Lyken17

from llava.wids import ShardListDataset

train_url = "https://storage.googleapis.com/webdataset/fake-imagenet/imagenet-train.json"
'''
{
  "__kind__": "wids-shard-index-v1",
  "wids_version": 1,
  "shardlist": [
    {

Lyken17

import time
from collections import defaultdict
import os, os.path as osp
from datasets import load_dataset
import torch
import json
from transformers import AutoTokenizer, AutoModel, AutoModelForCausalLM, AutoConfig
from tqdm import tqdm

Lyken17

In [42]: k
Out[42]: 'https://www.danezon.com/wp-content/uploads/2021/01/The-Man-From-Toronto-Kevin-Hart-Jacket-510x612.jpg'

In [43]: f[k]
Out[43]:
{'query': '<image> Can you briefly explain the content in the image?',
 'orig_text': 'Kevin Hart The Man From Toronto Bomber Jacket',
 'output': "In the image, the renowned actor Kevin Hart is captured in a moment of quiet intensity. He stands in front of a vibrant market stall, his gaze directed off to the side, suggesting a moment of deep thought or contemplation. Dressed in a navy blue jacket and a green sweater, he exudes a casual yet stylish vibe. His hands are tucked away in his pockets, a common gesture that often signifies relaxation or introspection.\n\nThe market stall behind him is a riot of colors, with various items on display. The exact nature of these items is not clear from the image, but they add to the lively atmosphere of the scene. The background is filled with people, their faces blurred, indicating that they are not the focus of the image.\n\

Lyken17

import glob
import tarfile
import json
import os, os.path as osp
from io import BytesIO
from PIL import Image, ImageFile
import hashlib
from torch.utils.data import Dataset, get_worker_info, ConcatDataset
from multiprocessing.pool import ThreadPool as Pool

Lyken17

import torch
from diffusers import DiffusionPipeline
from diffusers.models.attention_processor import LoRAAttnProcessor

from utils import print_gpu_utilization


# pipe = DiffusionPipeline.from_pretrained("stabilityai/stable-diffusion-xl-base-1.0")
pipe = DiffusionPipeline.from_pretrained(
    "runwayml/stable-diffusion-v1-5", torch_dtype=torch.float16, variant="fp16"

Lyken17

from typing import Any
import torch
import torch.nn as nn
import torch.nn.functional as F
from torch import autograd


def scientific_precision(number):
    suffix = ["KB", "MB", "GB", "TB", "PB"]
    for idx_, s in enumerate(suffix):

Lyken17

import torch
from torch import nn

# net = nn.Linear(500, 500)
# input = torch.randn(64, 500)

net = nn.Conv2d(3, 3, kernel_size=3, padding=1)
input = torch.randn(1, 3, 32, 32)

# only calculate input grad, prints ('_saved_mat2', torch.Size([500, 500]))

Lyken17

# e408b27

MODEL=facebook/opt-6.7b
torchrun --nproc_per_node=8 --master_port=24567 train.py \
    --model_name_or_path $MODEL \
    --data_path ./alpaca_data.json \
    --bf16 True \
    --output_dir ./output/$MODEL \
    --num_train_epochs 3 \
    --per_device_train_batch_size 2 \

Lyken17

batch_size, seq_length, embed_dim = x.size()
# B, T, D
qkv = self.qkv_proj(x) # B, T, 3xE
# head_dim = embed_dim // num_heads

# Separate Q, K, V from linear output
qkv = qkv.reshape(batch_size, seq_length, self.num_heads, 3 * self.head_dim) # B, T, H, 3xHD
qkv = qkv.permute(0, 2, 1, 3)  # B, H, T, 3xHD
q, k, v = qkv.chunk(3, dim=-1) # B, H, T, HD 

Lyken17

import numpy as np

from collections import Counter

import tvm
from tvm import relay
from tvm.relay import ExprFunctor, ExprMutator, ExprVisitor
from tvm.relay.expr_functor import ExprMutator, Call
from tvm.relay.dataflow_pattern import wildcard, is_op, is_constant, is_expr, rewrite, DFPatternCallback