sayakpaul

"""
NVFP4 Quantization Benchmark for Flux.1-Dev on Modal B200 GPU

This script benchmarks NVFP4 (4-bit floating point) quantization performance
for the Flux.1-Dev image generation model on NVIDIA B200 GPUs.

B200 GPUs have native FP4 Tensor Core support, making them ideal for
this workload.

Usage:

sayakpaul

import torch 
from diffusers import DiffusionPipeline 
import spaces
from spaces.zero.torch.aoti import ZeroGPUCompiledModel, ZeroGPUWeights
from time import perf_counter
import argparse 


CKPT_ID = "black-forest-labs/Flux.1-Dev"

sayakpaul

FA3 attention processor comes from https://gist.github.com/sayakpaul/ff715f979793d4d44beb68e5e08ee067.

Results from an H100:

latency=36.606 seconds. (AoT regional compilation)
latency=36.555 seconds. (JiT regional compilation)

sayakpaul

# Make sure you are using the latest `bitsandbytes` (at least 0.46.0) and PyTorch nightlies (at least 2.8).
# Put together by sayakpaul and anijain2305

from diffusers.quantizers import PipelineQuantizationConfig
from diffusers import FluxPipeline
import argparse
import json
import torch
import time
from functools import partial

sayakpaul

from diffusers import DiffusionPipeline
import torch.utils.benchmark as benchmark
import torch
import psutil
import os
import json
import argparse

def benchmark_fn(f, *args, **kwargs):
    t0 = benchmark.Timer(

sayakpaul

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

sayakpaul

"""
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

sayakpaul

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.

sayakpaul

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

sayakpaul

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),