Blackwell Attention Fix (RTX 50xx)

Overview

Runtime ComfyUI node that prevents the common xformers crash on NVIDIA Blackwell GPUs (RTX 5090 / 5080 / 5070):

the=is prevents you from having to disable xfromers, which may slow your machine down. If you prefer, add his to your .bat file:
--disable-xformers --use-pytorch-cross-attention

GitHub Repo: https://github.com/lonecatone23/comfyUI-BlackwellAttentionFix

NotImplementedError: No operator found for memory_efficient_attention_forward

… requires device with capability ≤ (9, 0) but your GPU has capability (12, 0) (too new)

It works by redirecting ComfyUI’s attention entry points to the built-in attention_pytorch (SDPA) implementation, which is fully compatible with compute capability 12.0.

Modes

bypass (default) : No-op. Model passes through unchanged. Safe for any GPU and shared workflows.

pytorch :Forces native PyTorch attention. Only applies the change when a Blackwell GPU is detected.

Usage

- Insert the node immediately after any model / UNET / Krea2 loader and before samplers or attention-heavy custom nodes.

- On RTX 50-series + Krea2 (or any model that triggers the xformers error) → set mode = pytorch.

- On all other systems or shared workflows → leave mode = bypass.

The node can stay permanently in a workflow; with bypass it never interferes with xformers, SageAttention, or other people’s setups.

What it patches

- comfy.ldm.modules.attention.optimized_attention

- optimized_attention_masked

- attention_xformers

- Direct imports captured at load time (notably comfy.ldm.krea2.model)

Compatibility

- Works with Krea2 / krea2edit, Flux, SDXL, SD1.5, Wan, LTX and most other models that go through ComfyUI’s standard attention path.

- Does not require launch flags --disable-xformers, etc.).

- Does not uninstall or break xformers; it only redirects the active functions for the current process when enabled.

Limitations

- VAE may still log “Using xformers attention” — this is harmless and separate from the diffusion attention path that was crashing.

- Performance is native PyTorch SDPA (usually close to xformers on Blackwell). SageAttention is not used by this node.

- Extremely rare custom nodes that hard-code their own xformers calls are not covered.

License / Status

Community utility node. No external dependencies beyond ComfyUI + PyTorch.

Instagram: https://www.instagram.com/synth.studio.models/

Buy me a☕ https://ko-fi.com/lonecatone

This represents several hours of work that I give away for free. If you enjoy it, Please make sure to 👍like and post to my workflows. That is how you support me!!! Also, feel free to ⚡tip 😉