ModelLoader: read failed for tensor lm.blk.0.attn_norm.weight with ASR q4_k.gguf

[tobias-mueller-seven2one]

Running vibevoice-cli on Windows always fails with ModelLoader: read failed for tensor lm.blk.0.attn_norm.weight immediately after the audio is loaded. The build itself succeeds without errors, only inference fails.
The issue reproduces with both vibevoice-asr-q4_k.gguf and vibevoice-asr-q8_0.gguf from the official mudler/vibevoice.cpp-models HF repository.

Reproduce

1. Clone the repository and check out v0.1
2. Build on Windows with MSYS2 / MinGW-w64 UCRT64:

   cmake -B build -G "Ninja" -DBUILD_SHARED_LIBS=OFF -DCMAKE_BUILD_TYPE=Release -DGGML_USE_MMAP=OFF
   cmake --build build -j
   

3. Copy required DLLs next to the executable:

   copy C:\msys64\ucrt64\bin\libwinpthread-1.dll build\bin\
   copy C:\msys64\ucrt64\bin\libgcc_s_seh-1.dll build\bin\
   copy C:\msys64\ucrt64\bin\libstdc++-6.dll build\bin\
   copy C:\msys64\ucrt64\bin\libgomp-1.dll build\bin\
   

4. Download models:

   hf download mudler/vibevoice.cpp-models --local-dir models
   

5. Run inference:

   .\build\bin\vibevoice-cli.exe asr --model ..\models\vibevoice-asr-q4_k.gguf --tokenizer ..\models\tokenizer.gguf --audio "audio.wav" --max-new-tokens 8192
   

Expected

The model loads successfully and transcribes the audio.

Actual

asr: loaded 142636 samples (5.94s)
[vv I] backend: CPU
[vv E] ModelLoader: read failed for tensor lm.blk.0.attn_norm.weight
asr: failed to load model

Environment

OS	Windows 11
Compiler	GCC 16.1.0 (MSYS2 UCRT64, MinGW-w64)
Branches tested	master (commit ad856bd) and v0.1 tag (commit 8ffffa1)
ggml version	0.10.0 (commit 8be60f83)
Model	vibevoice-asr-q4_k.gguf (10,392,063,296 bytes)
Model source	hf download mudler/vibevoice.cpp-models

Additional context

• The error occurs at the very first tensor (lm.blk.0.attn_norm.weight), which suggests a GGUF format mismatch between the uploaded HF models and the current model_loader.cpp parsing logic, or a Windows-specific file reading issue.
• Disabling mmap (-DGGML_USE_MMAP=OFF) does not resolve the issue.
• Building from both the latest master and the v0.1 tag (which was tagged at the same time the models were uploaded to HF) produces the same error.
• The build completes without errors. The CLI starts correctly and loads the audio file successfully, only model loading fails.

Has this been tested on Windows by the maintainers? Is there a known workaround or a specific build configuration required for Windows?