sd-turbo on --cpu mode produces gibberish when trying with 2 or 4 steps

### Expected Behavior

i expect it to generate a working image, like this without --cpu (on an rtx 3060)


### Actual Behavior

it generates gibberish on any step count that isnt 1


NOTE: it works normally on gpu mode (without --cpu)

### Steps to Reproduce

download comfyui

download vae,clip,unet or whole checkpoint from https://huggingface.co/stabilityai/sd-turbo

python main.py --cpu

use this workflow:

[sd2.1-turbo.json](https://github.com/user-attachments/files/19026984/sd2.1-turbo.json)

### Debug Logs

```powershell
python main.py --cpu --disable-all-custom-nodes
Checkpoint files will always be loaded safely.
Total VRAM 64139 MB, total RAM 64139 MB
pytorch version: 2.6.0+cu124
xformers version: 0.0.29.post3
Set vram state to: DISABLED
Device: cpu
Using sub quadratic optimization for attention, if you have memory or speed issues try using: --use-split-cross-attention
ComfyUI version: 0.3.18
[Prompt Server] web root: /home/user/AI/ComfyUI/web
Skipping loading of custom nodes
Starting server

To see the GUI go to: http://127.0.0.1:8188
got prompt
Using split attention in VAE
Using split attention in VAE
VAE load device: cpu, offload device: cpu, dtype: torch.float32
input_blocks.1.0.skip_connection.weight down_blocks.0.resnets.0.conv_shortcut.weight
input_blocks.1.0.skip_connection.bias down_blocks.0.resnets.0.conv_shortcut.bias
input_blocks.2.0.skip_connection.weight down_blocks.0.resnets.1.conv_shortcut.weight
input_blocks.2.0.skip_connection.bias down_blocks.0.resnets.1.conv_shortcut.bias
input_blocks.5.0.skip_connection.weight down_blocks.1.resnets.1.conv_shortcut.weight
input_blocks.5.0.skip_connection.bias down_blocks.1.resnets.1.conv_shortcut.bias
input_blocks.8.0.skip_connection.weight down_blocks.2.resnets.1.conv_shortcut.weight
input_blocks.8.0.skip_connection.bias down_blocks.2.resnets.1.conv_shortcut.bias
input_blocks.10.0.skip_connection.weight down_blocks.3.resnets.0.conv_shortcut.weight
input_blocks.10.0.skip_connection.bias down_blocks.3.resnets.0.conv_shortcut.bias
input_blocks.11.0.skip_connection.weight down_blocks.3.resnets.1.conv_shortcut.weight
input_blocks.11.0.skip_connection.bias down_blocks.3.resnets.1.conv_shortcut.bias
input_blocks.12.0.op.weight down_blocks.3.downsamplers.0.conv.weight
input_blocks.12.0.op.bias down_blocks.3.downsamplers.0.conv.bias
middle_block.0.skip_connection.weight mid_block.resnets.0.conv_shortcut.weight
middle_block.0.skip_connection.bias mid_block.resnets.0.conv_shortcut.bias
middle_block.2.skip_connection.weight mid_block.resnets.1.conv_shortcut.weight
middle_block.2.skip_connection.bias mid_block.resnets.1.conv_shortcut.bias
output_blocks.11.2.conv.weight up_blocks.3.upsamplers.0.conv.weight
output_blocks.11.2.conv.bias up_blocks.3.upsamplers.0.conv.bias
label_emb.0.0.weight class_embedding.linear_1.weight
label_emb.0.2.weight add_embedding.linear_2.weight
label_emb.0.0.bias add_embedding.linear_1.bias
label_emb.0.2.weight class_embedding.linear_2.weight
label_emb.0.2.bias add_embedding.linear_2.bias
label_emb.0.2.bias class_embedding.linear_2.bias
label_emb.0.0.weight add_embedding.linear_1.weight
label_emb.0.0.bias class_embedding.linear_1.bias
model weight dtype torch.float32, manual cast: None
model_type EPS
Requested to load SD2ClipModel
loaded completely 9.5367431640625e+25 675.263671875 True
CLIP/text encoder model load device: cpu, offload device: cpu, current: cpu, dtype: torch.float16
clip missing: ['text_model.encoder.layers.23.layer_norm1.weight', 'text_model.encoder.layers.23.layer_norm1.bias', 'text_model.encoder.layers.23.self_attn.q_proj.weight', 'text_model.encoder.layers.23.self_attn.q_proj.bias', 'text_model.encoder.layers.23.self_attn.k_proj.weight', 'text_model.encoder.layers.23.self_attn.k_proj.bias', 'text_model.encoder.layers.23.self_attn.v_proj.weight', 'text_model.encoder.layers.23.self_attn.v_proj.bias', 'text_model.encoder.layers.23.self_attn.out_proj.weight', 'text_model.encoder.layers.23.self_attn.out_proj.bias', 'text_model.encoder.layers.23.layer_norm2.weight', 'text_model.encoder.layers.23.layer_norm2.bias', 'text_model.encoder.layers.23.mlp.fc1.weight', 'text_model.encoder.layers.23.mlp.fc1.bias', 'text_model.encoder.layers.23.mlp.fc2.weight', 'text_model.encoder.layers.23.mlp.fc2.bias', 'text_projection.weight']
Requested to load BaseModel
loaded completely 9.5367431640625e+25 3303.187271118164 True
100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 4/4 [00:07<00:00, 1.96s/it]
Requested to load AutoencoderKL
loaded completely 9.5367431640625e+25 319.11416244506836 True
Prompt executed in 14.35 seconds
```

### Other

torch: 2.6.0