The thread tried to read from or write to a virtual address for which it does not have the appropriate access.

Hi,

I hope that all the provided information meets your criteria.

- Nuitka version, full Python version, flavor, OS, etc.:

> D:\>python -m nuitka --version

2.1.2 Commercial: None Python: 3.11.4 (tags/v3.11.4:d2340ef, Jun 7 2023, 05:45:37) [MSC v.1934 64 bit (AMD64)] Flavor: CPython Official Executable: C:\prgFiles\Python\Python311\python.exe OS: Windows Arch: x86_64 WindowsRelease: 10 Version C compiler: ~\AppData\Local\Nuitka\Nuitka\Cache\downloads\gcc\x86_64\13.2.0-16.0.6-11.0.1-msvcrt-r1\mingw64\bin\gcc.exe (gcc 13.2.0).

- How did you install Nuitka and Python Nuitka is installed into env using pip, as well as all the other packages:

> python -m pip freeze

D:>python -m pip freeze Nuitka==2.1.2 numpy==1.23.5 pandas==2.0.2 scipy==1.10.1 tensorboard==2.12.3 torch==2.0.1 tqdm==4.65.0 transformers==4.30.2 ....

-Description of issue: Executable file crashes after few seconds. I checked in C:\Users<user_nanne>\AppData\Local\CrashDumps\PretrainMODAI.exe.6012.dmp and I came across this information:

Process Name: <path_file>\PretrainMODAI.exe Process Architecture: x64 Exception Code: 0xc0000005 Exception Information: The thread tried to read from or write to a virtual address for wich it does not have the appropriate access. Heap Information: Not Present Error Information: OS Version: 10.0.14393

- Many times when you get an error from Nuitka, your setup may be special:

Hello World test works!

- Also supply a Short, Self Contained, Correct, Example: Unfortunately, I possess three tensor files totaling 3.4 GB located within the "pretrained_tokenizer/pretrained_tensors" directory, accompanied by two JSON files (vocab.json and config.json) also located within the "pretrained_tokenizer" directory. Additionally, there are 82 text files totaling 25 MB within the "source_tokenizer" directory. Regrettably, due to their extensive size, these files cannot be provided for testing purposes.

# Test version for converting in exe file
import os,sys
import transformers
import torch
from pathlib import Path
from tqdm import tqdm
import random
from torch.utils.tensorboard import SummaryWriter


def get_config():
    return {   
        "root": os.path.dirname(os.path.realpath(__name__)), 
        "file_name_source": "postgres.txt",        
        "folder_source": "source_ini",      
        "folder_source_tokenizer": "source_tokenizer",   
        "folder_tokenizer": "pretrained_tokenizer", 
        "folder_tensors": "pretrained_tensors",         
        "batch_size": 32,
        "num_epochs": 4,
        "lr": 1e-5,
        "seq_len": 350,
        "max_len": 512,
        "model_folder": "model_results",
        "model_pretrained": "pretrained_model",
        "model_trained": "trained_model",              
        "logging_dir": "logging_dir",        
        "model_basename": "tmodel_",
        "preload": None,
        "tokenizer_file": "tokenizer_{0}.json",
        "experiment_name": "runs/tmodel"
    }

    
def get_model_folder_base_path(config):
    model_root = config["root"]
    model_folder = config["model_folder"]    
    return str(Path('.') / model_root / model_folder)    

#Folder of pretrained tensors files
def get_model_pretrained_tensors_path(config):
    model_root = config["root"]
    model_folder = config["folder_tokenizer"]  
    model_basename = config["folder_tensors"]     
    return str(Path('.') / model_root / model_folder / model_basename ) 
    
#Folder for pretrained model    
def get_model_folder_base_pretrained_path(config):
    model_root = config["root"]
    model_folder = config["model_folder"]  
    model_basename = config["model_pretrained"]   
    folder=str(Path('.') / model_root / model_folder / model_basename ) 
    print(f"Root folder: {model_root}")
    return str(Path('.') / model_root / model_folder / model_basename )        
    
#Folder source for pretraining model
def get_folder_source_tokenizer_path(config):
    model_root = config["root"]
    model_basename = config["folder_source_tokenizer"]
    return str(Path('.') / model_root / model_basename) 

