kemingy

from struct import unpack
import numpy as np


def read_vec(filepath: str, vec_type: np.dtype = np.float32):
    """Read vectors from a file. Support `fvecs`, `ivecs` and `bvecs` format.

    Args:
        filepath: The path of the file.
        vec_type: The type of the vectors, choose from [np.float32, np.int32, np.dtype('u1')]

kemingy

Test disk speed

• https://cloud.google.com/compute/docs/disks/benchmarking-pd-performance

• write

# sequential write
sudo fio --name=write_throughput --directory=. --numjobs=4 \
      --size=2G --time_based --runtime=5m --ramp_time=2s --ioengine=libaio \

kemingy

# use devel since vllm need to compile the paged attention
ARG base=nvidia/cuda:11.8.0-cudnn8-devel-ubuntu20.04

FROM ${base}

ENV DEBIAN_FRONTEND=noninteractive LANG=en_US.UTF-8 LC_ALL=en_US.UTF-8
ARG commit=main
ARG CONDA_VERSION=py310_23.3.1-0

RUN apt update && \

kemingy

import json
import base64

import msgpack
import numpy as np

dim = 1536
emb = np.random.rand(dim).astype(np.float32)
emb_list = emb.tolist()

kemingy

How to use

requires:

• https://github.com/vi/websocat
• https://github.com/websocket-client/websocket-client

# terminal 1
python mosec_websocket.py
# terminal 2

kemingy

package main

import (
	"fmt"
	"sync"

	"go.starlark.net/starlark"
	"go.starlark.net/starlarkstruct"
)

kemingy

from flask import Flask, request


app = Flask(__name__)


@app.route("/inference", methods=["POST"])
def inference():
    return request.get_data()

kemingy

kemingy

local wezterm = require 'wezterm';

function recompute_padding(window)
  local window_dims = window:get_dimensions();
  local overrides = window:get_config_overrides() or {}

  if not window_dims.is_full_screen then
    if not overrides.window_padding then
      -- not changing anything
      return;

kemingy

三要素

• 用笔--五执笔+九用
• 识势--永字八法+五势
• 裹束--大圈+小圈

主要书体

• 篆
• 隶