fix: Add NVIDIA Blackwell (RTX 50xx, sm_120) GPU support

[Hasham-dev]

Summary

Minimal fixes to support NVIDIA Blackwell architecture GPUs (RTX 5050/5060/5070/5080/5090, compute capability sm_120) on Fooocus.

• bfloat16 UNet dtype for Blackwell GPUs (compute major >= 12) which have native bf16 tensor core support
• Skip manual_cast for bf16 weights to avoid unnecessary dtype casting overhead
• Fix numpy TypeError with bfloat16 tensors in modules/patch.py and extras/ip_adapter.py — numpy doesn't support bf16, so we convert to float32 before .numpy() calls

Changes (3 files, +13 -2 lines)

File	Change
ldm_patched/modules/model_management.py	Auto-detect Blackwell GPUs and use bf16 dtype; skip manual_cast for bf16
modules/patch.py	Fix bf16→numpy crash in patched_unet_forward
extras/ip_adapter.py	Fix bf16→numpy crash in IP-Adapter attention patcher

Testing

• GPU: NVIDIA GeForce RTX 5070 (sm_120, 11.5GB VRAM)
• CUDA: 12.8
• PyTorch: 2.12.0.dev (nightly, cu128)
• Result: Image generation works at ~3.2 it/s at 896x1152, including Image Prompt (IP-Adapter) mode
• VAE note: Users will also need madebyollin/sdxl-vae-fp16-fix VAE for stable bf16 decoding (the default SDXL VAE overflows in bf16)

Fixes

• Fixes NVIDIA GeForce RTX 5090 with CUDA capability sm_120 is not compatible with the current PyTorch installation. #3862 (RTX 5090 sm_120 not compatible)
• Fixes [Feature Request]: # RTX 5060 (sm_120) Not Supported - CUDA Compatibility Issue #4123 (RTX 5060 sm_120 not supported)
• Fixes [Bug]: ### Issue: CUDA error on RTX 5050 Laptop GPU (sm_120 not supported) **Description:** When running Fooocus on a system with an NVIDIA GeForce RTX 5050 Laptop GPU, I encounter the following error: #4141 (CUDA error on RTX 5050 sm_120)

Prerequisites

Users with Blackwell GPUs need PyTorch nightly with CUDA 12.8 support:

pip install --pre torch torchvision torchaudio --index-url https://download.pytorch.org/whl/nightly/cu128