takuma104

#import "YAJL.h" // YAJL ObjC
#import "NSString+SBJSON.h" // JSON Framework

// JSON Data: Twitter Home Timeline (count=200)
// iPod touch 3rd Gen -> test0:296ms test1:323ms test2:170ms test3:488ms

@implementation Test

- (void)test0:(NSData*)JSONData {
	double st = CFAbsoluteTimeGetCurrent();

takuma104

static void indentedMsg(int indent, NSString *str) {
	NSMutableString *idt = [NSMutableString string];
	for (int i = 0; i < indent; i++) {
		[idt appendString:@" "]; 
	}
	NSLog(@"%@%@", idt, str);
}

- (void)parserDidStartDictionary:(YAJLParser *)parser {
	indentedMsg(indent, @"dict>");

takuma104

/*
 PerfCheck: Stopwatch for Debugging (ObjC)

 Usage:
    PFBEGIN(@"hoge");
    // some process
    PFEND(@"hoge");

    PFBEGIN(@"fuga");
    // some process

takuma104

#import "flalib.h"

@interface TestSprite : FLSprite {

}

@end

@implementation TestSprite

takuma104

# SELF-ATTENTION DOES NOT NEED O(n2) MEMORY： https://arxiv.org/pdf/2112.05682.pdf
# https://github.com/google-research/google-research/blob/master/memory_efficient_attention/memory_efficient_attention.ipynb

# you may need to apply 'export XLA_PYTHON_CLIENT_PREALLOCATE=false' 
# https://github.com/google/jax/issues/7118#issuecomment-950183972

import functools
import jax
import jax.numpy as jnp
import math

takuma104

import xformers
import xformers.ops
import torch

device = 'cuda'
dtype = torch.float16
shape = (1, 1024, 16, 16)

torch.manual_seed(0)

takuma104

import torch
from diffusers import StableDiffusionPipeline, DPMSolverMultistepScheduler
import matplotlib.pyplot as plt
from PIL import Image
import numpy as np

plt.rcParams["figure.figsize"] = (10,5)
plt.rcParams['figure.facecolor'] = 'white'

def generate_tuxedo_cat_picture(fn_prefix, seed=0):

takuma104

import torch
from diffusers import StableDiffusionPipeline, DPMSolverMultistepScheduler, EulerAncestralDiscreteScheduler
import matplotlib.pyplot as plt
from PIL import Image
import numpy as np
import xformers
import xformers.ops

plt.rcParams["figure.figsize"] = (10,5)
plt.rcParams['figure.facecolor'] = 'white'

takuma104

Reproducible generation with xFormers for SD2.x or variant models

More details and background: Generating (almost) reproducible pictures using Diffusers with xFormers

Limitation

Unfortunately, Flash Attention won't accept SD1.x attention shapes, but will accept SD2.x or variants.

Test Result

takuma104

# from https://github.com/lllyasviel/ControlNet/blob/main/gradio_canny2image.py

from share import *
import config

import cv2
import einops
import gradio as gr
import numpy as np
import torch