<?xml version="1.0" encoding="utf-8" standalone="yes"?><rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom" xmlns:content="http://purl.org/rss/1.0/modules/content/"><channel><title>Heterogeneous Devices on Murtaza Rangwala</title><link>https://murtaza-hatim.com/tags/heterogeneous-devices/</link><description>Recent content in Heterogeneous Devices on Murtaza Rangwala</description><generator>Hugo -- 0.147.2</generator><language>en</language><lastBuildDate>Mon, 07 Sep 2026 00:00:00 +0000</lastBuildDate><atom:link href="https://murtaza-hatim.com/tags/heterogeneous-devices/index.xml" rel="self" type="application/rss+xml"/><item><title>Kafila: Serving Large Language Models on Heterogeneous Consumer Devices</title><link>https://murtaza-hatim.com/publications/kafila/</link><pubDate>Mon, 07 Sep 2026 00:00:00 +0000</pubDate><guid>https://murtaza-hatim.com/publications/kafila/</guid><description>Kafila serves LLMs on a small, trusted group of heterogeneous consumer devices by planning block placement and NAT-traversing ring order from each device&amp;#39;s measured bandwidth, capacity, and reachability. This cuts the pipeline&amp;#39;s slowest stage by 4.2x versus uniform division and enables serving models uniform division can&amp;#39;t fit at all.</description></item></channel></rss>