KSampler timings incorrect when using FP8 models in a Wan 2.2 I2V workflow

### Custom Node Testing

- [ ] I have tried disabling custom nodes and the issue persists (see [how to disable custom nodes](https://docs.comfy.org/troubleshooting/custom-node-issues#step-1%3A-test-with-all-custom-nodes-disabled) if you need help)

### Expected Behavior

The console shows the correct duration for the KSampler steps (High and Low noise)
For example :
0%| | 0/2 [00:00 Execution > Live preview method'.
[2026-01-05 16:44:37.758] ### ComfyUI Revision: 150 [f59f71cf] *DETACHED | Released on '2025-12-30'
[2026-01-05 16:44:37.764] [ComfyUI-Manager] All startup tasks have been completed.
[2026-01-05 16:44:37.765]
Import times for custom nodes:
[2026-01-05 16:44:37.765] 0.0 seconds: C:\Comfy\ComfyUI\custom_nodes\websocket_image_save.py
[2026-01-05 16:44:37.765] 0.0 seconds: C:\Comfy\ComfyUI\custom_nodes\ComfyUI-GGUF
[2026-01-05 16:44:37.765] 0.1 seconds: C:\Comfy\ComfyUI\custom_nodes\comfyui-manager
[2026-01-05 16:44:37.765]
[2026-01-05 16:44:38.075] Context impl SQLiteImpl.
[2026-01-05 16:44:38.075] Will assume non-transactional DDL.
[2026-01-05 16:44:38.076] No target revision found.
[2026-01-05 16:44:38.102] Starting server

