Updated from Forge Neo 2.7 to 2.11. Wan2.2 gives a black output.

[lucchesi590-coder]

Hi, am a layman and just as the title says, Wan2.2 is giving a black output now after I updated to 2.11.

Using launch args: --uv --cuda-stream --pin-shared-memory --cuda-malloc --nunchaku --sage
There are no errors whatsoever on the terminal.

I have tried changing:
Diffusion in low bits
Changing launch arguements
The text encoders
And other pretty basic stuff...

I am using the Quantized versions of the text encoder and the wan model. Using a rtx 5060 8gb(no hate im super new) and 32GB system ram.

Help appreciated thanks, I really enjoyed Wan.

[Haoming02]

Try --disable-sage

[fadedninna]

Try --disable-sage

or just put random gibberish in negative

[Haoming02]

or just put random gibberish in negative

That's only an issue when using SDXL, and has been fixed long ago

[lucchesi590-coder]

Try --disable-sage

No luck
I did see this however..

F:\StableDiffusion\Forge Neo UI\sd-webui-forge-neo\modules\processing.py:1049: RuntimeWarning: invalid value encountered in cast
x_sample = x_sample.astype(np.uint8)

Everything else is working

[fadedninna]

F:\StableDiffusion\Forge Neo UI\sd-webui-forge-neo\modules\processing.py:1049: RuntimeWarning: invalid value encountered in cast
x_sample = x_sample.astype(np.uint8)

In comfyui, i sometimes too get this error while using sage attention, i simply restart my system and it works fine after that

[lucchesi590-coder]

F:\StableDiffusion\Forge Neo UI\sd-webui-forge-neo\modules\processing.py:1049: RuntimeWarning: invalid value encountered in cast
x_sample = x_sample.astype(np.uint8)

In comfyui, i sometimes too get this error while using sage attention, i simply restart my system and it works fine after that

I've restarted 50+ times trying to troubleshoot lol

[Haoming02]

invalid value encountered in cast

That's just the result, not the cause

---

Post the full log

[lucchesi590-coder]

invalid value encountered in cast

That's just the result, not the cause

Post the full log

venv "F:\StableDiffusion\Forge Neo UI\sd-webui-forge-neo\venv\Scripts\Python.exe"
Python 3.11.9 (tags/v3.11.9:de54cf5, Apr  2 2024, 10:12:12) [MSC v.1938 64 bit (AMD64)]
Version: neo
Launching Web UI with arguments: --uv --pin-shared-memory --cuda-stream --cuda-malloc --nunchaku --sage
Total VRAM 8150 MB, total RAM 32681 MB                                                  memory_management.py :: INFO
PyTorch Version: 2.9.1+cu130                                                            memory_management.py :: INFO
VRAM State: NORMAL_VRAM                                                                 memory_management.py :: INFO
Device: NVIDIA GeForce RTX 5060 (cuda:0) - cudaMallocAsync                              memory_management.py :: INFO
Using async weight offloading with 2 streams                                            memory_management.py :: INFO
Pinned Memory: 14706 MB                                                                 memory_management.py :: INFO
Using SageAttention 2                                                                           attention.py :: INFO
Using PyTorch Attention for VAE                                                                 attention.py :: INFO
ControlNet preprocessor location: F:\StableDiffusion\Forge Neo UI\sd-webui-forge-neo\models\ControlNetPreprocessor
[-] ADetailer initialized. version: 25.3.0, num models: 11
[Forge Couple] INFO - Loaded Adv. Presets...
[Forge Couple] INFO - Loaded Adv. Presets...
[ControlNet] - INFO - ControlNet UI callback registered.
Model Selected:                                                                                main_entry.py :: INFO
{
  "checkpoint": "Wan2.2-I2V-A14B-LowNoise-Q3_K_S.gguf",
  "modules": [
    "Wan2.1_VAE.safetensors",
    "umt5-xxl-encoder-Q5_K_M.gguf"
  ],
  "dtype": "[torch.float16, torch.bfloat16]"
}
Patch LoRAs on-the-fly: True                                                                   main_entry.py :: INFO
Running on local URL:  http://127.0.0.1:7860

