FP8 WAN safetensors crash with OOM on RTX 5090

### Hardware

RTX 5090
64 GB system ram
Latest NVIDIA drivers installed (fully up to date)

### Environment

Fresh install of ComfyUI_windows_portable_nvidia_cu128.7z [v0.7.0]
Also tested ComfyUI_windows_portable_nvidia.7z (non‑cu128) [v0.7.0]
No manual Python environment
No custom patches
Clean, default portable installs

### Summary of the issue
Using the default Wan 2.2 I2V workflow FP8 WAN safetensors models consistently crash with CUDA OOM during sampling, even though VRAM usage is far below the 32 GB available. This happens on both portable builds (cu128 and non‑cu128). The crash usually occurs after the Wan 2.2 high ksampler finishes and passes it to the low ksampler. This default I2V workflow use to work previously a few months ago.

.gguf versions of WAN work fine, so the issue appears specific to FP8 safetensors.

### Models used
wan2.2_i2v_high_noise_14B_fp8_scaled.safetensors
wan2.2_i2v_low_noise_14B_fp8_scaled.safetensors
umt5_xxl_fp8_e4m3fn_scaled.safetensors
wan_2.1_vae.safetensors

### Behavior
Models load successfully (UNet + text encoder + VAE)
VRAM usage stays well below 32 GB
During sampling, ComfyUI attempts to unload/reload models
Then crashes with a fake OOM error
Browser sometimes crashes
Console shows repeated FP8‑related warnings

### Errors
unet unexpected: ['scaled_fp8']
CUDA error: out of memory
This happens identically on both portable builds.

### What I’ve already tested
fresh portable cu128 install → crashes
fresh portable non‑cu128 install → crashes
latest NVIDIA drivers → no change
weight_dtype=fp8 → crashes even with FP16 models
GGUF versions of WAN → work perfectly
same workflow months ago → worked perfectly

Additionally using the weight_dtype=default DOES work when using the full wan2.2_i2v_high_noise_14B_fp16.safetensors and wan2.2_i2v_low_noise_14B_fp16.safetensors but is extremely slow since using the full fp16 model.

### Expected behavior
FP8 WAN safetensors should load and sample without crashing, as they did in earlier ComfyUI versions and as the default workflow suggests.

### Actual behavior
FP8 WAN safetensors crash during sampling with a misleading OOM error, even though VRAM is not exhausted.

### Request
Can someone else with a 5090 (or really any Nvidia card that can run the standard fp8 .safetensors) test this? You just need to install the current Nvidia portable version [v0.7.0] and try running the default Wan 2.2 I2V workflow.