msaroufim

import torch

# CUDA kernel with inline PTX
kernel_source = """
__global__ void vector_add(const float* a, const float* b, float* c, int n) {
    int idx;
    asm("mov.u32 %0, %%ctaid.x;" : "=r"(idx));
    int tid;
    asm("mov.u32 %0, %%tid.x;" : "=r"(tid));
    int ntid;

ezyang

import torch
from torch import nn
from torch.distributed.tensor.placement_types import Replicate, Shard
from torch.testing._internal.distributed.fake_pg import FakeStore

import torch.distributed as dist
from torch.distributed.device_mesh import init_device_mesh
from torch.distributed.tensor import DTensor, Replicate

world_size = 4

Chillee

import torch
torch.set_default_device('cuda')
from triton.testing import do_bench
from collections import defaultdict
from functools import partial
import random
random.seed(0)

def get_flops(A, B):
    ms = do_bench(lambda: torch.mm(A, B))

shunting314

"""
Augment kernel metadata generated by kernel_metadata metric table in inductor.
For each row in input, use NCU to profile the kernel. The corresponding output row
contains more metadata gathered by NCU.

It can be super slow to run NCU. e.g. for the 10K kernels gathered from Huggingface,
it took almost a whole day to run NCU for each unique kernels. The script thus cache
the ncu output in the file system. If the ncu output is cached, we don't run NCU again.

Example input:  https://gist.github.com/shunting314/22995da0da8b66d4cf989cb7f0508399

fishmingyu

-------------------------------------------------------  ------------  ------------  ------------  ------------  ------------  ------------  
                                                   Name    Self CPU %      Self CPU   CPU total %     CPU total  CPU time avg    # of Calls  
-------------------------------------------------------  ------------  ------------  ------------  ------------  ------------  ------------  
graph_0_cpp_fused_add_exp_index_select_mul_scatter_a...        29.02%      11.966ms        29.02%      11.966ms      11.966ms             1  
graph_0_cpp_fused_add_clone_exp_index_select_mul_rel...        28.60%      11.794ms        28.60%      11.794ms      11.794ms             1  
graph_0_cpp_fused_add_clone_exp_index_select_mul_new...        27.49%      11.335ms        27.49%      11.335ms      11.335ms             1  
                                         aten::scatter_         5.92%       2.442ms         5.92%       2.442ms     814.000us             3  
                              

fishmingyu

import torch

import torch_geometric
from torch_geometric.profile import benchmark
from torch_geometric.testing import (
    disableExtensions,
    onlyFullTest,
    onlyLinux,
    withCUDA,
    withPackage,

Chillee

import torch
import torch._inductor.config
import time

torch._inductor.config.triton.cudagraphs = False
torch.set_float32_matmul_precision('high')

def bench(f, name=None, iters=100, warmup=5, display=True, profile=False):
    for _ in range(warmup):
        f()

ivan-loh

#!/bin/bash

sudo apt install build-essential texinfo libx11-dev libxpm-dev libjpeg-dev libpng-dev libgif-dev libtiff-dev libgtk2.0-dev libncurses-dev libgnutls28-dev

wget http://ftp.gnu.org/gnu/emacs/emacs-29.3.tar.gz
tar xvzf emacs-29.3.tar.gz

cd emacs-29.3
./configure --with-mailutils
make

Borda

"""
Parse all users from release notes according to PR reference.

Resources:
 - https://developer.github.com/v3/pulls/
 - http://zetcode.com/python/requests/
 - https://developer.github.com/v3/#authentication
"""

import json

kinu

Google に入ったときの話 (Kinuko)

祭っぽいので私も書いてみることにした！お手軽に gist で。

コンテキスト：https://togetter.com/li/1331865

対策とか（特になし）

と書き出したはいいが、私が受けたときは本も情報もあまりなく、かつ競プロ的なものの存在も知らなかったので、とりあえず家にあったアルゴリズムの本を２回くらい読み直した。そして受かった。最初っから情報があまりない方のパターンで申し訳ない 😄