Alright, so I've got my local SD instance (Automatic1111) humming along. It's generating images that are, objectively, quite nice. The problem? My GPU is practically idling while my CPU is pegged at 100%. The fans sound like a small aircraft preparing for takeoff, but for all the wrong reasons.
I've checked the obvious culprits. I'm not accidentally running on CPU-only mode in the settings. The NVIDIA GPU is selected. Yet, Task Manager tells a clear and frustrating story: the GPU usage spikes briefly during the actual generation, then flatlines, while the CPU carries the entire burden of... something. Queuing? Image decoding? The UI itself?
This feels like a classic case of a hidden configuration or a dependency bottleneck. Before I dive into the usual "reinstall PyTorch with CUDA" ritual, has anyone done a proper, systematic trace on where the CPU cycles are actually going during a typical txt2img workflow? I'm skeptical of the usual forum replies that just parrot the same three settings to check.
What I'm looking for is the *reproducible methodology*. What specific logs, performance counters, or stack traces actually pinpoint the stage where the CPU becomes the choke point? Is it in the VAE decode? The web UI's overhead? Some inefficient data movement between system RAM and VRAM?
Data skeptic, not a data cynic.