使用gguf版本的indextts2模型时，所有的模型都加载在内存里，占用也很大。When using the GGUF version of IndexTTS2, all model components are loaded into memory, resulting in high memory usage.

[asdong602]

使用index-tts2-f16.gguf 模型，文件大小4.5个G。现在我发现两个问题
第一个问题：模型都加载到内存里，而不是优先加载到显存里，我本地16G的显存一点没用。
第二个问题：模型加载到内存里，感觉占用了20到30G的空间

[asdong602]

补充：audiocpp-windows-cuda-balance-d3d7481 版本+ CUDA 参数 。 实际测试使用CUDA参数比CPU参数速度快很多，显卡应该是参与计算了，我的疑问就是模型加载没有优先加载到显存里，加载到显存里会不会更快？

[0xShug0]

@asdong602 There are two separate stages here: reading the model package, and uploading the model weights. audio.cpp does not load GGUF directly from disk into VRAM in one step. It first opens/maps the GGUF on the host side as a tensor source, then uploads the required tensors to the selected backend. If you run with --backend cuda, the loaded weights should be uploaded to CUDA/VRAM. If --backend cuda is not specified, the CLI defaults to CPU. It's noraml to see high transient host memory during GGUF load. The recent commit #97 acutally makes the host memory usage more efficient. GGUF cannot be “opened directly as VRAM.” It is a file format on disk. The loader must parse metadata and read tensor byte ranges on the host first.

[asdong602]

@0xShug0 Thanks so much for your detailed explanation!

[0xShug0]

@asdong602 Just a quick update. The high RAM usage for index-tts2 on Windows is acutally a bug, which was solved in #151