malfet

#include <stdio.h>
#include <chrono>

__global__ void noop() { }

int main(int argc, const char *argv[]) {
   cudaDeviceProp prop;
  auto rc = cudaGetDeviceProperties(&prop, 0);
  printf("Running on %s sm%d.%d  multiProcessorCount = %d maxBlocksPerMultiProcessor  = %d maxThreadsPerBlock  = %d\n",
         prop.name, prop.major, prop.minor, prop.multiProcessorCount, prop.maxBlocksPerMultiProcessor, prop.maxThreadsPerBlock);

malfet

a_cpp = """#include <iostream>

namespace foo::bar {
inline namespace baz {
int inc(int x) {
  std::cout << "do inc from lib_a" << std::endl;
  return x + 1;
}
} // inline namespace baz
void do_a(int x) {

malfet

import torch
import sys
import timeit

def add_repeat(x, y, repeat=10):
    rc = x.clone()
    for i in range(repeat):
       rc += y
    return rc

malfet

#include <stdio.h>

__global__ void print() {
    printf("Hello World of CUDA threadIdx.x=%d\n", threadIdx.x);
}

__global__ void noop() { }

int main(int argc, const char *argv[]) {
   cudaDeviceProp prop;

malfet

import torch
import os
os.environ["MTL_CAPTURE_ENABLED"]="1"
a = torch.ones(2, (1 << 31) + 5, dtype=torch.int8, device='mps')
index_0 = torch.tensor([0, -1, 0, 1], device=a.device)
index_1 = torch.tensor([-2, -1, 0, 1], device=a.device)
values = torch.tensor([12, 13, 10, 11], dtype=a.dtype, device=a.device)
with torch.mps.profiler.metal_capture("index_put"):
    a.index_put_((index_0, index_1), values, accumulate=True)
    b = a[1, -2].cpu()

malfet

#!/usr/bin/env python3

import sys
import subprocess
import urllib.request
import json

def get_latest_version(package_name: str) -> str:
    """Get latest version from PyPI"""
    api_url = f"https://pypi.org/pypi/{package_name}/json"

malfet

With cudnn-9.10.2.21
```
$ CUDNN_LOGINFO_DBG=3 RUN_SLOW=1 python3 -m pytest -v tests/models/vit/test_modeling_vit.py::ViTModelTest::test_batching_equivalence
========================================================================================== test session starts ===========================================================================================
platform linux -- Python 3.10.12, pytest-8.4.1, pluggy-1.6.0 -- /home/ubuntu/py3.10-nightly/bin/python3
cachedir: .pytest_cache
rootdir: /home/ubuntu/transformers
configfile: pyproject.toml
plugins: xdist-3.8.0, asyncio-1.1.0, rerunfailures-15.1, order-1.3.0, timeout-2.4.0, rich-0.2.0
asyncio: mode=strict, asyncio_default_fixture_loop_scope=function, asyncio_default_test_loop_scope=function

malfet

import torch
import torch.utils.cpp_extension as _ce
import tempfile
import os
import subprocess

src = """#include <c10/util/BFloat16.h>
#include <iostream>
int main() {
   std::cout << c10::BFloat16(3.14) << std::endl;

malfet

import ctypes
import torch
import time

def nvrtc_compile(source: str) -> str:
    from ctypes import CDLL, c_void_p, c_char_p, c_size_t, byref, create_string_buffer
    libnvrtc = CDLL('libnvrtc.so')
    def get_error_string() -> str:
        err_p = c_char_p()
        libnvrtc.nvrtcGetErrorString(result, byref(err_str))

malfet

"""
Example showing how to use the no_header mode with a TensorBase CUDA extension

This example creates a CUDA extension that directly includes ATen/core/TensorBase.h 
instead of torch/extension.h, resulting in faster compilation with no_header=True
"""
from datetime import datetime
import torch
import torch.utils.cpp_extension