Sayak Paul sayakpaul

When loading the LoRA params (that were obtained on a quantized base model) and merging them into the base model, it is recommended to first dequantize the base model, merge the LoRA params into it, and then quantize the model again. This is because merging into 4bit quantized models can lead to some rounding errors. Below, we provide an end-to-end example:

First, load the original model and merge the LoRA params into it:

from diffusers import FluxPipeline 
import torch 

ckpt_id = "black-forest-labs/FLUX.1-dev"
pipeline = FluxPipeline.from_pretrained(

Detailed writeup: https://huggingface2.notion.site/How-to-split-Flux-transformer-and-run-inference-aa1583ad23ce47a78589a79bb9309ab0

But TLDR is we split the models where possible and decouple the different stages of pipeline

	from google import genai
	from google.genai import types
	import typing_extensions as typing
	from PIL import Image
	import requests
	import io
	import json
	import os

	"""
	Implementation of the label generation part in https://danielvanstrien.xyz/posts/2025/deepseek/distil-deepseek-modernbert.html
	using `transformers` and DeepSeek.
	"""

	from transformers import AutoModelForCausalLM, AutoTokenizer
	import torch
	import re
	import contextlib
	import math

	from moviepy.editor import VideoFileClip, clips_array
	import glob

	def create_video_collage(video_paths, output_path="collage.mp4"):
	"""
	Combine four videos of the same resolution into a 2×2 collage.

	Args:
	video_paths (list[str]): List of paths to the four video files.
	output_path (str): Filename for the output collage video.

	from diffusers import DiffusionPipeline
	from diffusers import FluxTransformer2DModel, BitsAndBytesConfig
	from transformers import T5EncoderModel, BitsAndBytesConfig as BnbConfig
	from offloader import ModelOffloaderV2
	import torch.utils.benchmark as benchmark
	from pathlib import Path
	import os
	import sys
	import torch
	import json

	import torch
	from diffusers.utils import export_to_video
	from diffusers import LTXPipeline, LTXVideoTransformer3DModel, GGUFQuantizationConfig

	ckpt_path = (
	"https://huggingface.co/city96/LTX-Video-gguf/blob/main/ltx-video-2b-v0.9-Q3_K_S.gguf"
	)
	transformer = LTXVideoTransformer3DModel.from_single_file(
	ckpt_path,
	quantization_config=GGUFQuantizationConfig(compute_dtype=torch.bfloat16),

	"""
	Usage:

	python low_rank_lora.py --repo_id=glif/how2draw --filename="How2Draw-V2_000002800.safetensors" \
	--new_rank=4 --new_lora_path="How2Draw-V2_000002800_rank_4.safetensors"
	"""

	import torch
	from huggingface_hub import hf_hub_download
	import safetensors.torch

	# Copyright 2024 Black Forest Labs and The HuggingFace Team. All rights reserved.
	#
	# Licensed under the Apache License, Version 2.0 (the "License");
	# you may not use this file except in compliance with the License.
	# You may obtain a copy of the License at
	#
	# http://www.apache.org/licenses/LICENSE-2.0
	#
	# Unless required by applicable law or agreed to in writing, software
	# distributed under the License is distributed on an "AS IS" BASIS,