Pintos HTML

A web-based interface for Pintos, the educational operating system from UC Berkeley's CS 162. Boots a real Pintos kernel in your browser using the v86 JavaScript x86 emulator.

Now featuring Llama 2 (15M) LLM inference running directly on Pintos, ported from karpathy/llama2.c.

Live Demo

https://www.ocf.berkeley.edu/~briancpark/pintos/

Features

• Boots the full Pintos kernel (VM project) in-browser via WebAssembly x86 emulation
• LLM inference: Run Llama 2 (stories15M) to generate text, entirely within the Pintos OS
• Interactive shell with user programs: ls, cat, echo, cp, mkdir, rm, and more
• PC speaker audio emulation: Pokemon Red/Blue theme (contest), I have Been Working on My Pintos (railroad162 "I've Been Working on My Pintos")
• Serial output rendering for clean terminal display
• PS/2 keyboard input via scancode translation
• Mobile-friendly with virtual keyboard support

Running Llama 2 on Pintos

Once the shell boots, run:

llama2 model -i "Once upon a time" -n 64

This loads the 15M parameter Llama 2 model (split into 8MB chunks to fit the Pintos filesystem limit) and generates text token by token. The model weights (~60MB) are baked into the disk image and loaded at runtime.

Options:

• -i <prompt> - input prompt
• -n <steps> - number of tokens to generate (default: 256)
• -t <temp> - temperature (default: 1.0, 0.0 = greedy)
• -s <seed> - random seed (default: 42)

Setup

Prerequisites

• A built Pintos source tree (with llama2 in examples)
• GCC with 32-bit support (gcc -m32)
• QEMU (qemu-system-i386)
• Perl, wget

Build & Deploy

# 1. Build the Pintos kernel and create a bootable disk image (~80MB with LLM weights)
./build.sh ~/pintos/src

# 2. Download v86 dependencies and deploy to a web directory
./deploy.sh ~/public_html/pintos

How It Works

1. build.sh compiles the Pintos VM project and runs bootable-cs162.sh to create a disk image with the kernel, shell, user programs, and LLM model weights
2. deploy.sh downloads the v86 emulator (JS + WASM + BIOS files) and copies the HTML + disk image to the deployment directory
3. index.html loads v86 with 128MB RAM, boots the disk image, captures serial output for display, and translates browser keyboard events to PS/2 scancodes
4. llama2 loads split model weights (model.00-model.07) and tokenizer.bin from the Pintos filesystem and runs transformer inference using custom math routines (no math.h in Pintos)

Available Shell Commands

Command	Description
llama2 model [opts]	Run Llama 2 (15M) text generation
ls	List files
cat <file>	Print file contents
echo <text>	Echo text
cp <src> <dst>	Copy a file
mkdir <dir>	Create a directory
rm <file>	Remove a file
hex-dump <file>	Hex dump a file
contest	Play Pokemon Red/Blue title screen theme
railroad162	Play "I've Been Working on My Pintos"
halt	Shut down Pintos

Created by Brian Park