Linux implementation bottleneck for Flux

I did the same installation steps for Linux and Windows and used the same generation parameters

OS: Linux

**Model: Flux dev nf4

UI: latest
Cuda: 12.1
Python: 3.10**
console output during generation:

Moving model(s) has taken 8.43 seconds
7%|██▉ | 2/30 [00:22<05:26, 11.67s/it

time: 11 seconds per iteration (varying too much between 8s and 15s)

OS: Windows

**Model: Flux dev nf4 (same)

UI: latest (same)
Cuda: 12.1 (same)
Python: 3.10 (same)**
console output during generation:

Moving model(s) has taken 13.71 seconds
60%|█████████████████████████████████████████████████▏ | 12/30 [00:42<00:28, 3.59s/it]
Total progress: 65%|██████████████████████████████████████████▉ | 13/30 [00:42<00:24, 3.56s/it]

time : 3.5 seconds per iteration (stable)

No source file, test, or entry point is named. Reproduce Flux dev nf4 generation on Linux and Windows with CUDA 12.1, Python 3.10, and the same parameters, then inspect the generation logs and iteration timings to identify where Linux performance diverges. Done means the bottleneck is isolated and covered by an appropriate fix or documented diagnosis.