Quantized weights we publish
Collection
FP8, INT8, NF4 and precision-cast variants of open models. Each card says exactly which tensors changed and which were left alone. • 16 items • Updated
import torch
from diffusers import DiffusionPipeline
# switch to "mps" for apple devices
pipe = DiffusionPipeline.from_pretrained("wavespeed/Wan2.1-T2V-14B-Diffusers-fp16", dtype=torch.bfloat16, device_map="cuda")
prompt = "Astronaut in a jungle, cold color palette, muted colors, detailed, 8k"
image = pipe(prompt).images[0]Wan2.1-T2V-14B-Diffusers-fp16
An fp16 cast of Wan2.1-T2V-14B-Diffusers.
What was changed
Upstream ships the transformer in fp32. Here every component is stored in fp16, which halves the download and the resident weight footprint — the transformer goes from ~57.2 GB to ~28.6 GB — with no change to tensor names or shapes. Nothing is quantized; this is a straight dtype cast.
fp16 rather than bf16 is deliberate: it is the dtype our fp8/nf4 variants are built on top of, so the whole family shares one baseline.
Usage
import torch
from diffusers import WanPipeline
pipe = WanPipeline.from_pretrained(
"wavespeed/Wan2.1-T2V-14B-Diffusers-fp16", torch_dtype=torch.float16
).to("cuda")
frames = pipe("a paper boat drifting down a gutter after rain", num_frames=81).frames[0]
Related
wavespeed/Wan2.1-T2V-14B-Diffusers-fp16-nf4— the same baseline with NF4-quantized transformer blocks.
License
Apache-2.0, inherited from Wan2.1.
- Downloads last month
- 33
Model tree for wavespeed/Wan2.1-T2V-14B-Diffusers-fp16
Base model
Wan-AI/Wan2.1-T2V-14B-Diffusers