import torch
from diffusers import DiffusionPipeline
# switch to "mps" for apple devices
pipe = DiffusionPipeline.from_pretrained("codemichaeld/Wan1.3b_fp8", dtype=torch.bfloat16, device_map="cuda")
prompt = "Astronaut in a jungle, cold color palette, muted colors, detailed, 8k"
image = pipe(prompt).images[0]FP8 Model with Delta Compensation
- Source:
https://huggingface.co/Kijai/WanVideo_comfy - File:
Wan2_1-T2V-1_3B_fp32.safetensors - FP8 Format:
E5M2 - Delta File:
Wan2_1-T2V-1_3B_fp32-fp8-delta.safetensors
Usage (Inference)
To restore near-original precision:
import torch
from safetensors.torch import load_file
fp8_state = load_file("Wan2_1-T2V-1_3B_fp32-fp8-e5m2.safetensors")
delta_state = load_file("Wan2_1-T2V-1_3B_fp32-fp8-delta.safetensors")
restored_state = {}
for key in fp8_state:
if f"delta.{key}" in delta_state:
fp8_weight = fp8_state[key].to(torch.float32)
delta = delta_state[f"delta.{key}"]
restored_state[key] = fp8_weight + delta
else:
restored_state[key] = fp8_state[key].to(torch.float32)
Requires PyTorch ≥ 2.1 for FP8 support.
- Downloads last month
- 10
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support