feat: support independent sampler rng

[leejet]

In ComfyUI, when using samplers that include ancestral sampling steps (such as euler_a), an independent RNG is used to generate noise. Moreover, ComfyUI uses the RNG of the device where the current tensor resides. This means that if ComfyUI is running on CUDA, the initial noise is generated on the CPU, but the noise added during sampling is generated on CUDA. If ComfyUI is running on another device, such as the CPU, the results will be different.

The ComfyUI commit that introduced this issue: comfyanonymous/ComfyUI@916d1e1.

I don’t really understand why ComfyUI is designed this way, as it breaks cross-device consistency and causes different versions of ComfyUI to produce different outputs. However, for users who want to reproduce ComfyUI’s results in sd.cpp, this PR is necessary.

.\bin\Release\sd.exe -m ..\..\stable-diffusion-webui\models\Stable-diffusion\v1-5-pruned-emaonly.safetensors -p "a lovely cat BREAK" -v --diffusion-fa --rng cpu --sampler-rng cuda

sd.cpp result

comfyui result