#Folder source of tokenizer
def get_tokenizer_folder_path(config):
    model_root = config["root"]
    model_basename = config["folder_tokenizer"]
    return str(Path('.') / model_root / model_basename) 
    
#The file json of tokenizer (vocabulary)    
def get_tokenizer_file_path(config,version):
    model_root = config["root"]
    model_basename = config["folder_tokenizer"]
    file_basename = config["tokenizer_file"].format(version)
    return str(Path('.') / model_root / model_basename / f'{file_basename}')


def mlm(tensor,seed):
    # create random array of floats with equal dims to tensor
    torch.manual_seed(seed)
    rand = torch.rand(tensor.shape)
    # mask random 15% where token is not 0 <s>, 1 <unk>,2 <s/> or 4 <pad>
    mask_arr = (rand < .15) * (tensor != 0) * (tensor != 1) * (tensor != 2) * (tensor != 4)
    # loop through each row in tensor (cannot do in parallel)
    for i in range(tensor.shape[0]):
        # get indices of mask positions from mask array
        selection = torch.flatten(mask_arr[i].nonzero()).tolist()
        # mask tensor
        tensor[i, selection] = 3 #mask id
    return tensor


class Dataset(torch.utils.data.Dataset):
    def __init__(self, encodings):
        # store encodings internally
        self.encodings = encodings

    def __len__(self):
        # return the number of samples
        return self.encodings['input_ids'].shape[0]

    def __getitem__(self, i):
        # return dictionary of input_ids, attention_mask, and labels for index i
        return {key: tensor[i] for key, tensor in self.encodings.items()}
	
	
def main():
    
	config=get_config()
	SEED=42#By default
	torch.manual_seed(SEED)
	NUMBER_OF_EPOCHS=1


	"""
	Pretrain model 
	"""

	# initialize the tokenizer using the tokenizer we initialized and saved to file
	##Folder with trained tokenizers
	tokenizer_pretrained_folder = get_tokenizer_folder_path(config)
	##Folder with splited source file
	tokenizer_source_folder=get_folder_source_tokenizer_path(config)
	print(f"tokenizer_source_folder: {tokenizer_source_folder}")
	max_len = config["max_len"]
	tokenizer = transformers.RobertaTokenizer.from_pretrained(tokenizer_pretrained_folder, max_len=max_len)
	VOCAB_SIZE=tokenizer.vocab_size
	print ("VOCAB_SIZE",VOCAB_SIZE)


	#set folder base model
	model_base_name=get_model_folder_base_path(config)

	if not os.path.exists(model_base_name):
		#create folder target
		os.makedirs(model_base_name,exist_ok=True)
		print(f"Nothing to do. The model folder [ {model_base_name} ] doesn't exists!")
		#prepare pretraining folder
		model_pretrainig_name=get_model_folder_base_pretrained_path(config)
		os.makedirs(model_base_name,exist_ok=True)
		sys.exit() 

	path_pretrained_model=get_model_folder_base_pretrained_path(config)
	#print("path_pretrained_model: ",path_pretrained_model)

	if os.path.exists(path_pretrained_model):
		print(f'The folder model [ {path_pretrained_model} ] already exists!')
		sys.exit()


	#Source of file text for training
	paths = [str(x) for x in Path(tokenizer_source_folder).glob('**/*.txt')]
	# initialize lists of tensors
	input_ids = []
	mask = []
	labels = []
	# open all files, encode and add to single dataset
	for path in tqdm(paths[:200]):
		# :50
		# open the file and split into list by newline characters
		with open(path, 'r', encoding='utf-8') as fp:
			lines = fp.read().split('\n')
		# encode
		sample = tokenizer(lines, max_length=max_len, truncation=True, padding='max_length')#config["max_len"]
		# convert tokens to tensor
		labels.append(torch.tensor(sample.input_ids))
		# create attention mask tensor
		mask.append(torch.tensor(sample.attention_mask))
		# mask ~15% of tokens to create inputs
		input_ids.append(mlm(labels[-1].detach().clone(),SEED))
		
