vadimkantorov

import datetime
import re

# https://regexlib.com/REDetails.aspx?regexp_id=3344&AspxAutoDetectCookieSupport=1
# http://stackoverflow.com/questions/3143070/javascript-regex-iso-datetime/7221570#7221570
re_iso8601date = r'^(?:(?=[02468][048]00|[13579][26]00|[0-9][0-9]0[48]|[0-9][0-9][2468][048]|[0-9][0-9][13579][26])\d{4}(?:(-|)(?:(?:00[1-9]|0[1-9][0-9]|[1-2][0-9][0-9]|3[0-5][0-9]|36[0-6])|(?:01|03|05|07|08|10|12)(?:\1(?:0[1-9]|[12][0-9]|3[01]))?|(?:04|06|09|11)(?:\1(?:0[1-9]|[12][0-9]|30))?|02(?:\1(?:0[1-9]|[12][0-9]))?|W(?:0[1-9]|[1-4][0-9]|5[0-3])(?:\1[1-7])?))?)$|^(?:(?![02468][048]00|[13579][26]00|[0-9][0-9]0[48]|[0-9][0-9][2468][048]|[0-9][0-9][13579][26])\d{4}(?:(-|)(?:(?:00[1-9]|0[1-9][0-9]|[1-2][0-9][0-9]|3[0-5][0-9]|36[0-5])|(?:01|03|05|07|08|10|12)(?:\2(?:0[1-9]|[12][0-9]|3[01]))?|(?:04|06|09|11)(?:\2(?:0[1-9]|[12][0-9]|30))?|(?:02)(?:\2(?:0[1-9]|1[0-9]|2[0-8]))?|W(?:0[1-9]|[1-4][0-9]|5[0-3])(?:\2[1-7])?))?)$'

for yyyy in range(1000, 3001):
    for mm in range(1, 13):
        for dd in range(1, 32):

vadimkantorov

# git clone https://gist.github.com/vadimkantorov/07a421f676f8af94457732f8e9127dde
# cd 07a421f676f8af94457732f8e9127dde
# uv venv
# uv sync
# source .venv/bin/activate # uv run python

# https://github.com/pytorch/pytorch/issues/160875#issuecomment-4183789164
# https://github.com/vllm-project/vllm/issues/24218

[build-system]

vadimkantorov

# python html2xhtml.py test.html test.xhtml

import html.parser
import xml.dom.minidom

class HTMLToMinidom(html.parser.HTMLParser):
    # https://developer.mozilla.org/en-US/docs/Glossary/Void_element
    def __init__(self, void_elements = ('area', 'base', 'br', 'col', 'embed', 'hr', 'img', 'input', 'link', 'meta', 'param', 'source', 'track', 'wbr')):
        super().__init__()
        self.doc = xml.dom.minidom.Document()

vadimkantorov

alias cmdline='ps --no-headers -o comm,args -p'
cmdline 1

# can produce JSON, but without process args
alias notbrokenjson='ps --no-headers -o '\''{"pcpu": %C, "group": "%G", "PPID": %P, "user": "%U", "comm": "%c", "rgroup": "%g", "nice": "%n", "pid": %p, "pgid": %r, "etime": "%t", "ruser": "%u", "time": "%x", "tty": "%y", "vsz": %z}'\'' -p'
notbrokenjson 1 > notbroken.json

# does not remove escape sequences or quotes in args - this will break JSON format for some processes
alias brokenjson='ps --no-headers -o '\''{"pcpu": %C, "group": "%G", "PPID": %P, "user": "%U", "args": "%a", "comm": "%c", "rgroup": "%g", "nice": "%n", "pid": %p, "pgid": %r, "etime": "%t", "ruser": "%u", "time": "%x", "tty": "%y", "vsz": %z}'\'' -p'
brokenjson 1 > brokenjson.json

vadimkantorov

# git clone https://gist.github.com/vadimkantorov/fe63f8628ff6cad460e934e1d7ed650b
# cd fe63f8628ff6cad460e934e1d7ed650b
# uv venv
# uv sync

# https://github.com/vllm-project/vllm/pull/20358#issuecomment-3247178818
# https://github.com/vllm-project/vllm/issues/9244
# https://github.com/astral-sh/uv/issues/8082
# https://github.com/vllm-project/vllm/issues/24126

vadimkantorov

# https://stackoverflow.com/questions/71944328/how-to-implement-json-format-logs-in-python

import json
import logging

# https://docs.python.org/3/library/logging.html#logrecord-attributes
formatter = type('JsonFormatter', (logging.Formatter, ), dict(format = lambda self, record: json.dumps(dict(time = self.formatTime(record), level = record.levelname, message = record.getMessage(), module = record.module, lineno = record.lineno)) ))()

# simpler version below does not escape quotes in message and does not delete newlines in message
# formatter = logging.Formatter('{\"time\": \"%(asctime)-s\", \"level\": \"%(levelname)-s\", \"message\": \"%(message)s\", \"module\": \"%(module)s\", \"lineno\": %(lineno)d}')

vadimkantorov

echo '{"site": {"hello": "world"}}' > foo.json
echo 'Hello {{ site.hello }}' > foo.txt.j2

bash j2.sh -t foo.txt.j2 -i foo.json -o foo.txt
cat foo.txt

source function.sh
j2 -t foo.txt.j2 -i foo.json -o foo.txt
cat foo.txt
unset -f j2

vadimkantorov

# Extracted and simplified the two-level aggregation approach (first, parallel aggregation in blocks, then final sequential aggregation) from https://github.com/volcengine/verl/blob/main/verl/utils/kernel/kernels.py
# Examples of single-level sequential, online aggregation approaches:
# - https://github.com/linkedin/Liger-Kernel/blob/main/src/liger_kernel/ops/cross_entropy.py
# - https://github.com/Dao-AILab/flash-attention/blob/main/flash_attn/ops/triton/cross_entropy.py
# logsumexp_torch has some eager pseudo/code in PyTorch which emulates what Triton does, except that BLOCK_SIZE_M equials to M
# tl.program_id(axis=0).to(tl.int64) is used for https://arxiv.org/abs/2410.10989 and https://github.com/linkedin/Liger-Kernel/blob/05b43a14913ced3776aa3fc50020089b8c0d63c1/src/liger_kernel/ops/cross_entropy.py#L77-L79
# sample_verl.pt is derived from the inputs (logits = torch.matmul(hidden, weights) uploaded by @WindowsXP-Beta in https://github.com/volcengine/verl/issues/2656#issuecomment-3131136498 )
# created for

vadimkantorov

nvidia-smi -q -x | grep "</pid>" | tr -d "</pid>\t"

vadimkantorov

# https://github.com/pytorch/pytorch/issues/158710
# https://github.com/pytorch/pytorch/issues/158698
# https://github.com/pytorch/pytorch/issues/69431

import torch

def to_(tensor1d, dtype, *, chunks = 0, split_size = 0):
    # TODO: instead of clone() maybe could copy_ into a buffer, clone() does not allow using a buffer
    # TODO: unclear if these codes can support autograd, and if so, will it remember too much in saved_for_backward