[bug]: FLUX generation crash with 6.4.0

Is there an existing issue for this problem?

• I have searched the existing issues

Install method

Invoke's Launcher

Operating system

Windows

GPU vendor

Nvidia (CUDA)

GPU model

RTX 4070 Ti SUPER

GPU VRAM

16GB

Version number

6.4.0

Browser

No response

System Information

{
"version": "6.4.0",
"dependencies": {
"absl-py" : "2.3.1",
"accelerate" : "1.10.0",
"annotated-types" : "0.7.0",
"anyio" : "4.10.0",
"attrs" : "25.3.0",
"bidict" : "0.23.1",
"bitsandbytes" : "0.47.0",
"blake3" : "1.0.5",
"certifi" : "2022.12.7",
"cffi" : "1.17.1",
"charset-normalizer" : "2.1.1",
"click" : "8.2.1",
"colorama" : "0.4.6",
"coloredlogs" : "15.0.1",
"compel" : "2.1.1",
"contourpy" : "1.3.3",
"CUDA" : "12.8",
"cycler" : "0.12.1",
"Deprecated" : "1.2.18",
"diffusers" : "0.33.0",
"dnspython" : "2.7.0",
"dynamicprompts" : "0.31.0",
"einops" : "0.8.1",
"fastapi" : "0.116.1",
"fastapi-events" : "0.12.2",
"filelock" : "3.13.1",
"flatbuffers" : "25.2.10",
"fonttools" : "4.59.1",
"fsspec" : "2024.6.1",
"gguf" : "0.17.1",
"h11" : "0.16.0",
"httptools" : "0.6.4",
"huggingface-hub" : "0.34.4",
"humanfriendly" : "10.0",
"idna" : "3.4",
"importlib_metadata" : "7.1.0",
"invisible-watermark" : "0.2.0",
"InvokeAI" : "6.4.0",
"jax" : "0.7.1",
"jaxlib" : "0.7.1",
"Jinja2" : "3.1.4",
"kiwisolver" : "1.4.9",
"MarkupSafe" : "2.1.5",
"matplotlib" : "3.10.5",
"mediapipe" : "0.10.14",
"ml_dtypes" : "0.5.3",
"mpmath" : "1.3.0",
"networkx" : "3.3",
"numpy" : "1.26.3",
"onnx" : "1.16.1",
"onnxruntime" : "1.19.2",
"opencv-contrib-python": "4.11.0.86",
"opencv-python" : "4.11.0.86",
"opt_einsum" : "3.4.0",
"packaging" : "24.1",
"picklescan" : "0.0.28",
"pillow" : "11.0.0",
"prompt_toolkit" : "3.0.51",
"protobuf" : "4.25.8",
"psutil" : "7.0.0",
"pycparser" : "2.22",
"pydantic" : "2.11.7",
"pydantic-settings" : "2.10.1",
"pydantic_core" : "2.33.2",
"pyparsing" : "3.2.3",
"PyPatchMatch" : "1.0.2",
"pyreadline3" : "3.5.4",
"python-dateutil" : "2.9.0.post0",
"python-dotenv" : "1.1.1",
"python-engineio" : "4.12.2",
"python-multipart" : "0.0.20",
"python-socketio" : "5.13.0",
"PyWavelets" : "1.9.0",
"PyYAML" : "6.0.2",
"regex" : "2025.7.34",
"requests" : "2.28.1",
"safetensors" : "0.6.2",
"scipy" : "1.16.1",
"semver" : "3.0.4",
"sentencepiece" : "0.2.1",
"setuptools" : "70.2.0",
"simple-websocket" : "1.1.0",
"six" : "1.17.0",
"sniffio" : "1.3.1",
"sounddevice" : "0.5.2",
"spandrel" : "0.4.1",
"starlette" : "0.47.2",
"sympy" : "1.13.3",
"tokenizers" : "0.21.4",
"torch" : "2.7.1+cu128",
"torchsde" : "0.2.6",
"torchvision" : "0.22.1+cu128",
"tqdm" : "4.66.5",
"trampoline" : "0.1.2",
"transformers" : "4.55.4",
"typing-inspection" : "0.4.1",
"typing_extensions" : "4.12.2",
"urllib3" : "1.26.13",
"uvicorn" : "0.35.0",
"watchfiles" : "1.1.0",
"wcwidth" : "0.2.13",
"websockets" : "15.0.1",
"wrapt" : "1.17.3",
"wsproto" : "1.2.0",
"zipp" : "3.19.2"
},
"config": {
"schema_version": "4.0.2",
"legacy_models_yaml_path": null,
"host": "0.0.0.0",
"port": 9090,
"allow_origins": [],
"allow_credentials": true,
"allow_methods": [""],
"allow_headers": [""],
"ssl_certfile": null,
"ssl_keyfile": null,
"log_tokenization": false,
"patchmatch": true,
"models_dir": "models",
"convert_cache_dir": "models\.convert_cache",
"download_cache_dir": "models\.download_cache",
"legacy_conf_dir": "configs",
"db_dir": "databases",
"outputs_dir": "M:\SD_out\images",
"custom_nodes_dir": "nodes",
"style_presets_dir": "style_presets",
"workflow_thumbnails_dir": "workflow_thumbnails",
"log_handlers": ["console"],
"log_format": "color",
"log_level": "info",
"log_sql": false,
"log_level_network": "warning",
"use_memory_db": false,
"dev_reload": false,
"profile_graphs": false,
"profile_prefix": null,
"profiles_dir": "profiles",
"max_cache_ram_gb": 48,
"max_cache_vram_gb": null,
"log_memory_usage": false,
"device_working_mem_gb": 4,
"enable_partial_loading": true,
"keep_ram_copy_of_weights": true,
"ram": null,
"vram": null,
"lazy_offload": true,
"pytorch_cuda_alloc_conf": "backend:cudaMallocAsync",
"device": "auto",
"precision": "auto",
"sequential_guidance": false,
"attention_type": "auto",
"attention_slice_size": "auto",
"force_tiled_decode": false,
"pil_compress_level": 1,
"max_queue_size": 10000,
"clear_queue_on_startup": false,
"allow_nodes": null,
"deny_nodes": null,
"node_cache_size": 512,
"hashing_algorithm": "blake3_single",
"remote_api_tokens": [
{"url_regex": "civitai.com", "token": "REDACTED"},
{"url_regex": "^[a-z]{3}whatever.*\.com$", "token": "REDACTED"}
],
"scan_models_on_startup": false,
"unsafe_disable_picklescan": false
},
"set_config_fields": [
"device_working_mem_gb", "legacy_models_yaml_path", "remote_api_tokens", "pytorch_cuda_alloc_conf",
"host", "max_cache_ram_gb", "enable_partial_loading", "outputs_dir"
]
}

