ComfyUI cuDNN Frontend Error: No Valid Execution Plans Built (RTX 50-Series / Blackwell Fix)

One of the more confusing ComfyUI failures to hit a brand-new GPU is a Python traceback that ends in cuDNN Frontend error: [cudnn_frontend] Error: No valid execution plans built. It tends to appear the moment a workflow reaches the sampler, produces no image, and sends people down a rabbit hole of reinstalling PyTorch, cuDNN, and drivers — none of which actually fixes it. This guide explains what the error actually means, why it specifically targets RTX 50-series (Blackwell) cards, and the one launch flag that restores generation.

What the Error Actually Means

The error is not about a missing or broken cuDNN installation. It is about a mismatch between PyTorch’s default attention path and a GPU architecture that is newer than what that path knows how to handle. By default, PyTorch uses its scaled dot-product attention (SDPA) implementation, which can dispatch the computation to the cuDNN backend under the hood. On Blackwell-generation GPUs, that cuDNN backend has no registered execution plan for the new compute capability, so it returns the message you see in the traceback.

The binding nuance is worth spelling out because it explains a lot of failed “fixes.” Calling torch.backends.cuda.enable_cudnn_sdp(False) from a sitecustomize.py file does not work in this situation. The flag is not retained through CUDA initialization, so by the time PyTorch would honor it, the decision point has already passed. As the ComfyUI troubleshooting documentation states plainly, the only reliable fix is the launch flag.

Why It Hits Flux Hardest

With FLUX models specifically, the failure lands on the very first attention block of the UNet. That means the error is not a slow, intermittent degradation — it is an immediate hard stop. No image is ever produced, which makes the failure especially visible and often makes people think their model is corrupted, when the model is completely fine. The same underlying mechanism can affect any model that routes attention through the cuDNN SDPA backend on a Blackwell card, but FLUX makes it impossible to miss.

If your workflow produces nothing and the console shows only the No valid execution plans built line followed by a traceback that mentions execution.py around the sampler, this is very likely the cause rather than a broken checkpoint or invalid prompt.

!!! Exception during processing !!! cuDNN Frontend error: [cudnn_frontend] Error: No valid execution plans built.
Traceback (most recent call last):
  File "/home/user/ComfyUI/execution.py", line 496, in execute
  ...
RuntimeError: cuDNN Frontend error: [cudnn_frontend] Error: No valid execution plans built.

The Fix: –use-split-cross-attention

The recommended solution is to launch ComfyUI with a hand-written attention implementation that bypasses SDPA and its cuDNN backend entirely:

python main.py --use-split-cross-attention

This flag forces ComfyUI onto the split-cross-attention path, which uses an explicit attention implementation instead of letting PyTorch dispatch to the broken cuDNN backend. For desktop and portable installs, add the flag to whatever launcher you normally use:

• Manual install: run python main.py --use-split-cross-attention from the ComfyUI directory.
• Portable install: edit run_nvidia_gpu.bat (or create a copy) and add --use-split-cross-attention to the python invocation line.
• Desktop app: add the flag to the custom launch arguments in settings, or start the server manually with the flag.

Because the flag simply sidesteps the problematic backend rather than changing model or sampler settings, your FLUX (or any other) workflow should start producing images again without altering your quality settings, LoRAs, or sampling configuration.

The CUDA 12.8 Requirement Most People Miss

There is a second, related failure that often gets tangled up with this one. Blackwell cards will not run correctly with PyTorch builds compiled against older CUDA versions. As the NVIDIA developer forums and PyTorch discussions confirm, builds compiled against cu121 will not run on RTX 50-series cards at all. You need a PyTorch build compiled with CUDA 12.8 or newer to properly support the Blackwell compute capability.

This matters because the symptoms can look similar to the cuDNN error even though the root cause is different. If you are on a Blackwell card and see errors that reference an unsupported compute capability (for example sm_120 or sm_103) or experience a silent CPU fallback, check your PyTorch build’s CUDA version before anything else:

python -c "import torch; print(torch.__version__); print(torch.version.cuda)"

If the printed CUDA version is 12.1, you need to reinstall a CUDA 12.8+ build of PyTorch. Only after you are on a CUDA 12.8+ build should you expect the --use-split-cross-attention flag to be the remaining piece. These are two distinct problems with the same fix checklist, and treating them as one will waste time.

Diagnostic Checklist

Work through these in order. Each step either rules out a cause or pinpoints the exact one:

1. Confirm the card is Blackwell. An RTX 5070, 5080, or 5090 (or a laptop 50-series chip) is Blackwell. Older generations rarely trigger this specific cuDNN message.
2. Check your PyTorch CUDA build. Print torch.version.cuda. If it reads 12.1, upgrade to a CUDA 12.8+ build first; the cuDNN flag alone will not help you on a cu121 build.
3. Apply the launch flag. Start ComfyUI with --use-split-cross-attention and rerun the workflow.
4. Isolate from custom nodes. If it still fails, rule out custom nodes before blaming core ComfyUI.
5. Weigh SDPA quality vs compatibility. The split-cross-attention flag changes the attention path, so if you notice any subtle output difference, that is expected — but on Blackwell it is the path that actually works.

Why This Matters for Local AI Users

Blackwell cards are exactly the GPUs that local AI enthusiasts buy specifically to run model-heavy workflows like FLUX and video generation. The irony is that the newest hardware is precisely where the default attention path breaks, because the cuDNN backend has not yet caught up to the architecture. Understanding the difference between a driver problem, a PyTorch build problem, and a backend-dispatch problem is what separates a five-minute fix from an afternoon of reinstalling software.

The relevant facts here come from ComfyUI’s official troubleshooting documentation and from NVIDIA’s own developer forums, both of which identify the SDPA-to-cuDNN dispatch and the CUDA 12.8 build requirement as the two controlling variables. If you are running a Blackwell card and your FLUX workflow dies at the sampler, the flag in the next section is the fix that makes it work again.

Conclusion

The No valid execution plans built error is a Blackwell-specific dispatch problem, not a broken install and not a corrupted model. The fix is a single launch flag, --use-split-cross-attention, which routes attention through an explicit implementation instead of the cuDNN backend that has no plan for the new architecture. Pair it with a CUDA 12.8+ PyTorch build and the failure disappears. For the full official guidance, see the ComfyUI troubleshooting documentation, and if you are chasing related VRAM or offloading behavior, our guide to ComfyUI attention backends explains how the different paths trade speed, VRAM, and compatibility.