An fp8 or fp8_scaled ComfyUI Krea 2 file (13 GB) on a GPU with no fp8 runtime
is widened to bf16 in host RAM (about 26 GB) before the denoiser is placed.
With too little free memory the process was killed (Linux OOM killer, Windows
commit limit), taking the whole app down instead of failing the load.

Price the new host memory from the safetensors header (tensors already at the
compute dtype stay memory-mapped and cost nothing) and refuse with a readable
message before any widening, in both the original-layout loader and the
ComfyUI-quantized loader. UNSLOTH_DIFFUSION_HOST_RAM_CHECK=0 skips the check.