Saar Eliad
bertashusbend.bsky.social
Saar Eliad
@bertashusbend.bsky.social
Researcher & Tech Lead, AI inference systems
A nice read this morning: megaparse code.
What this trending repo does eventually, is parsing documents by converting them to images and sending them to a vision llm.
(Similar to the ColPali technique)
github.com/quivrhq/mega...
MegaParse/libs/megaparse/src/megaparse/parser/megaparse_vision.py at main · QuivrHQ/MegaParse
File Parser optimised for LLM Ingestion with no loss 🧠 Parse PDFs, Docx, PPTx in a format that is ideal for LLMs. - QuivrHQ/MegaParse
github.com
December 6, 2024 at 5:31 AM
Reposted by Saar Eliad
Hi, just sharing an updated version of the PyTorch 2 Internals slides: drive.google.com/file/d/18YZV.... Content: basics, jit, dynamo, Inductor, export path and executorch. This is focused on internals so you will need a bit of C/C++. I show how you can export and run a model on a Pixel Watch too.
November 19, 2024 at 11:05 AM
Switched lately to uv & ruff. Nice work, astral-sh. A satisfying feeling to see such speedups, and it also comes with a better user experience in the vscode addon (e.g. auto fixes.)
December 1, 2024 at 12:37 PM
As we all migrate to bluesky, I'll share that I took some time lately to migrate an (LLM related) internal Excel workbook to a deployed streamlit app.

UI is often neglected for technical research, engineering, or architecture activities. Yet including even a minimalistic UI turned to be beneficial.
November 30, 2024 at 7:47 AM