PyData posts talk on d9d — an open framework for distributed LLM training

🛠 PyData on YouTube: talk on d9d — a framework for distributed LLM training

PyData's channel posted a recording of a 50-minute talk by Maxim Afanasyev from Tochka on August 23 from PyData & PyCon Yerevan 2026: how d9d works — an open-source Apache-2.0 framework for distributed training, built natively on PyTorch 2.0 and tested when scaling MoE models to 235 billion parameters.

🌍 d9d occupies a niche between monoliths like Megatron-LM and building your own infrastructure: native PyTorch 2.0 primitives, scaling from a single GPU to 6D parallelism, streaming checkpointing, and pipeline parallelism via torch.distributed.pipelining.

👤 The recording is a practical breakdown of distributed LLM training: from 6D parallelism to specific APIs. The code is open (Apache-2.0): you can clone it, read it, and borrow engineering techniques.

Source 1: https://www.youtube.com/watch?v=BGTtk3geHoY Source 2: https://pretalx.pycon.am/pydata-pycon-2026/talk/E7DZBU/