[2026-01-05 16:44:38.103] To see the GUI go to: http://127.0.0.1:8188
[2026-01-05 16:44:38.587] [DEPRECATION WARNING] Detected import of deprecated legacy API: /scripts/ui.js. This is likely caused by a custom node extension using outdated APIs. Please update your extensions or contact the extension author for an updated version.
[2026-01-05 16:44:38.591] [DEPRECATION WARNING] Detected import of deprecated legacy API: /extensions/core/groupNode.js. This is likely caused by a custom node extension using outdated APIs. Please update your extensions or contact the extension author for an updated version.
[2026-01-05 16:44:38.780] [DEPRECATION WARNING] Detected import of deprecated legacy API: /scripts/ui/components/buttonGroup.js. This is likely caused by a custom node extension using outdated APIs. Please update your extensions or contact the extension author for an updated version.
[2026-01-05 16:44:38.783] [DEPRECATION WARNING] Detected import of deprecated legacy API: /scripts/ui/components/button.js. This is likely caused by a custom node extension using outdated APIs. Please update your extensions or contact the extension author for an updated version.
[2026-01-05 16:44:52.655] got prompt
[2026-01-05 16:44:52.669] Using split attention in VAE
[2026-01-05 16:44:52.670] Using split attention in VAE
[2026-01-05 16:44:52.793] VAE load device: cuda:0, offload device: cpu, dtype: torch.bfloat16
[2026-01-05 16:44:52.881] Found quantization metadata version 1
[2026-01-05 16:44:52.882] Using MixedPrecisionOps for text encoder
[2026-01-05 16:44:53.813] CLIP/text encoder model load device: cuda:0, offload device: cpu, current: cpu, dtype: torch.float16
[2026-01-05 16:44:53.814] Requested to load WanTEModel
[2026-01-05 16:44:59.631] loaded completely; 22891.42 MB usable, 6419.49 MB loaded, full load: True
[2026-01-05 16:45:01.697] Requested to load WanVAE
[2026-01-05 16:45:02.721] loaded completely; 21330.75 MB usable, 242.03 MB loaded, full load: True
[2026-01-05 16:45:06.778] Found quantization metadata version 1
[2026-01-05 16:45:06.778] Detected mixed precision quantization
[2026-01-05 16:45:06.779] Using mixed precision operations
[2026-01-05 16:45:06.791] model weight dtype torch.float16, manual cast: torch.float16
[2026-01-05 16:45:06.791] model_type FLOW
[2026-01-05 16:45:06.845] unet unexpected: ['scaled_fp8']
[2026-01-05 16:45:06.914] Requested to load WAN21
[2026-01-05 16:45:26.361] loaded completely; 20567.30 MB usable, 13631.43 MB loaded, full load: True
[2026-01-05 16:45:38.633]
100%|████████████████████████████████████████████████████████████████████████████████████| 2/2 [00:12<00:00, 7.09s/it]
100%|████████████████████████████████████████████████████████████████████████████████████| 2/2 [00:12<00:00, 6.09s/it]
[2026-01-05 16:45:50.505] Found quantization metadata version 1
[2026-01-05 16:45:50.505] Detected mixed precision quantization
[2026-01-05 16:45:50.505] Using mixed precision operations
[2026-01-05 16:45:50.680] model weight dtype torch.float16, manual cast: torch.float16
[2026-01-05 16:45:50.680] model_type FLOW
[2026-01-05 16:45:50.743] unet unexpected: ['scaled_fp8']
[2026-01-05 16:45:50.823] Requested to load WAN21
[2026-01-05 16:46:12.139] loaded completely; 20433.93 MB usable, 13631.43 MB loaded, full load: True
[2026-01-05 16:46:23.908]
100%|████████████████████████████████████████████████████████████████████████████████████| 2/2 [00:11<00:00, 6.88s/it]
100%|████████████████████████████████████████████████████████████████████████████████████| 2/2 [00:11<00:00, 5.87s/it]
[2026-01-05 16:46:35.686] Requested to load WanVAE
[2026-01-05 16:46:36.043] loaded completely; 20779.94 MB usable, 242.03 MB loaded, full load: True
[2026-01-05 16:46:42.665] Prompt executed in 110.01 seconds
```

### Other

The console shows:
loaded completely; 20567.30 MB usable, 13631.43 MB loaded, full load: True
50%|██████████ | 1/2 [00:00<00:00, 2.21it/s]
100%|████████████████████| 2/2 [00:12<00:00, 6.09s/it]
=> sugesting this High Noise KSampler step took 12 seconds

But in the comfyui.log you can see:
[2026-01-05 16:45:06.914] Requested to load WAN21
[2026-01-05 16:45:26.361] loaded completely; 20567.30 MB usable, 13631.43 MB loaded, full load: True
[2026-01-05 16:45:38.633]
100%|████████████████████████████████████████████████████████████████████████████████████| 2/2 [00:12<00:00, 7.09s/it]
100%|████████████████████████████████████████████████████████████████████████████████████| 2/2 [00:12<00:00, 6.09s/it]
[2026-01-05 16:45:50.505] Found quantization metadata version 1
=> 16:45:26.3 - 16:45:50.5 = 24.3 seconds

The GUI has the same issue:
the progress bar in the KSampler node immediately jumps to 50%.
Then it does 12 seconds of calculation, and then jumps to 100%
But then it does another 12 seconds of calculation.

I also tried a Z-Image Turbo workflow using the BF16 models, but there I do not have this issue.
So for me, I only have this behaviour when using the FP8 models in a Wan 2.2 flow.

my hardware setup:
CPU : AMD Ryzen 9 7950X (16c/32t)
GPU : AMD Radeon RX 7900 XTX (24 GB VRAM)
System Memory : 64 GB
OS : Windows 11
installed driver : https://www.amd.com/en/resources/support-articles/release-notes/RN-AMDGPU-WINDOWS-PYTORCH-7-1-1.html

Reproduce the Wan 2.2 I2V workflow with run_amd_gpu.bat and the listed FP8 models, then compare KSampler progress output with the timestamps in comfyui.log. Trace how the KSampler steps report progress and timing. Done means both high- and low-noise steps show accurate progress and duration in the console and GUI.

Written by the indexing model from the issue text.