julien-c

#!/usr/bin/env python3
"""Traffic hammer + canary for reproducing llama-server output corruption.

Mimics the production traffic pattern that preceded the 2026-06-03 '/////'
corruption: small chats with a growing shared prefix (checkpoint restores),
long generations cancelled mid-stream, 4-way parallel bursts, slot swapping,
and >10s idle gaps (CUDA/HIP graph-cache eviction boundary).

After every round, a temperature-0 canary probe checks output sanity.
On canary failure the harness logs the round + action history and exits 42.

julien-c

import type { ExtensionAPI } from "@earendil-works/pi-coding-agent";
import { Type } from "typebox";
import { Compile } from "typebox/compile";

const DEFAULT_BASE_URL = "http://localhost:8080/v1";
const PROPS_TIMEOUT_MS = 120_000;

const ModelsResponseSchema = Type.Object({
	data: Type.Optional(
		Type.Array(

julien-c

import { McpServer } from "@modelcontextprotocol/sdk/server/mcp.js";
import { StdioServerTransport } from "@modelcontextprotocol/sdk/server/stdio.js";
import { z } from "zod";
import { spawn } from "node:child_process";
import puppeteer, { Browser, BrowserContext, Page } from "puppeteer-core";
import { fileURLToPath } from "url";
import { dirname } from "path";
import { setTimeout } from "node:timers/promises";

const LIGHTPANDA_PORT = 9222;

julien-c

Map(98) {
  'unspecified' => 6519,
  'unknown' => 230,
  'cc-by-4.0' => 131,
  'cc-by-sa-4.0' => 76,
  'mit' => 73,
  'apache-2.0' => 41,
  'cc0-1.0' => 40,
  'cc-by-nc-sa-4.0' => 36,
  'cc-by-nc-4.0' => 21,

julien-c

import os
import subprocess
from huggingface_hub.hf_api import HfApi
from huggingface_hub.repository import Repository

os.makedirs("./Helsinki-NLP/", exist_ok=True)

os.environ["GIT_LFS_SKIP_SMUDGE"] = "1"
# Important, to not download the large files

julien-c

export class ModelInfo {
	/**
	 * Key to config.json file.
	 */
	key:          string;
	etag:         string;
	lastModified: Date;
	size:         number;
	modelId:      ModelId;
	author?:      AuthorId;

julien-c

API Name	Memory	vCPUs	Physical Processor	Network Performance	Linux On Demand cost	Linux Reserved cost
a1.2xlarge	16.0 GiB	8 vCPUs	AWS Graviton Processor	Up to 10 Gigabit	$148.92 monthly	$93.80 monthly
a1.4xlarge	32.0 GiB	16 vCPUs	AWS Graviton Processor	Up to 10 Gigabit	$297.84 monthly	$187.61 monthly
a1.large	4.0 GiB	2 vCPUs	AWS Graviton Processor	Up to 10 Gigabit	$37.23 monthly	$23.43 monthly
a1.medium	2.0 GiB	1 vCPUs	AWS Graviton Processor	Up to 10 Gigabit	$18.61 monthly	$11.75 monthly
a1.metal	32.0 GiB	16 vCPUs	AWS Graviton Processor	Up to 10 Gigabit	$297.84 monthly	$187.61 monthly
a1.xlarge	8.0 GiB	4 vCPUs	AWS Graviton Processor	Up to 10 Gigabit	$74.46 monthly	$46.93 monthly
c1.medium	1.7 GiB	2 vCPUs	Intel Xeon Family	Moderate	$94.90 monthly	$66.43 monthly

julien-c

import logging

from transformers.modeling_bart import BartForSequenceClassification
from transformers.pipelines import TextClassificationPipeline
from transformers.tokenization_bart import BartTokenizer


logging.basicConfig(level=logging.INFO)

julien-c

To get it to work please follow the steps:

1) Create test user by toggling the Authorize test users for this app and grant permissions "manage_pages" and "page_messaging".
2) Use the Edit Button and get an access token for this user (using v2.6). Please save this for later.
3) Use edit button to login as the test user
4) After login, create page as the test user
5) Use the user access token for the test user to get the page access token for this user. You can do this with the following call:
~~~
https://graph.facebook.com/v2.6/me/accounts?access_token=<TEST_USER_ACCESS_TOKEN>
~~~

julien-c

import spacy
nlp = spacy.load('en')
doc = nlp(text)
# At this point entity annotations are in `doc.ents`

for token in doc:
  print tok.ent_iob_ + '-' + mapping[tok.ent_type_]  # `I-PER`, etc.