apivovarov

a = 5
b = 6
y = a/b

h = 0.00001

def dy_da_f():
  a2 = a + h
  y2 = a2 / b
  dy_da = (y2 - y) / h

apivovarov

import jax
from jax import Array
import jax.numpy as jnp

def init_params(key: Array, shape) -> Array:
    return jax.random.normal(key, shape).astype(jax.dtypes.bfloat16)

def softmax(x):
    mx = x.max(axis=-1, keepdims=True)
    mx = jax.lax.stop_gradient(mx)

apivovarov

HloModule xla_computation_ff, entry_computation_layout={(f32[1,224,224,3]{3,2,1,0})->(f32[1,224,224,3]{3,2,1,0})}

ENTRY main.20 {
  Arg_0.1 = f32[1,224,224,3]{3,2,1,0} parameter(0)
  multiply.10 = f32[1,224,224,3]{3,2,1,0} multiply(Arg_0.1, Arg_0.1)
  multiply.11 = f32[1,224,224,3]{3,2,1,0} multiply(Arg_0.1, multiply.10)
  constant.8 = f32[] constant(0.044715)
  broadcast.9 = f32[1,224,224,3]{3,2,1,0} broadcast(constant.8), dimensions={}
  multiply.12 = f32[1,224,224,3]{3,2,1,0} multiply(multiply.11, broadcast.9)
  add.13 = f32[1,224,224,3]{3,2,1,0} add(Arg_0.1, multiply.12)

apivovarov

HloModule xla_computation_ff, entry_computation_layout={(f32[4,1000]{1,0})->(f32[4,1000]{1,0})}

region_0.4 {
  Arg_0.5 = f32[] parameter(0)
  Arg_1.6 = f32[] parameter(1)
  ROOT maximum.7 = f32[] maximum(Arg_0.5, Arg_1.6)
}

region_1.15 {
  Arg_0.16 = f32[] parameter(0)

apivovarov

#include <iostream>
#include <vector>

std::vector<int> testNRVO(int value, size_t size, const std::vector<int> **localVec)
{
   std::vector<int> vec(size, value);

   *localVec = &vec;

   /* Do something here.. */

apivovarov

import torch
from transformers import RobertaTokenizer, RobertaModel

torch.set_grad_enabled(False)

class RobertaTraceWrapper(torch.nn.Module):
    def __init__(self, model):
        super().__init__()
        self.model = model
    def forward(self, x):

apivovarov

// To compile - nvcc cuda_check.cu -o cuda_check -lcuda
// To run ./cuda_check
// set g++ path to older g++ if needed - export NVCC_PREPEND_FLAGS='-ccbin
// /usr/local/gcc-11/bin/g++-11'
#include <cuda.h>
#include <cuda_runtime_api.h>
#include <stdio.h>

/* Outputs some information on CUDA-enabled devices on your computer,
 * including compute capability and current memory usage.

apivovarov

import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.optim as optim
import torchvision
from torchinfo import summary

in_sz = 28*28

n_epochs = 1

apivovarov

VV=8.5.3-1+cuda11.8

sudo apt install \
libcudnn8=8.9.0.131-1+cuda11.8 \
libcudnn8-dev=8.9.0.131-1+cuda11.8 \
libnvinfer-bin=$VV \
libnvinfer-dev=$VV \
libnvinfer-plugin-dev=$VV \
libnvinfer-plugin8=$VV \
libnvinfer8=$VV \

apivovarov

import math

import numpy as np

import aitemplate
from aitemplate.frontend import nn, Tensor
from aitemplate.compiler.model import AITData
from aitemplate.compiler.base import IntVar, IntImm

class M1AIT(nn.Module):