[lucialuka]

Project

OpenAssembler

Checklist

• I read CONTRIBUTING.md - awesome-python is a shortlist, not a catalog
• One project per PR
• PR title format: Add project-name
• Entry format: - [pypi-name](https://github.com/owner/repo) - Description ending with period.
• Display name is the PyPI package name
• Placed in an existing use case (new sections and subcategories are maintainer-only)
• Meets all Quality Requirements: active, stable, documented, at least 1 month old

Which Tier

Pick one:

• Obvious choice - a tool an experienced Python developer would name when asked "what do I use for this?"
• Challenger - not yet the obvious choice, but a credible successor to one. Give adoption-trajectory evidence, not popularity alone.

Explain:
While ptxas (an important nvidia compiler that compile ptx to assembler language) is closed source and OpenAI Triton reach merely ~90% of the speed comparing to native nvidia toolset, we created OpenAssembler, an open source compiler that can compile your whole ML repo (constitute of Python, CUDA C/C++ and even custom ptx) into nvidia hardware executable cuda binary. It achieve 3-10% speedup comparing to OpenAI Triton for ML training/inferencing, still do not completely exceed the closed source nvidia native toolset. We expose the complete compiler details through reverse engineering ptxas (closed source by nvidia) with clear documentations in Reverse Engineering Results , so if you want to customize your kernel with a higher training speed comparing to OpenAI Triton, you can also do it with this.

Displacement

N/A

[putriazuraananda25-sudo]

#3406