FindHao

The complete minimal reproducer follows. Both source files are identical to the files compiled and run in this experiment, with no omissions. Save them in the same directory, then use the build and run commands below. The shared-memory size and compilation target match the tested B200 configuration.

Save the following CUDA kernel as shared_memory_kernel.cu:

extern "C" __global__ void shared_memory_probe(unsigned int *output,
                                                unsigned int count) {
  extern __shared__ unsigned int scratch[];
  for (unsigned int i = threadIdx.x; i < count; i += blockDim.x)

FindHao

{
    "version": "2.0.0",
    "tasks": [
        {
            "label": "build",
            "type": "shell",
            "command": "make -j 4",
            "group": {
                "kind": "build",
                "isDefault": true

FindHao

from ctypes import c_void_p, c_long
import torch
import math
import random
import os
import tempfile
from math import inf, nan
from torch._inductor.hooks import run_intermediate_hooks
from torch._inductor.utils import maybe_profile
from torch import empty_strided, as_strided, device

FindHao

#0  c10::get_backtrace[abi:cxx11](unsigned long, unsigned long, bool) (frames_to_skip=<optimized out>, frames_to_skip@entry=1, maximum_number_of_frames=maximum_number_of_frames@entry=64, 
    skip_python_frames=skip_python_frames@entry=true) at /scratch/findhao/pytorch/c10/util/Backtrace.cpp:285
#1  0x00007fff836f0821 in c10::(anonymous namespace)::<lambda()>::operator() (__closure=<optimized out>) at /scratch/findhao/pytorch/c10/util/Logging.cpp:28
#2  std::_Function_handler<std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> >(), c10::(anonymous namespace)::GetFetchStackTrace()::<lambda()> >::_M_invoke(const std::_Any_data &) (
    __functor=...) at /usr/include/c++/9/bits/std_function.h:286
#3  0x00007fff836f1d53 in std::function<std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> > ()>::operator()() const (this=<optimized out>) at /usr/include/c++/9/bits/std_function.h:683
#4  c10::Error::Error (this=0x558e620, source_location=..., msg=...) at /scratch/f

FindHao

import time
from torch import profiler
import torch
import argparse

out_channels = 64
input_shape = [1, 3, 224, 224]
weight_shape = [64, 3, 7, 7]
stride = (2, 2)

FindHao

#!/bin/bash
# Put this script to the same folder with Icalingua++-XX.AppImage or where you want to download the appimage.

# get script path
file_path=$(dirname $(readlink -f $0))
cd $file_path
# check github releases for new version
url=$(curl -s https://api.github.com/repos/Icalingua-plus-plus/Icalingua-plus-plus/releases/latest | grep "browser_download_url.*AppImage" | cut -d : -f 2,3 | tr -d \" | sed -n '3p')

# check if url is empty

FindHao

# https://github.com/phlippe/uvadlc_notebooks_benchmarking/blob/main/PyTorch/Tutorial5_Inception_ResNet_DenseNet.py
from torchvision.datasets import CIFAR10
from torchvision import transforms
from torch.utils.tensorboard import SummaryWriter
import torch.utils.data as data
import torch
from flax.training import train_state, checkpoints
from flax import linen as nn
from jax import random
import jax.numpy as jnp

FindHao

#!/usr/bin/env bash
# This script is based on Copyright (C) 2013 - 2020 Teddysun <i@teddysun.com>
# Modified by FindHao <find@findhao.net>

# https://findhao.net/easycoding/2605.html

# You must modify the config before run it!!! 
 

# [[ $EUID -ne 0 ]] && echo "Error: This script must be run as root!" && exit 1

FindHao

import torch
from torch import profiler
import numpy as np


def _len_and_dim_norm(vectors):
    """
    length and attention head size dim normalization
    """
    vectors = vectors * torch.rsqrt(

FindHao

from torch import profiler
import torch
import torch.nn.functional as F
import argparse

def run_conv2d(input_shape, weight_shape, other_args, profile_folder):
    input = torch.ones(input_shape, dtype=torch.float32, device='cuda')
    weight = torch.ones(weight_shape, dtype=torch.float32, device='cuda')
    bias = other_args[0]
    stride = other_args[1]