	# convert lists of tensors into tensors
	input_ids = torch.cat(input_ids)
	mask = torch.cat(mask)
	labels = torch.cat(labels)

	pretrained_tensors_path = get_model_pretrained_tensors_path(config)

	if not os.path.exists(pretrained_tensors_path):
		os.mkdir(pretrained_tensors_path)
		#save tensors
		torch.save(input_ids, f'{pretrained_tensors_path}\\input_ids.pt')
		torch.save(mask, f'{pretrained_tensors_path}\\attention_mask.pt')
		torch.save(labels, f'{pretrained_tensors_path}\\labels.pt')

		del input_ids, mask, labels
	   
	#reload tensors from saved data
	input_ids = torch.load(f'{pretrained_tensors_path}\\input_ids.pt')
	mask = torch.load(f'{pretrained_tensors_path}\\attention_mask.pt')
	labels = torch.load(f'{pretrained_tensors_path}\\labels.pt')

	#building dataset
	print("Building dataset ...")
	encodings = {'input_ids': input_ids, 'attention_mask': mask, 'labels': labels}


	#initialise dataset
	dataset = Dataset(encodings)
	#initilise dataloader
	loader = torch.utils.data.DataLoader(dataset, batch_size=64, shuffle=True)

	#create a RoBERTa config object, which will describe which features we want to initialize our RoBERTa model with.

	DROPOUT_PROB = 0.15      # Set the dropout probability ( force neurons to learns more robust)
	configRoBERTa = transformers.RobertaConfig(
		vocab_size=VOCAB_SIZE,  # we align this to the trained_tokenizer vocab
		max_position_embeddings=514,
		hidden_size=768, # (hidden_size must be multiple of num_attention_heads)
		num_attention_heads=16,
		num_hidden_layers=8,
		type_vocab_size=2,
		hidden_dropout_prob=DROPOUT_PROB,
		attention_dropout_prob=DROPOUT_PROB
	)

	
	#define folder for pretraining model
	model_folder_pretraining = get_model_folder_base_pretrained_path(config)

	if not os.path.exists(model_folder_pretraining):
		os.mkdir(model_folder_pretraining)
	#initialize a RoBERTa model with a language modeling head.
	model = transformers.RobertaForMaskedLM(configRoBERTa)

	#prepare pretraining
	device = torch.device('cuda') if torch.cuda.is_available() else torch.device('cpu')
	# and move our model over to the selected device
	model.to(device)

	# activate pretraining mode
	model.train()
	# initialize optimizer
	optim = transformers.AdamW(model.parameters(), lr=1e-5)

	# optionally, if using tensorboard, initialize writer object 
	writer = SummaryWriter(log_dir=model_folder_pretraining)      

	epochs = NUMBER_OF_EPOCHS  # pretrained for 4 ideally
	step = 0

	
	# Pretraining loop by epochs from here ....
	print ("Done")


if __name__ == "__main__":
    main()

- Provide in your issue the Nuitka options used: py -m nuitka --standalone --nofollow-import-to=tensorflow --nofollow-import-to=keras --noinclude-default-mode=warning --noinclude-numba-mode=nofollow --module-parameter=numba-disable-jit=yes --assume-yes-for-downloads --module-parameter=torch-disable-jit=yes --include-data-dir=../PretrainTestNuitka/model_results=model_results --include-data-dir=../PretrainTestNuitka/pretrained_tokenizer=pretrained_tokenizer --include-data-dir=../PretrainTestNuitka/source_tokenizer=source_tokenizer MyScript2Test.py

-Messages during conversion:

