rootonchair/MiniMax-H3-nunchaku-lite-int4

MiniMax-H3 t2va — SVDQuant int4 (nunchaku-lite)

SVDQuant W4A4 quantization of the transformer/ partition (t2va / fl2va workflows) of MiniMaxAI/MiniMax-H3: 312 SVDQ targets (int4, group 64, rank-32 low-rank branch, smoothing search) plus 50 AdaLN modulation projections as AWQ W4A16. 18.5GB vs ~66GB bf16. Embedded nunchaku_lite.runtime_manifest v1. Calibration: 8 prompts x 20 steps, default canvas.

Load

import torch
from diffusers import ModularPipeline
from diffusers.models.transformers.transformer_minimax_h3 import (
    MiniMaxH3RotaryPosEmbed, MiniMaxH3Transformer3DModel)
from nunchaku_lite import patch_transformer
from huggingface_hub import hf_hub_download

ckpt = hf_hub_download("rootonchair/MiniMax-H3-nunchaku-lite-int4",
                       "svdq-int4_r32-minimax-h3-t2va.safetensors")
config = MiniMaxH3Transformer3DModel.load_config("MiniMaxAI/MiniMax-H3", subfolder="transformer")
with torch.device("meta"):
    transformer = MiniMaxH3Transformer3DModel.from_config(config)
transformer.rope.inv_freq = MiniMaxH3RotaryPosEmbed(
    rope_freq_dim=transformer.config.rope_freq_dim,
    rope_theta=transformer.config.rope_theta).inv_freq
# patch with assign for meta-built modules (nunchaku_lite.core._patch_component(assign=True))

pipe = ModularPipeline.from_pretrained("MiniMaxAI/MiniMax-H3", workflow="t2va")

Produced with diffuse-compressor (examples/text_to_video/quantize_minimax_h3.py).