https://guilhermeleobas.github.io/categories/pytorch/ Recent content in PyTorch on Hugo en-us Tue, 30 Jun 2026 00:00:00 +0000 https://guilhermeleobas.github.io/blog/tp_as_number-slot/ Tue, 30 Jun 2026 00:00:00 +0000 https://guilhermeleobas.github.io/blog/tp_as_number-slot/ <p><strong>Note</strong>: This article originally appeared in <a href="https://openteams.com/tp-as-number-slot-and-binary-dispatch/">OpenTeams blog</a>.</p> <p>Recently Animesh Jain from Meta published a <a href="https://docs.pytorch.org/devlogs/dynamo/2026-05-13-agent-friendly-dynamo/">blogpost</a> about the work we are doing in Dynamo. The gist of it is that we are refactoring Dynamo&rsquo;s object model to mirror CPython <a href="https://docs.python.org/3/c-api/typeobj.html#tp-slots">PyTypeObject slots</a>. Given that, I thought it would be interesting to write a follow-up post about one small corner of that work: the <code>tp_as_number</code> slot, and how binary operations are dispatched in CPython.</p> <h2 id="background">Background</h2> <h3 id="torchdynamo">TorchDynamo</h3> <p>TorchDynamo (or simply Dynamo) is the JIT compiler that PyTorch uses to make PyTorch programs faster. Dynamo hooks into the eval frame API (<a href="https://peps.python.org/pep-0523/">PEP 523</a>) to intercept the Python bytecode before it runs and symbolically trace it to an intermediate representation called <em>FX Graph</em>. It is worth mentioning that Dynamo is not a general-purpose Python implementation. It is designed to optimize a very specific class of programs and it works quite well for that purpose.</p>