emaadmanzoor

Train nanochat on 2 NVIDIA DGX Sparks

Last updated: November 3, 2025

This guide is about training nanochat on 2 DGX Sparks linked via QSFP/CX7. I estimate the training to take about 5 days (about half the training time on 1 DGX Spark).

Prerequisites

Follow the following NVIDIA tutorials to link and test your Spark cluster:

emaadmanzoor

MATCH (source:Place {id: "Amsterdam"}), (destination:Place {id: "London"}) 
CALL gds.shortestPath.dijkstra.stream({
  nodeProjection: 'Place',
  relationshipProjection: {
    ROAD: {
      type: 'EROAD',
      properties: 'distance',
      orientation: 'UNDIRECTED'
    }
  },

emaadmanzoor

[Unit]
Requires=zookeeper.service
After=zookeeper.service

[Service]
Type=simple
User=gb760
ExecStart=/bin/sh -c '/home/gb760/kafka/bin/kafka-server-start.sh /home/gb760/kafka/config/server.properties > /home/gb760/kafka/kafka.log 2>&1'
ExecStop=/home/gb760/kafka/bin/kafka-server-stop.sh
Restart=on-abnormal

emaadmanzoor

emaadmanzoor

95865 Recitation - Model Evaluation Demo

• Lecture slides: https://speakerdeck.com/emaadmanzoor/introduction-to-classification
• Course website: http://www.andrew.cmu.edu/user/georgech/95-865/
• Other recitations: http://www.eyeshalfclosed.com/teaching/

Contact:

• Emaad Ahmed Manzoor (emaad[at]cmu.edu)
• George Chen (georgechen[at]cmu.edu)

emaadmanzoor

#!/usr/bin/env python

# Copyright 2016 Emaad Ahmed Manzoor
# License: Apache License, Version 2.0
# http://www.eyeshalfclosed.com/blog/2016/07/22/spark-streaming-statistics/

"""
Get Spark Streaming microbatch statistics:
    - Batch start time
    - Scheduling delay (in seconds) for each microbatch

emaadmanzoor

Contents

1. README
   1. StreamSpot format
   2. CDM format
      1. CDM events
      2. CDM subjects, objects and principals
      3. CDM simple edges
   3. Combining CDM records into a StreamSpot edge
2. [Assigning a graph ID to each StreamSpot edge](https://gist.github.com/emaadmanzoor/b23d46c9d4d064746e1ef667cb8070bc#assi

emaadmanzoor

StreamSpot Bootstrap Clusters

www3.cs.stonybrook.edu/~emanzoor/streamspot/

Below are the bootstrap clusters used for the experiments in the StreamSpot paper for each of following datasets:

• all (01-C50_k10_all.txt): Chunk length of 50, 10 clusters.
• ydc (02-C25_k5_ydc.txt): Chunk length of 25, 5 clusters.
• gfc (03-C50_k5_gfc.txt): Chunk length of 50, 5 clusters.

emaadmanzoor

See the project website for more details.

Please report any issues to emaadahmed.manzoor@kaust.edu.sa.

Execution

Running this requires having the following files in the same directory as calculate_cluster_statistics.py:

• all_links.p
• all_tweets.p

emaadmanzoor

See the project website for more details.

Please report any issues to emaadahmed.manzoor@kaust.edu.sa.

Correlation Results

The attention potential (as estimated in section 4), when evaluated on this Twitter dataset:

• Is 73.61% correlated with the retweets obtained.
• Is significantly correlated (p < 0.05).