What happened

Clicked the Invoke button in Generate mode using FLUX.1 dev model. Shortly afterwards the process ends and the app hangs:

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

What you expected to happen

I expected an image to be generated.

How to reproduce the problem

1. Open the Invoke app with Launcher 1.7.0
2. Choose FLUX.1 dev model
3. Put in a simple prompt, a few words
4. Choose 1024 x 1024 frame
5. Click INVOKE.

Additional context

This happens only with FLUX models. Issue is not apparent using SDXL models.
I re-installed Invoke using 6.3.0 and 1.6.0 Launcher; the issue remains, but I did not have the problem when using FLUX models before the upgrade to 6.4.0

Discord username

VE1LEB

[psychedelicious]

From our brief discussion on discord:

I cannot reproduce the issue on Linux.

I've reviewed the changes from v6.3.0 to v6.4.0 and cannot find anything that touched the FLUX text encoder or model loading code.

The problem is occurring during the load of the tokenizer. That happens here:
https://github.com/invoke-ai/InvokeAI/blob/main/invokeai/app/invocations/flux_text_encoder.py#L77-L81

When we call context.models.load(self.t5_encoder.tokenizer) internally we are using the transformers library to load the tokenizer. Here's the upstream class https://github.com/huggingface/transformers/blob/4f9b4e62bc52a52b19a6a4a1a6bfc61a3f5b65b1/src/transformers/models/t5/tokenization_t5_fast.py

The warning message about add_prefix_space happens during initialization of the tokenizer class:
https://github.com/huggingface/transformers/blob/4f9b4e62bc52a52b19a6a4a1a6bfc61a3f5b65b1/src/transformers/models/t5/tokenization_t5_fast.py#L113-L117

