ugovaretto

Monitoring the Two-Host TP2 Server

Where the logs live and how to see concurrency and speculative-decoding health for the gufo Qwen3.8-Flash-Next TP2 pair (see TP2-SETUP.md).

Log files

What	Where
rank 0 — all requests and events	strix-halo-1:~/tp2-swap.log (llama-swap wrapper tees gufo output here)

ugovaretto

Local TP2 build on the twin Strix Halo hosts (CachyOS)

How the gufo executables on strix-halo-1 and strix-halo-2 were built, and why the naive build fails on these machines.

The README is written for two toolchains: Nix (not installed here) or "With Nix-less" with the qualified toolchain GCC 15.3 + ROCm 7.2.3. These CachyOS boxes have GCC 16.2.1 and the rocm-gfx1151-bin package layout and no Nix, so three extra problems had to be worked around. Those are almost certainly why a naive build attempt fails.

ugovaretto

// Tampermonkey scriot: USE EXISTING HEADER, REMOVE THIS HEADER AND ONLY COPY THE CODE AFTER 'user strict'
// ==UserScript==
// @name         DeepSeek Local Nerd Font Canvas Enforcer
// @match        *://*:3080/*
// @run-at       document-start
// @grant        none
// ==UserScript==

(function() {
    'use strict';

ugovaretto

MANDATORY RULES

• Do not use XML tags like <tool_call> or <function> or any XML tag at all.
• You must respond strictly using standard OpenAI JSON tool calling syntax.
• If you need to call a tool, output a JSON object matching this schema: {"name": "tool_name", "arguments": {"arg1": "value", "arg2":"value",...,"argN":"value"}}.
• Stick to the plan, do not go off and do things which are not related to the current task.
• Follow user instructions e.g. if the user asks you to load a skill just load, undersrand and memorize the skill DO NOT USE IT until it's required.
• You can only access resources in the current directory; you are not allowed to access resources outside the current directory.
• Always think before making any chage to the code.
• Create a mental sandbox to simulate the impact of the chages you make before making them.

ugovaretto

 ssh -J <username>@<remote-cluster-login-address> -R 7777:localhost:7777 <username>@<compute-node name e.g. nid20080> -N

ugovaretto

# uv venv
# uv pip install gguf
# uv run gguf-jinja-dump.py <gguf ..0001 file>

from gguf import GGUFReader
import sys

def get_jinja_template(file_path):
    reader = GGUFReader(file_path)
    # The key is standard for tokenizer chat templates

ugovaretto

import asyncio
import time
from concurrent.futures import ThreadPoolExecutor, as_completed
import sys

async def do_async_work(task_id: int, duration: float = 0.1) -> str:
    await asyncio.sleep(duration)
    return f"Task {task_id} completed"

async def run_asyncio(tasks: int=5):

ugovaretto

{
  "$schema": "https://charm.land/crush.json",
  "mcp": {
    "websearch-fetch-mcp": {
      "type": "stdio",
      "command": "uv",
      "args": [
        "run",
        "--with", "fastmcp",
        "--with", "requests",

ugovaretto

#!/usr/bin/env python3
import os
import sys
import requests
import json
import argparse

def generate_bash_command(prompt):
    api_key = os.getenv('BASH_AI_KEY')
    api_url = os.getenv('BASH_URL_KEY', "http://localhost:11434/v1/chat/completions")

ugovaretto

from selenium import webdriver
from selenium.webdriver.common.by import By
from selenium.webdriver.chrome.service import Service
from webdriver_manager.chrome import ChromeDriverManager

# 1. Setup the WebDriver (using webdriver_manager for simplicity)
# This automatically downloads and manages the correct ChromeDriver version
service = Service(ChromeDriverManager().install())
driver = webdriver.Chrome(service=service)