Wan2.1 LowVram Common Errors

🧩 SageAttention: ImportError

Error message:
ImportError: cannot import name 'sageattn_qk_int8_pv_fp16_cuda' from 'sageattention'

Root Causes & Solutions:

• Version mismatch: The function might not exist in your installed version.

• Fix: Reinstall the correct SageAttention:

• Clone the official repo and install:

git clone https://github.com/thu-ml/SageAttention.git cd SageAttention pip install -e .

This ensures all kernels, including sageattn_qk_int8_pv_fp16_cuda, are built GitHub+4Reddit+4GitHub+4arXiv+15Reddit+15Reddit+15GitHubYouTube+12048.csdn.net+1.

1. Install Triton first; it's required for many SageAttention variants Patreon+1YouTube+1.

2. For v2, use CUDA ≥ 12.4 and Torch ≥ 2.3 Reddit+2GitHub+22048.csdn.net+2.

Reddit users confirm this approach solved the import error CSDN Blog+9Reddit+9GitHub+9.

🛠 Missing or Misplaced GGUF Model

Error message:
LoaderGGUF: gguf_name 'None' not in []

Cause: Model file not named, recognized, or placed correctly.

Solution:

• Download the .gguf file.

Put it in ComfyUI’s correct folder:

ComfyUI/models/diffusion_models/

• Refresh the UI or restart, and the model should appear.

🧩 “Failed to find C compiler”

Cause: SageAttention or Triton can't build native components.

Solution:

• Windows: Install Visual Studio C++ tools.

• Linux: sudo apt install build-essential.

• Then reinstall SageAttention.

🖥 Low VRAM GPUs (e.g., GTX 960)

Symptoms: Workflow fails or crashes due to memory limitations.

Suggestions:

• Use low-res (e.g. 256×256) single-frame runs or text-to-video only.

• Try lightweight models or versions with reduced complexity (e.g., Wan 2.1).

• Follow tips shared by users for 2 GB or 4 GB setups.

⚙️ Triton & Torch Compatibility

Common issue:
“Too old GPU for Triton” or “compile_transformer_blocks” errors.

Fixes:

• In older GPUs (e.g., GTX 1070), disable Triton-related options:

  • Set compile_transformer_blocks = false in “Torch WAN Compiler” node arXiv+10GitHub+10GitHub+10GitHub+6Reddit+6YouTube+6.

• If using newer GPUs (RTX 30–40 series):

Ensure Triton and SageAttention are both installed using:

python_embeded\python.exe -m pip install -U triton-windows python_embeded\python.exe -m pip install -U sageattention

You might also need to manually add include/ and libs/ folders into python_embeded GitHub+4Patreon+4Reddit+4.

• Restart ComfyUI after the initial compilation (a black screen is normal on first run) GitHub+7Reddit+7GitHub+7.

🧱 Model File Placement & Loader Errors

Issues:

• “Unknown model architecture!” in UnetLoaderGGUFAdvanced.

• Missing safetensors for encoders.

Solutions:

• Use supported GGUF variants.

• Install required components:

  • Encoder .safetensors (e.g., umt5_xxl_fp8_e4m3fn_scaled.safetensors) must be in the right folder.

  • Place files under: models/diffusion_models/, clip/, vae/.

• Refresh the interface after placement.

• ComfyUI Manager can automate placing files.

🌠 i2v Output Quality Drops

Problem: Video interpolation (i2v) quality is poor vs t2v.

Recommended settings by the_frizz1:

• Sampler: uni_pc_bh2

• Denoise level: 1
  These settings significantly improve i2v results.

📦 Additional Tools & Tips

• TeaCache: Optional caching extension—additional 20–30% speed boost if installed from KJNodes/ComfyUI manager 2048.csdn.netGitHubReddit+3Reddit+3GitHub+3.

• Frame Interpolation (RIFE/VFI): Use plugins like ComfyUI-Frame-Interpolation by Fannovel16 for smoother outputs.

• Start/End Frame Support: Requires high VRAM (12 GB+) or different workflow nodes.

• Zluda + AMD GPUs (7800 XT) – may still face compatibility issues.

✅ Step-by-Step Setup Checklist

1. Install Triton first (triton-windows or pip install triton).

2. Clone and install SageAttention from source.

3. Install TeaCache (optional via ComfyUI Manager).

4. Place models properly: .gguf, encoders, VAEs, etc.

5. Disable compile_transformer_blocks if on older Nvidia hardware.

6. Restart ComfyUI after first run of SageAttention node.

7. Use recommended sampler & denoise settings for stable video (i2v).