#3526 [Engine] Performance degradation among other issues on V0.2.1

Descrição

### System Specifications

- **Graphics Card & VRAM:** NVIDIA RTX 5080 16GB + RTX 2080 + AMD RADEON (iGPU) (used 5080 + 2080 for inference, Radeon for display)
- **System RAM:** 64GB
- **Operating System:** [Windows 11]

### Issue

I tried to upgrade to V0.2.1 but found many issues:

- When I was able to run the engine, the performance was very slow compared to V 0.1 (engine 0.1.44.1). With the old version, the system achieves ~1200 t/s for prefill, while with the new 0.1.46.3 only gets to ~500 t/s. Also, generation fropped from ~65 t/s to 53 t/s.
- Tuning breaks: with the computer doing nothing I got a result that the PC "was busy" and refused the tuning, tried again, said that it was already optimized and then the engine crashed when I queried the model with the error "the fused layout's MoE buffers are too small for layer 45".
- Disabling vision ignores the setting
- Desktop app ignores most of the changes to the settings (multi-gpu select, optimize for speed, etc).

Rolling back to V0.1.44 fixed all the issues.