the bug that taught me more about PyTorch than years of using it

The blog post recounts how a silent PyTorch MPS backend bug caused training loss plateaus by failing to update model weights on non-contiguous tensors, leading the author through a deep debugging journey that revealed PyTorch internals and the complexity of device-specific GPU kernel implementations. This experience taught more about PyTorch's workings than years of usage, emphasizing th...

https://elanapearl.github.io/blog/2025/the-bug-that-taught-me-pytorch

IsaacSim

NVIDIA Isaac Sim™ is an open-source application on NVIDIA Omniverse for developing, simulating, and testing AI-driven robots in realistic virtual environments.

https://github.com/isaac-sim/IsaacSim

Codetracer - a time-travelling debugger for Python

Codetracer is a new way to debug and code-review Python. Demo:
https://youtu.be/e2ctwMbpuCA

https://github.com/metacraft-labs/codetracer

codelion / openevolve

Open-source implementation of AlphaEvolve

https://github.com/codelion/openevolve

AI-Trader

Five AIs battle for NASDAQ 100 supremacy. Zero human input. Pure competition.

https://github.com/HKUDS/AI-Trader

torchforge

PyTorch-native post-training at scale.

https://github.com/meta-pytorch/torchforge

Introducing FlashPack: Lightning-Fast Model Loading for PyTorch

The FlashPack package dramatically speeds up PyTorch model loading by flattening all weights into a single contiguous stream, memory-mapping the file, and overlapping disk, CPU, and GPU operations with CUDA streams. This approach yields 3-6× faster loading compared to traditional methods like loadstatedict(), reducing GPU idle time and improving overall performance, especially on syste...

https://blog.fal.ai/introducing-flashpack-lightning-fast-model-loading-for-pytorch

The Building Blocks of Agentic AI: From Kernels to Clusters

The PyTorch Native Agentic Stack is a scalable, PyTorch-integrated framework designed for building and deploying autonomous AI agents across thousands of GPUs. It simplifies complex distributed reinforcement learning workflows by orchestrating large-scale models, providing abstractions for services, fault tolerance, and efficient state management to accelerate AI research and deployment.

https://ai.meta.com/blog/introducing-pytorch-native-agentic-stack

Nyno

Create and Run Linux Workflows without Limits. Short for "nine" / "yaml" / "no-code" / "automation".

https://github.com/empowerd-cms/nyno

Detecting object wrappers

The wrapt library's version 2.0.0 changed its object proxy class hierarchy, breaking checks that detect if an object is already wrapped, causing repeated wrapping and performance issues. The best practice is to use custom wrapper types and traverse wrapper chains via the wrappedattribute to reliably detect wrapping, emphasizing careful version pinning and cautious monkey patching in ...

https://grahamdumpleton.me/posts/2025/10/detecting-object-wrappers/

pip 25.3

The pip 25.3 release is the final major pip update of 2025, featuring the removal of non-PEP 517 package build support and non-PEP 660 editable installs, meaning pip no longer calls legacy setup.py commands. It adds the new --build-constraint option for specifying build-time constraints separately, improves caching and metadata handling, supports editable requirements as Direct URLs, and...

https://discuss.python.org/t/announcement-pip-25-3-release/104550

Kanchi

A realtime celery task monitoring and management system.

https://github.com/getkanchi/kanchi

EuroPython 2025 Videos

Here are all the videos for the conference, brought to you by the EuroPython 2025 Team and the EuroPython Society.

https://www.youtube.com/playlist?list=PL8uoeex94UhFQY9cYBQOVkj9fSHMHf5x9

Helion: A High-Level DSL for Performant and Portable ML Kernels

Helion is a Python-embedded high-level DSL that compiles to optimized Triton kernels, blending the simplicity of PyTorch syntax with powerful autotuning to deliver high-performance, portable machine learning kernels across hardware architectures. It automates complex tasks like tensor indexing, memory management, and hardware-specific tuning, enabling developers to write efficient kernel...

https://pytorch.org/blog/helion/

The Best Way to Share Code Between Python Apps

The video explains how to use UV workspaces to manage multiple Python applications with shared code and dependencies in a single repository. It demonstrates structuring a project with CLI and FastAPI apps, extracting shared logic into internal packages, and managing dependencies efficiently to avoid duplication and conflicting environments. The approach simplifies development and scaling...

https://www.youtube.com/watch?v=N_ypJwV8Q8I