of the transcription

To Reproduce

C:\Users\USER\Downloads\funasr-llamacpp-windows-x64-cuda>llama-funasr-sensevoice.exe -m sensevoice-small-q8.gguf -a sample.wav --backend cuda
ggml_cuda_init: found 1 CUDA devices (Total VRAM: 16310 MiB):
  Device 0: NVIDIA GeForce RTX 5060 Ti, compute capability 12.0, VMM: yes, VRAM: 16310 MiB

C:\Users\USER\Downloads\funasr-llamacpp-windows-x64-cuda>dir sample.wav
 Volume in drive C has no label.
 Volume Serial Number is ---

 Directory of C:\Users\USER\Downloads\funasr-llamacpp-windows-x64-cuda

07/19/2026  05:15 AM         1,077,326 sample.wav
               1 File(s)      1,077,326 bytes
               0 Dir(s)  89,731,526,656 bytes free

C:\Users\USER\Downloads\funasr-llamacpp-windows-x64-cuda>

C:\Users\USER\Downloads\funasr-llamacpp-windows-x64-cuda>llama-funasr-sensevoice.exe -m sensevoice-small-q8.gguf -a sample.wav
I'll record myself in English, no, you don't want.
[sensevoice] done 1.55s

Code sample

C:\Users\USER\Downloads\funasr-llamacpp-windows-x64-cuda>llama-funasr-sensevoice.exe -m sensevoice-small-q8.gguf -a sample.wav --backend cuda
ggml_cuda_init: found 1 CUDA devices (Total VRAM: 16310 MiB):
  Device 0: NVIDIA GeForce RTX 5060 Ti, compute capability 12.0, VMM: yes, VRAM: 16310 MiB

C:\Users\USER\Downloads\funasr-llamacpp-windows-x64-cuda>dir sample.wav
 Volume in drive C has no label.
 Volume Serial Number is FA17-FC47

 Directory of C:\Users\USER\Downloads\funasr-llamacpp-windows-x64-cuda

07/19/2026  05:15 AM         1,077,326 sample.wav
               1 File(s)      1,077,326 bytes
               0 Dir(s)  89,731,526,656 bytes free

C:\Users\USER\Downloads\funasr-llamacpp-windows-x64-cuda>

C:\Users\USER\Downloads\funasr-llamacpp-windows-x64-cuda>llama-funasr-sensevoice.exe -m sensevoice-small-q8.gguf -a sample.wav
I'll record myself in English, no, you don't want.
[sensevoice] done 1.55s

Expected behavior

Should return the audio like the CPU implementation

Error logs

No errors

Environment

• OS: W11

Audio details

It just works with CPU so it should work with GPU.

Start with the llama-funasr-sensevoice.exe entry point and reproduce the CPU and CUDA commands using sample.wav. Verify the CUDA command shown in the report and compare its output and logs with the working CPU run. Done means the CUDA backend returns the same transcription for the provided audio.