About this repository

FP8 (e4m3) quantized Ovi fusion model, published by WaveSpeed AI. Ovi generates video and its matching audio in one pass; both towers are quantized here — 30 video blocks and 30 audio blocks — with e4m3_e4m3_dynamic_per_tensor on attention and FFN layers and e4m3_weightonly elsewhere. Compute stays in bf16. FP8 matmul needs Hopper (H100/H200) or newer to be faster than bf16.

The original model is Ovi by Character.AI, released under Apache-2.0; that license carries over to these weights.

Ovi FusionModel - FP8 Quantized

This is the Ovi FusionModel quantized with FP8 (e4m3_e4m3_dynamic_per_tensor) for faster inference.

Quantization Details

Video Model Blocks: 30 blocks quantized
Audio Model Blocks: 30 blocks quantized
Attention/FFN layers: e4m3_e4m3_dynamic_per_tensor
Other layers: e4m3_weightonly

Usage

import sys
import os
import torch
from omegaconf import OmegaConf
from huggingface_hub import hf_hub_download

OVI_PATH = "./workspace/Ovi"
sys.path.insert(0, OVI_PATH)
os.chdir(OVI_PATH)

from ovi.ovi_fusion_engine import OviFusionEngine

# Download quantized weights
model_path = hf_hub_download(
    repo_id="wavespeed/Ovi-e4m3_e4m3_dynamic_per_tensor",
    filename="model.pth"
)

config = OmegaConf.load("config.yaml")
engine = OviFusionEngine(config=config, device="cuda", target_dtype=torch.bfloat16)

# Load quantized weights
engine.model.load_state_dict(torch.load(model_path))

# Model is already quantized, ready for inference