Nuitka-Options: Used command line options: --standalone --nofollow-import-to=tensorflow --nofollow-import-to=keras Nuitka-Options: --noinclude-default-mode=warning --noinclude-numba-mode=nofollow --module-parameter=numba-disable-jit=yes Nuitka-Options: --assume-yes-for-downloads --module-parameter=torch-disable-jit=yes Nuitka-Options: --include-data-dir=../PretrainTestNuitka/model_results=model_results Nuitka-Options: --include-data-dir=../PretrainTestNuitka/pretrained_tokenizer=pretrained_tokenizer Nuitka-Options: --include-data-dir=../PretrainTestNuitka/source_tokenizer=source_tokenizer MyScript2Test.py Nuitka: Starting Python compilation with Nuitka '2.1.2' on Python '3.11' commercial grade 'not installed'. Nuitka-Options:WARNING: No data files in directory '../PretrainTestNuitka/model_results.' Nuitka: Completed Python level compilation and optimization. Nuitka: Generating source code for C backend compiler. Nuitka: Running data composer tool for optimal constant value handling. Nuitka: Running C compilation via Scons. Nuitka-Scons: Backend C compiler: gcc (gcc 13.2.0). Nuitka-Scons: Backend linking program with 4768 files (no progress information available for this stage). Nuitka-Scons: Compiled 4738 C files using ccache. Nuitka-Scons: Cached C files (using ccache) with result 'cache hit': 4736 Nuitka-Scons: Cached C files (using ccache) with result 'cache miss': 2 Nuitka-Options: Included data file 'pretrained_tokenizer\config.json' due to specified data dir Nuitka-Options: '../PretrainTestNuitka/pretrained_tokenizer' on command line. Nuitka-Options: Included data file 'pretrained_tokenizer\merges.txt' due to specified data dir Nuitka-Options: '../PretrainTestNuitka/pretrained_tokenizer' on command line. Nuitka-Options: Included data file 'pretrained_tokenizer\vocab.json' due to specified data dir Nuitka-Options: '../PretrainTestNuitka/pretrained_tokenizer' on command line. Nuitka-Options: Included data file 'pretrained_tokenizer\pretrained_tensors\attention_mask.pt' due to specified data dir Nuitka-Options: '../PretrainTestNuitka/pretrained_tokenizer' on command line. Nuitka-Options: Included data file 'pretrained_tokenizer\pretrained_tensors\input_ids.pt' due to specified data dir Nuitka-Options: '../PretrainTestNuitka/pretrained_tokenizer' on command line. Nuitka-Options: Included data file 'pretrained_tokenizer\pretrained_tensors\labels.pt' due to specified data dir Nuitka-Options: '../PretrainTestNuitka/pretrained_tokenizer' on command line. Nuitka-Options: Included 82 data files due to specified data dir '../PretrainTestNuitka/source_tokenizer' on command line. Nuitka-Plugins:data-files: Included data file 'certifi\cacert.pem' due to package data for 'certifi'. Nuitka-Plugins:data-files: Included data file 'grpc_cython_credentials\roots.pem' due to package data for 'grpc'. Nuitka-Plugins:data-files: Included data file 'pandas\io\formats\templates\html.tpl' due to package data directory 'templates' for Nuitka-Plugins:data-files: 'pandas.io.formats'. Nuitka-Plugins:data-files: Included data file 'pandas\io\formats\templates\html_style.tpl' due to package data directory Nuitka-Plugins:data-files: 'templates' for 'pandas.io.formats'. Nuitka-Plugins:data-files: Included data file 'pandas\io\formats\templates\html_table.tpl' due to package data directory Nuitka-Plugins:data-files: 'templates' for 'pandas.io.formats'. Nuitka-Plugins:data-files: Included data file 'pandas\io\formats\templates\latex.tpl' due to package data directory 'templates' for Nuitka-Plugins:data-files: 'pandas.io.formats'. Nuitka-Plugins:data-files: Included data file 'pandas\io\formats\templates\latex_longtable.tpl' due to package data directory Nuitka-Plugins:data-files: 'templates' for 'pandas.io.formats'. Nuitka-Plugins:data-files: Included data file 'pandas\io\formats\templates\latex_table.tpl' due to package data directory Nuitka-Plugins:data-files: 'templates' for 'pandas.io.formats'. Nuitka-Plugins:data-files: Included data file 'pandas\io\formats\templates\string.tpl' due to package data directory 'templates' Nuitka-Plugins:data-files: for 'pandas.io.formats'. Nuitka-Plugins:data-files: Included 602 data files due to package data directory 'zoneinfo' for 'pytz'. Nuitka-Plugins:data-files: Included data file 'scipy\stats_sobol_direction_numbers.npz' due to package data for 'scipy'. Nuitka-Plugins:data-files: Included 8101 data files due to package data directory 'include' for 'torch'. Nuitka-Plugins:dll-files: Found 1 file DLLs from numpy installation. Nuitka-Plugins:dll-files: Found 2 files DLLs from sklearn._distributor_init installation. Nuitka-Plugins:dll-files: Found 11 files DLLs from torch installation. Nuitka: Keeping build directory 'MyScript2Test.build'. Nuitka: Successfully created 'MyScript2Test.dist\MyScript2Test.exe'.