P.M.SALMAN KHAN SALMANKHANPM

Compilers for Machine Learning

A hands-on one semester course where students build their own compiler from scratch, starting from elementwise programs and ending with training SOTA LLMs on GPUs. This course aggressively builds on the previous week, and is an exercise in slop management. If you let any slop in early, it will compound and you will not finish the class.

Course description: This course covers the design and implementation of a modern machine learning compiler, and examines the interaction between IR design, hardware capabilities, and the structure of machine learning programs. Topics covered include term rewriting, code generation, movement operators, kernel fusion, memory hierarchies, GPU architecture, automatic differentiation, and flash attention. It is a project course, providing experience with performance-oriented programming, managing a codebase that grows all semester, and working in 1 or 2 person teams, culminating in a compiler capable of training modern LLMs.

Prerequisites: This c

LLM Wiki

A pattern for building personal knowledge bases using LLMs.

This is an idea file, it is designed to be copy pasted to your own LLM Agent (e.g. OpenAI Codex, Claude Code, OpenCode / Pi, or etc.). Its goal is to communicate the high level idea, but your agent will build out the specifics in collaboration with you.

The core idea

Most people's experience with LLMs and documents looks like RAG: you upload a collection of files, the LLM retrieves relevant chunks at query time, and generates an answer. This works, but the LLM is rediscovering knowledge from scratch on every question. There's no accumulation. Ask a subtle question that requires synthesizing five documents, and the LLM has to find and piece together the relevant fragments every time. Nothing is built up. NotebookLM, ChatGPT file uploads, and most RAG systems work this way.

Oh my zsh Setup (Fast + Minimal)

Install Zsh

Ubuntu

sudo apt update
sudo apt install zsh curl git

Note

I’ve written a newer optimized 2026 version of this guide with a faster setup using Zinit + Starship and improved plugin loading.

This version avoids slow shell startups and works better with modern Zsh setups.

Updated guide: https://gist.github.com/n1snt/2cccc8aa5f7b645a7628d3512c70deb6

Oh my zsh.

Generate framework compatibility matrix from fqnames

Download these 2 files
Compile AOSP without fcm from stock and wait for check_vintf to error out
Delete Python prefix from all lines (e.g. checkvintf E 06-24 00:30:22 49120 49120 check_vintf.cpp:554])
Paste the result in fqnames.txt
Launch the script

Migrated with additional information to my blog: https://msfjarvis.dev/posts/understanding-and-resolving-selinux-denials-on-android/

--------------

Denial in question

avc: denied { read write } for pid=29059 comm="i.tetherservice" name="ipa" dev="tmpfs" ino=11991 scontext=u:r:system_app:s0 tcontext=u:object_r:ipa_dev:s0 tclass=chr_file permissive=0

sepolicy fix

	MODEL_TO_USE = "NanoV2" # or "MistralNemo"
	IMPORT_MODEL_AND_DATA = False # set to True to import the model from HF hub, only needs to be run once
	REASONING_STYLE = "gpt_oss_120b" # or DeepSeek_R1_0528
	NUM_NODES = 2
	GPUS_PER_NODE = 8

	import nemo_run as run
	from nemo.collections import llm
	from nemo.collections.llm.gpt.model.mistral import MistralModel, MistralNeMoConfig12B
	from nemo.collections.llm.gpt.model.ssm import MambaModel, NemotronNano12Bv2

	RAR registration data
	WinRAR
	Unlimited Company License
	UID=4b914fb772c8376bf571
	6412212250f5711ad072cf351cfa39e2851192daf8a362681bbb1d
	cd48da1d14d995f0bbf960fce6cb5ffde62890079861be57638717
	7131ced835ed65cc743d9777f2ea71a8e32c7e593cf66794343565
	b41bcf56929486b8bcdac33d50ecf773996052598f1f556defffbd
	982fbe71e93df6b6346c37a3890f3c7edc65d7f5455470d13d1190
	6e6fb824bcf25f155547b5fc41901ad58c0992f570be1cf5608ba9