To create a public link, set `share=True` in `launch()`.
Startup time: 23.6s (prepare environment: 3.5s, launcher: 0.5s, forge init: 6.9s, shared init: 0.3s, misc. imports: 4.7s, load scripts: 3.2s, create ui: 2.7s, gradio launch: 1.9s).
Loading Model: {'checkpoint_info': {'filename': 'F:\\StableDiffusion\\Forge Neo UI\\sd-webui-forge-neo\\models\\Stable-diffusion\\WAN2.2\\Wan2.2-I2V-A14B-LowNoise-Q3_K_S.gguf', 'hash': '1862bfe0'}, 'additional_modules': ['F:\\StableDiffusion\\Forge Neo UI\\sd-webui-forge-neo\\models\\VAE\\Wan2.1_VAE.safetensors', 'F:\\StableDiffusion\\Forge Neo UI\\sd-webui-forge-neo\\models\\text_encoder\\umt5-xxl-encoder-Q5_K_M.gguf'], 'unet_storage_dtype': None}
Using Detected T5XXL Data Type: gguf (pre-quant)                                                   loader.py :: INFO
GGUF state dict: {'Q5_K': 144, 'Q6_K': 24}
Using Detected Model Data Type: gguf (pre-quant)                                                   loader.py :: INFO
GGUF state dict: {'Q3_K': 400}
Diffusion Model: {storage: gguf, computation: torch.float16}                                      k_model.py :: INFO
Model loaded in 44.4s (unload existing model: 0.2s, forge model load: 44.2s).
[NeverOOM] Enabled for UNet (always offload)                                              forge_never_oom.py :: INFO
Changed VRAM State to NO_VRAM                                                             forge_never_oom.py :: INFO
Requested to load WanVAE                                                                memory_management.py :: INFO
loaded partially; 0.00 MB usable, 0.00 MB loaded, 242.00 MB offloaded, 22.78 MB buffer reserved,     base.py :: INFO
lowvram patches: 0
Moving model(s) has taken 0.04 seconds                                                  memory_management.py :: INFO
[LORA] Loading wan2.2_i2v_A14b_low_noise_lora_rank64_lightx2v_4step_1022.safetensors for KModel  networks.py :: INFO
with 290 unmatched keys
[LORA] Loaded wan2.2_i2v_A14b_low_noise_lora_rank64_lightx2v_4step_1022.safetensors for          networks.py :: INFO
KModel-UNet with 1054 keys at weight 1.0 (skipped 0 keys) with on_the_fly = True
Skipping unconditional conditioning when CFG = 1. Negative Prompts are ignored.
Requested to load JointTextEncoder                                                      memory_management.py :: INFO
loaded completely; 0.00 MB usable, 0.00 MB loaded, full load: False                                  base.py :: INFO
Moving model(s) has taken 0.05 seconds                                                  memory_management.py :: INFO
Requested to load KModel                                                                memory_management.py :: INFO
loaded partially; 0.00 MB usable, 0.00 MB loaded, 3.76 MB offloaded, 400.08 MB buffer reserved,      base.py :: INFO
lowvram patches: 0
Moving model(s) has taken 0.97 seconds                                                  memory_management.py :: INFO
100%|████████████████████████████████████████████████████████████████████████████████████| 4/4 [01:20<00:00, 20.14s/it]
Requested to load WanVAE                                                                memory_management.py :: INFO
loaded partially; 0.00 MB usable, 0.00 MB loaded, 242.00 MB offloaded, 22.78 MB buffer reserved,     base.py :: INFO
lowvram patches: 0
Moving model(s) has taken 0.85 seconds                                                  memory_management.py :: INFO
F:\StableDiffusion\Forge Neo UI\sd-webui-forge-neo\modules\processing.py:1049: RuntimeWarning: invalid value encountered in cast
  x_sample = x_sample.astype(np.uint8)
Total progress: 100%|████████████████████████████████████████████████████████████████████| 4/4 [01:09<00:00, 17.40s/it]

Please note I've tried:

• Running it with/without the Lora
• With/Without the high noise model with the refiner
• With/Without Sage

[Haoming02]

Are you sure you were using umt5-xxl-encoder-Q5_K_M.gguf before?

[lucchesi590-coder]

Are you sure you were using umt5-xxl-encoder-Q5_K_M.gguf before?

100%, infact its the only one I had.

[Haoming02]

It's just that... umt5-xxl-encoder-Q5_K_M.gguf is giving me NaN (i.e. black outputs) too even after downgrading to 2.7

---

While on 2.11, using any other text encoder (e.g. umt5-xxl-fp8-scaled.safetensors) generates just fine

[lucchesi590-coder]

It's just that... umt5-xxl-encoder-Q5_K_M.gguf is giving me NaN (i.e. black outputs) too even after downgrading to 2.7

While on 2.11, using any other text encoder (e.g. umt5-xxl-fp8-scaled.safetensors) generates just fine

Odd, I can say for certain because I only had SD1.5 and WAN2.2. I tried umt5-xxl-encoder-Q3_K_S.gguf shortly after updating when I was getting the black outputs but to no avail.

But anyways I've downloaded umt5_xxl_fp8_e4m3fn_scaled.safetensors now, and its working.. Thank you