gradjitta

Flux Dev (dynamic shapes) Benchmark

Setup

import os
import torch
from einops import rearrange
from transformers import pipeline
from flux.sampling import denoise, get_noise, get_schedule, unpack, prepare

Chillee

import torch
from torch.nn.attention._flex_attention import _create_block_mask, _create_mask
from functools import partial
from torch.nn.attention._flex_attention import _flex_attention
from triton.testing import do_bench
import torch.nn.functional as F
from functools import lru_cache
torch.set_default_device('cuda')

# Example usage

matijagrcic

Ref: Exclusive Q&A: John Carmack’s ‘Different Path’ to Artificial General Intelligence

"So I asked Ilya Sutskever, OpenAI’s chief scientist, for a reading list. He gave me a list of like 40 research papers and said, ‘If you really learn all of these, you’ll know 90% of what matters today.’ And I did. I plowed through all those things and it all started sorting out in my head."

Ref: https://x.com/ID_AA_Carmack/status/1622673143469858816

I rather expected @ilyasut to have made a public post by now after all the discussion of the AI reading list he gave me. A canonical list of references from a leading figure would be appreciated by many. I would be curious myself about what he would add from the last three years.

Papers

geohot

#include <stdlib.h>
#include <stdbool.h>
#include <tgmath.h>
#define max(x,y) ((x>y)?x:y)
#define half __fp16

void E_(int* data0) {
  int val0 = data0[0];
  data0[0] = (val0+1);
}

geohot

[55883.721977] amdgpu:   map VA 0x702eae9d2000 - 0x702eae9d3000 in entry 0000000072d2b750
[55883.721996] amdgpu:   INC mapping count 1               
[55883.722133] kfd kfd: amdgpu: ioctl cmd 0xc0184b0c (#0xc), arg 0x7ffe16172bef
[55883.722238] gmc_v11_0_process_interrupt: 6 callbacks suppressed         
[55883.722250] amdgpu 0000:c3:00.0: amdgpu: [gfxhub] page fault (src_id:0 ring:24 vmid:8 pasid:32774, for process python3 pid 356134 thread python3 pid 356134)
[55883.722343] amdgpu 0000:c3:00.0: amdgpu:   in page starting at address 0x00000000aabbc000 from client 10
[55883.722391] amdgpu 0000:c3:00.0: amdgpu: GCVM_L2_PROTECTION_FAULT_STATUS:0x00800A30
[55883.722429] amdgpu 0000:c3:00.0: amdgpu:      Faulty UTCL2 client ID: CPC (0x5)
[55883.722466] amdgpu 0000:c3:00.0: amdgpu:      MORE_FAULTS: 0x0   
[55883.722497] amdgpu 0000:c3:00.0: amdgpu:      WALKER_ERROR: 0x0         

geohot

# -*- coding: utf-8 -*-
#
# TARGET arch is: ['-D__HIP_PLATFORM_AMD__', '-I/opt/rocm/include']
# WORD_SIZE is: 8
# POINTER_SIZE is: 8
# LONGDOUBLE_SIZE is: 16
#
import ctypes

geohot

# tiny@tiny9:~/tinygrad$ python3 examples/benchmark_copies.py
# CPU copy 6.18 ms, 16.28 GB/s
# GPU copy 4.38 ms, 23.00 GB/s
# GPU  6x  1.85 ms, 54.54 GB/s

import time
def timeit(fxn):
  tms = []
  for _ in range(10):
    st = time.perf_counter()

geohot

__kernel void matmul(__global float* data0, const __global float* data1, const __global float* data2) {
  int gidx0 = get_group_id(1); /* 512 */
  int gidx1 = get_group_id(0); /* 512 */
  float2 acc0 = (float2)(0.0f,0.0f);
  float2 acc1 = (float2)(0.0f,0.0f);
  for (int ridx0 = 0; ridx0 < 512; ++ridx0) {
    float2 val0 = (float2)(*((__global float2*)(data1+(gidx0*2048)+(ridx0*2))));
    float2 val1 = (float2)(*((__global float2*)(data1+(gidx0*2048)+(ridx0*2)+1024)));
    float2 val2 = (float2)(*((__global float2*)(data2+(gidx1*2)+(ridx0*2048))));
    float2 val3 = (float2)(*((__global float2*)(data2+(gidx1*2)+(ridx0*2048)+1024)));

thecharlieblake

from typing import Callable, List

import numpy as np
import torch
from torch._dynamo.backends.common import aot_autograd
from torch.fx.graph_module import GraphModule

# NOTE: requires torch >= 2.1.0

def np2torch(fn: Callable) -> Callable:

thesephist

import torch
import torch.nn as nn
import torch.nn.functional as F

from transformers import GPT2Tokenizer, GPT2LMHeadModel

ppl_model_name = 'gpt2-xl' if device == 'cuda' else 'gpt2'
ppl_tokenizer = GPT2Tokenizer.from_pretrained(ppl_model_name)
load_opts = {
    'device_map': 'auto',