--fast issues with Flux

### Expected Behavior

Performance boost without errors

### Actual Behavior

1. In multi-gpu configuration ```supports_fp8_compute``` returns ```False```, optimizations not applied, however main GPU actually support it (I have 4090 and 3090)
2. After force ```True``` it works fine, but with loras terminal displays errors in single_blocks from 30 to 37 at each sampling step. Loras themselves work correctly.

### Steps to Reproduce

-

### Debug Logs

```powershell
ERROR lora diffusion_model.single_blocks.30.linear1.weight "addmm_cuda" not implemented for 'Float8_e4m3fn'
```

### Other

_No response_