Skip to main content
ModelPilot

ComfyUI Kijai WanVideoWrapper crashes at model load (fp16_fast / sageattention) — fix

Also seen as: "WanVideoModelLoader crash", "fp16_fast torch 2.7", "sageattention not installed", "allow_fp16_accumulation"

Direct answer

How do I fix "WanVideoWrapper crash at load (fp16_fast / sageattention)"?

A Kijai WanVideoWrapper "advanced" graph that is supposed to be tested crashes at model load. The cause is usually a backend the graph assumes but your install does not have: base_precision=fp16_fast needs torch 2.7+, and attention_mode=sageattn needs the sageattention package. They are gated separately, so a graph moved between machines breaks when the backend is absent.

Key facts

  • ifp16_fast sets torch.backends.cuda.matmul.allow_fp16_accumulation, which does not exist before torch 2.7 — on older torch it errors at load.
  • iattention_mode=sageattn requires the sageattention package installed; without it the loader fails.
  • itorch.compile in the graph needs Triton present.

Next step

  1. 1Fastest path: set base_precision to fp16 and attention_mode to sdpa, and run eager (drop torch.compile) — this works on stock torch with no extra packages.
  2. 2If you specifically want fp16_fast, upgrade to torch 2.7+ and match your other CUDA wheels.
  3. 3If the graph hardcodes sageattn, install a sageattention wheel that matches your torch/CUDA build.
  4. 4When you do not need the wrapper features, prefer a native (non-wrapper) Wan i2v graph — it uses stock nodes with no backend gating.

If that didn't fix it

  • A "tested and working" advanced graph is tested on the author’s torch/attention build — this crash is environment mismatch, not a broken workflow.

Not sure this workflow will run before you rent a GPU?

Paste your ComfyUI workflow JSON into the free checker — it flags known dependencies and an estimated hardware requirement. It does not run the workflow, verify GPU capacity or judge output quality. No signup.

Check my workflow freeSee GPU pricing

Other ComfyUI errors

Fix

GPU Memory Exhausted

Fix

Model Not Found

Fix

Missing Python Package

Fix

Disk Space Exhausted

Fix

Gated Model Access

Fix

missing custom nodes (node type not found)

Frequently Asked Questions

What causes "WanVideoWrapper crash at load (fp16_fast / sageattention)" in ComfyUI?

fp16_fast sets torch.backends.cuda.matmul.allow_fp16_accumulation, which does not exist before torch 2.7 — on older torch it errors at load.

How do I fix "WanVideoWrapper crash at load (fp16_fast / sageattention)"?

Fastest path: set base_precision to fp16 and attention_mode to sdpa, and run eager (drop torch.compile) — this works on stock torch with no extra packages. If you specifically want fp16_fast, upgrade to torch 2.7+ and match your other CUDA wheels.