Environment
Windows 11 24H2
RTX 3060 12 GB (desktop)
16 GB RAM
ComfyUI Windows Portable (NVIDIA)
ComfyUI-FramePackWrapper_PlusOne 2.0.1
VideoHelperSuite installed and working
Using the bundled framepack_F1_example.json
Model:
FramePackI2V_HY_fp8_e4m3fn.safetensors
Problem
The workflow starts normally.
The model begins loading.
Progress consistently reaches approximately 7%.
ComfyUI then terminates with:
Windows fatal exception: access violation
The browser reports Reconnecting because the backend crashes.
The crash occurs during construction of the Hunyuan model.
Stack trace includes:
torch\nn\modules\linear.py
diffusers\models\attention_processor.py
hunyuan_video_packed.py
nodes.py
Specifically around:
self.proj_out = nn.Linear(hidden_size + mlp_dim, hidden_size)
Things already tested
Fresh ComfyUI Windows Portable
VideoHelperSuite reinstalled
Missing dependencies installed
Correct FramePack model selected
Workflow loads correctly
Video Combine fixed
base_precision
bf16
fp16
load_device
offload_device
main_device
Transformers downgraded from 5.9.0 to 4.49.0
CUDA test:
torch.nn.Linear(4096,4096,device="cuda",dtype=torch.float16)
passes successfully.
Memory is not exhausted.
No CUDA OOM occurs.
The crash is always a native Windows access violation.
Question
Is this a known issue with:
FramePackWrapper_PlusOne 2.0.1
HunyuanVideoClipModel
Torch 2.12 / CUDA 13
or RTX 3060 systems?
Has anyone reproduced this or found a workaround?
Environment
Windows 11 24H2
RTX 3060 12 GB (desktop)
16 GB RAM
ComfyUI Windows Portable (NVIDIA)
ComfyUI-FramePackWrapper_PlusOne 2.0.1
VideoHelperSuite installed and working
Using the bundled framepack_F1_example.json
Model:
FramePackI2V_HY_fp8_e4m3fn.safetensors
Problem
The workflow starts normally.
The model begins loading.
Progress consistently reaches approximately 7%.
ComfyUI then terminates with:
Windows fatal exception: access violation
The browser reports Reconnecting because the backend crashes.
The crash occurs during construction of the Hunyuan model.
Stack trace includes:
torch\nn\modules\linear.py
diffusers\models\attention_processor.py
hunyuan_video_packed.py
nodes.py
Specifically around:
self.proj_out = nn.Linear(hidden_size + mlp_dim, hidden_size)
Things already tested
Fresh ComfyUI Windows Portable
VideoHelperSuite reinstalled
Missing dependencies installed
Correct FramePack model selected
Workflow loads correctly
Video Combine fixed
base_precision
bf16
fp16
load_device
offload_device
main_device
Transformers downgraded from 5.9.0 to 4.49.0
CUDA test:
torch.nn.Linear(4096,4096,device="cuda",dtype=torch.float16)
passes successfully.
Memory is not exhausted.
No CUDA OOM occurs.
The crash is always a native Windows access violation.
Question
Is this a known issue with:
FramePackWrapper_PlusOne 2.0.1
HunyuanVideoClipModel
Torch 2.12 / CUDA 13
or RTX 3060 systems?
Has anyone reproduced this or found a workaround?