✅ [SOLVED] CUDA error: no kernel image on RTX 5070 Ti (PyTorch 2.8dev)

[RyuKwangChoon]

Environment:

• GPU: RTX 5070 Ti
• CUDA: 12.8 / 12.9
• PyTorch: 2.8.0.dev + cu128
• ComfyUI version: latest (May 2025)

✅ ComfyUI CUDA Error Analysis & Final Resolution Summary (MARKONE Version)

🎯 Objective

Resolve the error:

"CUDA error: no kernel image is available for execution on the device"
that occurs when running ComfyUI on RTX 5070 Ti (Ada Lovelace / Compute 12.0)

---

🧩 1. Problem Diagnosis (Initial Attempts)

[1] GPU & Environment

• GPU: NVIDIA GeForce RTX 5070 Ti
• CUDA Compute Capability: (12, 0)
• Issue: Official PyTorch builds did not yet support sm_120, leading to kernel execution failures.

[2] PyTorch Reinstallation Attempts

• Versions tested:
  • 2.5.1+cu121, 2.2.2+cu121, and Nightly builds
• Installation command example:

pip install --pre torch torchvision torchaudio --index-url ...

• Result: All builds failed with the same error due to sm_120 not supported.

[3] Execution Parameter Tweaks

• Tried parameters:
  • TORCH_USE_CUDA_DSA=1
  • CUDA_LAUNCH_BLOCKING=1
  • --disable-cuda-malloc
• Result: Useful for debugging, but did not resolve core CUDA kernel issue.

[4] ComfyUI Template/Node Issues

• Failed nodes: CLIPTextEncode, TextPrompt
• Even with custom JSON templates or simplified node graphs, CUDA kernel issue persisted

[5] System & Path Info

• ComfyUI location: C:\StabilityMatrix\Packages\ComfyUI
• Models: C:\StabilityMatrix\Models\...
• Launch command: python main.py --preview-method auto
• Observation: GUI opened, but crashed when encoding prompts using GPU.

⚠️ Preliminary Conclusion

• RTX 5070 Ti (sm_120) not yet supported in stable PyTorch CUDA kernels
• Key nodes such as CLIPTextEncode are unusable on GPU

Suggested Workarounds at the Time

1. Run in CPU mode
2. Wait for PyTorch sm_120 kernel support
3. Compile custom CUDA kernels (advanced)
4. Design workflows that bypass CLIP dependencies

---

🚀 2. Final Working Solution (Our Debugging Session)

✅ Strategy

• Switch to latest PyTorch nightly build with CUDA 12.8, which includes sm_120 kernel support.
• Resolve missing Python module errors by reinstalling dependencies via requirements.txt

🛠️ Step-by-Step Actions

Step	Description
✅ PyTorch Update	Installed 2.8.0.dev + cu128 (with sm_120 kernel)
✅ Driver Check	nvidia-smi confirmed CUDA 12.9 driver support
✅ Dependency Restore	pip install -r requirements.txt executed successfully
✅ GUI Fix	Accessed manually via http://127.0.0.1:8188
✅ Execution Script	Customized .bat file with CUDA config and precision flags
✅ Performance Test	GPU fully functional, <1s response time

---

🔍 Additional Discovery

Initial Installation Path

• Likely installed a lightweight ComfyUI distribution (e.g., comfyui-lightpack),
  which omitted core modules like aiohttp, yarl, uvicorn, etc.

Resolution

• Switched from minimal install to full dependency setup via requirements.txt
• Recognized the lightpack shortcut as a partial cause of errors

---

🎉 Final Conclusion

• RTX 5070 Ti + PyTorch 2.8.0.dev + CUDA 12.8 = Fully Compatible
• Manual fixes included:
  • Switching to the right PyTorch dev version
  • Installing missing modules
  • Adjusting GUI launch behavior
  • Reviewing .bat configuration
• This solution is not a workaround — it's a fully working custom environment setup

---

📁 Suggested .bat Script for RTX 5070 Ti

@echo off
set PYTORCH_CUDA_ALLOC_CONF=max_split_size_mb:128
set "CUDA_VISIBLE_DEVICES=0"
python main.py --precision float16 --preview-method direct
pause