[bug]: App crash after every generation

[vrubzov1957]

Is there an existing issue for this problem?

• [x] I have searched the existing issues

Install method

Invoke's Launcher

Operating system

Windows

GPU vendor

Nvidia (CUDA)

GPU model

1080 Ti 11G

GPU VRAM

11

Version number

6.4.0

Browser

Chrome

System Information

No response

What happened

When I try generate image (I use different flux models) - app crashes every time

I get every time:

[2025-08-23 22:28:42,995]::[InvokeAI]::INFO --> Executing queue item 17429, session d440903f-2df6-40a8- 86b8-2a01cd25b271 [2025-08-23 22:29:17,439]::[ModelManagerService]::INFO --> [MODEL CACHE] Loaded model 'dff561c5-d978-4b e6-a4ba-876c294e218e:text_encoder_2' (T5EncoderModel) onto cuda device in 32.99s. Total model size: 466 7.39MB, VRAM: 4667.39MB (100.0%) You set add_prefix_space. The tokenizer needs to be converted from the slow tokenizers [2025-08-23 22:29:17,981]::[ModelManagerService]::INFO --> [MODEL CACHE] Loaded model 'dff561c5-d978-4b e6-a4ba-876c294e218e:tokenizer_2' (T5TokenizerFast) onto cuda device in 0.00s. Total model size: 0.03MB , VRAM: 0.00MB (0.0%) D:\InvokeAI_640.venv\Lib\site-packages\bitsandbytes\autograd_functions.py:186: UserWarning: MatMul8bi tLt: inputs will be cast from torch.bfloat16 to float16 during quantization warnings.warn(f"MatMul8bitLt: inputs will be cast from {A.dtype} to float16 during quantization") Error named symbol not found at line 448 in file D:\a\bitsandbytes\bitsandbytes\csrc\ops.cu Invoke process exited with code 1

So, I cant generate even one image in InvokeAI 6.4.0 on my machine, where 5.6.0 is worked

What you expected to happen

I expected - normal generation of images

How to reproduce the problem

Start launcher, start app (with Server mode), try generate image

Additional context

Crashes even models (like Flux Dev NF4), which works in InvokeAI 5.6.0 version on this machine

Discord username

No response

[firesign]

I have a similar problem, but only with FLUX text-to-image checkpoints:

Started Invoke process with PID 1900 [2025-08-21 08:30:07,539]::[InvokeAI]::INFO --> PyTorch CUDA memory allocator: cudaMallocAsync [2025-08-21 08:30:25,810]::[InvokeAI]::INFO --> Using torch device: NVIDIA GeForce RTX 4070 Ti SUPER [2025-08-21 08:30:26,847]::[InvokeAI]::INFO --> cuDNN version: 90701 [2025-08-21 08:30:29,424]::[InvokeAI]::INFO --> Patchmatch initialized [2025-08-21 08:30:30,029]::[InvokeAI]::INFO --> Loading node pack chroma_invoke-main [2025-08-21 08:30:30,160]::[InvokeAI]::INFO --> Loading node pack my_node_pack [2025-08-21 08:30:30,238]::[InvokeAI]::INFO --> Loaded 2 node packs from C:\Users\MBL\invokeai\nodes: chroma_invoke-main, my_node_pack [2025-08-21 08:30:30,324]::[InvokeAI]::INFO --> InvokeAI version 6.4.0 [2025-08-21 08:30:30,324]::[InvokeAI]::INFO --> Root directory = C:\Users\MBL\invokeai [2025-08-21 08:30:30,325]::[InvokeAI]::INFO --> Initializing database at C:\Users\MBL\invokeai\databases\invokeai.db [2025-08-21 08:30:30,575]::[ModelManagerService]::INFO --> [MODEL CACHE] Calculated model RAM cache size: 13303.50 MB. Heuristics applied: [1, 2]. [2025-08-21 08:30:33,734]::[InvokeAI]::INFO --> Cleaned database (freed 0.22MB) [2025-08-21 08:30:33,734]::[InvokeAI]::INFO --> Invoke running on http://0.0.0.0:9090 (Press CTRL+C to quit) [2025-08-21 08:32:25,833]::[InvokeAI]::INFO --> Executing queue item 45571, session e505b435-feeb-4e72-9a4b-019ec31c1b2b Loading checkpoint shards: 100% ███████████████████████████████████████████████████████████████████████████████████████████████████████████| 2/2 [00:00<00:00, 5.61it/s] [2025-08-21 08:32:56,962]::[ModelManagerService]::INFO --> [MODEL CACHE] Loaded model 'f8da12df-ea4a-470a-bb54-7cec8c6b2cd7:text_encoder_2' (T5EncoderModel) onto cuda dev ice in 30.60s. Total model size: 9334.39MB, VRAM: 9334.39MB (100.0%) You set add_prefix_space. The tokenizer needs to be converted from the slow tokenizers Invoke process exited with code 3221225477

The issue started after moving to the new launcher and invoke v6.4.0 (from v6.3.0, which worked fine). I've also tried rolling back to v6.3.0, but the issue was not resolved.

[bug-philosopher]

You are running into a bitsandbytes problem on Windows. That Flux model is trying to use a CUDA kernel that your install doesn't have, so Invoke just quits.

I assume that's because of the 1080Ti, its based on Pascal architecture, and bitsandbytes either dropped or didn't really support pascal cards for the quantized kernels. I ran into same issues using older nvidia gpus because these models depend on CUDA features that only exist on newer nvidia gpu. My take is that bitsandbytes can’t find the right CUDA functions, so it throws the “named symbol not found” error and crashes Invoke.

Can you still generate images with SDXL or other models?