The SDXL global stack (fp16 UNet, Canny ControlNet, both text encoders, VAE)
is ~8.8 GiB and was always moved to CUDA with .to(device), so it could not
fit an 8 GiB card and --cpu-offload never reached it. Under the Windows
display driver, WSL2 included, the overflow spilled into shared system
memory and the four-step pass ran for over ten minutes instead of failing.
Residency now follows the same rule as the Qwen stack: an explicit request
wins, otherwise the card's total VRAM decides against
RESIDENT_SDXL_GLOBAL_MIN_VRAM_GIB. Below it the Lightning LoRA is fused on
the CPU and the stack is streamed with enable_sequential_cpu_offload, the
fastest placement that completed on every run of an 8 GiB card with 16 GiB
of system RAM. sdxl-zimage joins GLOBAL_OFFLOAD_PROFILES, so the
--cpu-offload warning now fires for chroma-zimage only.
Giuseppe G. Di Chiara
c45e6db3e4