wahengchang

Telegram Bot E2E 互動檢查表（跨專案範本 v1）

使用專用的測試 bot、測試帳號和測試聊天室。截圖與 log 不得含 bot token 或真實個資。勾選前，任何項目都不代表已通過。

0. 測試設定卡（每次測試填寫）

欄位	內容
專案／bot username
版本（commit／tag）

rh-operator-bundle-bot

Pipeline still running ...
PipelineRun is still running: Tasks Completed: 29 (Failed: 0, Cancelled 0), Incomplete: 6, Skipped: 19
[get-pr-number : parse-pr-url] 2026-10-13T01:25:15.03Z + echo -n 4270
[get-pr-number : parse-pr-url] 2026-10-13T01:25:15.03Z + tee /tekton/results/git_pr_number
[get-pr-number : parse-pr-url] 2026-10-13T01:25:15.06Z 4270

[acquire-lease : create-lease] 2026-10-13T01:25:24.83Z + calculate_duration_in_seconds 90m
[acquire-lease : create-lease] 2026-10-13T01:25:24.83Z + '[' m == m ']'
[acquire-lease : create-lease] 2026-10-13T01:25:24.83Z + TOTAL_DURATION_IN_SECONDS=5400
[acquire-lease : create-lease] 2026-10-13T01:25:24.83Z + export TOTAL_DURATION_IN_SECONDS

CrossGen-ai

ruvector 2026: Rotation-Ensemble RaBitQ — Faster, Higher-Recall 1-Bit Vector Search in Rust

TL;DR (150 chars): Rotation-Ensemble RaBitQ is a Rust 1-bit vector-quantization index that averages K independent rotations to cut variance 1/K at matched storage.

ruvector is a high-performance Rust vector-search runtime. This nightly experiment pits a new axis for RaBitQ compression — ensembling K independent rotations — against the more familiar multibit axis, on the same bit budget. All numbers below are from cargo run --release on Apple M4 Max / macOS 15.6.

Introduction

RaBitQ (SIGMOD 2024) crushes a d-float vector into d bits by storing sign(R·x) after a random orthogonal rotation R. It's unbiased, needs no codebook, has a Johnson–Lindenstrauss-style error bound, and runs on popcount+rotate kernels that fit in a few hundred lines of Rust. The follow-up multibit RaBitQ (VLDB 2025) spends K bits per coordinate on a non-uniform codebook under one rotat

sharpiethreesixtymedia

Last updated: October 12, 2026

SharpieBot is a private Discord bot run by Sharpie360 for The Sharpness Syndicate Discord server.

What we store

Your Discord user ID and server ID, to give and remove roles (Guest, Access, and the roles you pick in #welcome). Caravan Survival game data: your user ID, the role you play, your votes and actions, and whether you've played before. IDs of videos already posted, so YouTube alerts don't repeat. What we don't store

oxyc

AetherEngine I-frame scrub-preview performance findings on Apple TV

Context

• Client: Den on Apple TV / tvOS, using the stock AVPlayerViewController transport controls.
• Playback engine: AetherEngine 7.34.0.
• Source: a 1080p MKV obtained through a debrid provider.
• AetherEngine was configured with LoadOptions.serveIFramePlaylist = true, allowing AVKit to obtain native scrub thumbnails from the loopback HLS server.
• The stream URL and credentials have been omitted deliberately.

martandabhishek

<!DOCTYPE html>
<html>
<head>
<meta charset="UTF-8">
<meta name="viewport" content="width=device-width, initial-scale=1, shrink-to-fit=no">
<title>VBC Mobile E2E Test Results</title>

SYLX99999

import subprocess

def sh(c):
    p = subprocess.run(c, shell=True, capture_output=True, text=True)
    return ((p.stdout or '') + (p.stderr or '')).strip()

P = "systemctl show vas.service -p MemoryHigh -p MemoryMax -p MemoryCurrent -p CPUWeight -p ActiveState -p MainPID -p NRestarts"

print("[A] BEFORE")
print(" ", sh(P).replace(chr(10), ' | '))

chen-ganggang

# 随机生成问候语
import random

greet_list = ["你好", "Hi", "哈喽", "幸会"]
name = "朋友"

def random_greet():
    msg = random.choice(greet_list) + "，" + name
    return msg

andywang0191-stack

Running MoE LLMs Larger Than Your RAM: A Practical Guide to mmap, Expert Offloading and SSD Streaming with llama.cpp

Tuần này trên Hacker News có một project khá thú vị tên Janas: chạy LLM lớn hơn RAM bằng cách stream các expert từ SSD. Đọc xong mình nhớ lại hồi đầu năm, khi cố nhét một model MoE hơn 100B params vào con laptop 32GB RAM. Lần đầu thì máy treo cứng, swap kín ổ. Lần sau thì chạy được khoảng 8 tokens/s, vì lúc đó mình đã hiểu model MoE thực sự dùng bộ nhớ thế nào. Ở Việt Nam, card 24GB VRAM vẫn còn đắt, thuê GPU cloud tính theo USD cũng không rẻ, nên biết cách vắt kiệt phần cứng đang có là một kỹ năng đáng học. Bài này mình chia sẻ cách làm thực tế với llama.cpp: cái gì nên nằm trên GPU, cái gì để ở RAM, cái gì có thể để trên SSD.

Vì sao MoE chạy được khi lớn hơn RAM, còn dense model thì không

Với một dense model như Llama 3.1 70B, token nào cũng phải đi qua toàn bộ weights. Nếu weights không nằm hết trong RAM, mỗi token sẽ phải đọc lại hàng chục GB từ ổ đĩa. Kể cả NVMe Gen4 đọc được khoản

nonprofittechy