vadimkantorov

# Nebius s3
# sudo apt-get install s3cmd # export the vars below or configure access_key / secret_key
# export AWS_ACCESS_KEY_ID=...
# export AWS_SECRET_ACCESS_KEY=...
# https://s3tools.org/usage
# s3cmd -c ~/.nebius.s3cfg ls

[default]
# access_key=...
# secret_key=...

vadimkantorov

# PYTHONPATH=. python ...

import os
import torch

def cuda_oom_hook(device, alloc, device_alloc, device_free, info = dict(counter = 0), snapshot_dump_file_pattern = './memory_snapshot_{pid}_{oom_counter}.pt'):
    memory_summary = torch.cuda.memory_summary(device = device)
    memory_snapshot = torch.cuda.memory._snapshot(device = device)
    pid = os.getpid()
    print('device:', device, 'oom#:', info['oom_counter'], 'pid:', pid, 'alloc:', alloc, 'device_alloc:', device_alloc, 'device_free:', device_free)

vadimkantorov

import multiprocessing
import itertools

inputs = list(range(111))
batchsize = 10
num_workers = 4

batches = itertools.batched(inputs, batchsize)

def reducer(xs):

vadimkantorov

# Usage: bash cache_hf_model.sh Qwen/Qwen3-8B
# export HF_HOME=/my/cache/HF_HOME
python -c 'import sys, transformers; transformers.AutoModel.from_pretrained(sys.argv[-1], trust_remote_code=True, device_map="meta")' $@

vadimkantorov

def load_dotenv(dotenv_path = '.env'):
    # https://snarky.ca/use-toml-for-env-files/
    # https://github.com/theskumar/python-dotenv
    '''
    # such simple key-value files are toml subset and can be read via tomllib without external packages or hacks
    a="b"
    c="d"
    '''
    import os, tomllib
    os.environ.update(tomllib.load(open(dotenv_path, 'rb')))

vadimkantorov

import sys, fsspec
with fsspec.open(sys.argv[1], 'rt') as f: # must pass 'rt' explicitly, as in fsspec the default mode is 'rb'
    print(f.read()) # msut use context manager as in fsspec the result of fsspec.open(...) does not have method read()

# echo world > hello.txt
# python catfsspec.py hello.txt
# python catfsspec.py file://hello.txt
# python catfsspec.py s3://mybucket/hello.txt

vadimkantorov

# reference: https://gist.github.com/0xjac/85097472043b697ab57ba1b1c7530274

git clone --bare git@github.com:volcengine/verl.git
cd verl.git

# create a bare repo vadimkantorov/verl
git push --mirror git@github.com:vadimkantorov/verl.git
cd .. && rm -rf verl.git

# set up upstream remote

vadimkantorov

# Save as tqdm.py in project dir, then `from tqdm import tqdm; from tqdm.auto import tqdm` should pick up this class, if fails use export PYTHONPATH=.
# Test run: python tqdm.py

import os, sys

# huggingface_hub/hf_api.py:
# from tqdm.auto import tqdm as base_tqdm
# from tqdm.contrib.concurrent import thread_map
# https://tqdm.github.io/docs/shortcuts/#tqdmauto
sys.modules['tqdm.auto'] = sys.modules[__name__]

vadimkantorov

# https://github.com/pypa/pip/issues/11440
# https://github.com/pypa/pip/issues/7822
# https://stackoverflow.com/a/79598932/445810
# tomllib is available starting from python --version >= 3.11

python -m pip install $(python -c 'import tomllib;print(*tomllib.load(open("pyproject.toml","rb"))["project"]["dependencies"])') # --user --break-system-packages

vadimkantorov

import http.client
http.client.HTTPConnection.debuglevel = 1