Therefore, we know that we are getting past that point in the tokenizer init. But we don't get a log about loading the tokenizer successfully. This is what we should see:

[2025-08-25 14:40:10,100]::[ModelManagerService]::INFO --> [MODEL CACHE] Loaded model '186402f1-8a71-4c03-bb73-b3d2dafc8068:text_encoder_3' (T5EncoderModel) onto cuda device in 2.53s. Total model size: 11003.39MB, VRAM: 11003.39MB (100.0%)
You set `add_prefix_space`. The tokenizer needs to be converted from the slow tokenizers
# We don't make it past this point!
[2025-08-25 14:40:10,199]::[ModelManagerService]::INFO --> [MODEL CACHE] Loaded model '186402f1-8a71-4c03-bb73-b3d2dafc8068:tokenizer_3' (T5TokenizerFast) onto cuda device in 0.00s. Total model size: 0.03MB, VRAM: 0.00MB (0.0%)

With no changes on our end, it could be that the transformers library, which handles the tokenizer, is the problem.

Let's try rolling back to a previous version of transformers. In the Launcher, open & start the dev console (bottom left corner). Run this command:

uv pip install transformers==4.54.1

It should show something like this, indicating it uninstalled the current version and installed an earlier version:

Resolved 18 packages in 110ms
Prepared 1 package in 327ms
Uninstalled 1 package in 29ms
Installed 1 package in 33ms
 - transformers==4.55.4
 + transformers==4.54.1

Then start up Invoke and see if the issue persists

---

Then you tried the rollback, but it didn't fix the issue

[psychedelicious]

Next step is to test things without the Launcher to check if it is somehow interfering.

• Open PowerShell
• Navigate to your Invoke install directory by running cd C:\Users\MBL\invokeai\
• Run .venv\Scripts\Activate.ps1 to activate the virtual environment
• Run invokeai-web to start Invoke
• Look for a message saying the server is ready w/ a URL
• Go to that URL in a web browser
• Try to generate with FLUX

Does it crash the same way?

[psychedelicious]

User Tom Schülke on discord used ChatGPT to troubleshoot and was able to fix the issue. Summarizing their changes:

• Downgrade several packages with uv pip install --force-reinstall "transformers==4.44.2" "tokenizers==0.19.1" "safetensors==0.4.5" "sentencepiece==0.2.0"
• Replace T5TokenizerFast with T5Tokenizer in the Invoke codebase. The "fast" version is a tokenizer implemented in rust instead of python and is supposed to be a safe drop-in replacement.

So, @firesign , could you please try the more thorough downgrade command?

Also, I'll try to reproduce the issue on Windows at some point today.

Next step is to test things without the Launcher to check if it is somehow interfering.

• Open PowerShell
• Navigate to your Invoke install directory by running cd C:\Users\MBL\invokeai\
• Run .venv\Scripts\Activate.ps1 to activate the virtual environment
• Run invokeai-web to start Invoke
• Look for a message saying the server is ready w/ a URL
• Go to that URL in a web browser
• Try to generate with FLUX

Does it crash the same way?

I tried running the .venv script in PowerShell (as Admin) but "running scripts is disabled on this system".

[psychedelicious]

Ah, right. Run this before the .venv script:

Set-ExecutionPolicy RemoteSigned -Scope CurrentUser

Or, you can run the .venv script like this - a slightly less permissive way to run it:

powershell -ExecutionPolicy Bypass -File .venv\Scripts\Activate.ps1

Got the .venv script to run and started it up. I initiated a FLUX gen but it ended with the same crash as before.

I'll try your other suggestion of replacing the tokenizers, if I can figure out how to do that...
"I'm an artist, not a software developer, Jim!"

I got stopped at replacing the tokenizer. I quit invoke and did a repair, but when I started up from Launcher, Invoke gave lots of red error lines after announcing database initialization. 😒
I've saved the log but I'll try rebooting my machine and see if it improves its behaviour.

[psychedelicious]

Thanks for testing.

PS: Sorry I didn't get to testing on Windows today (it's the same computer I do my normal dev work on, but I rarely boot into Windows).

Thank you for helping us solve this. BTW, my invoke_demo install still seems broken. Fortunately it's not my production version so I can try flipping to a backup copy, and worst-case do a relatively quick reinstall.
I think I'll take a break though and let you and Tom work on it.