#544: Wheel Next + Packaging PEPs
Get the full intelligence
Search transcripts, export clips, track mentions, and explore all topics from “#544: Wheel Next + Packaging PEPs” inside PodZeus.
In this episode of Talk Python to Me, Michael Kennedy dives deep into the transformative 'Wheel Next' initiative—a collaborative effort by NVIDIA, QuantSight, Astral, and other major tech players to solve long-standing limitations in Python packaging. The core problem: current wheels are built for the lowest common denominator of hardware, typically CPU features from 2009, leaving massive performance gains on the table for modern CPUs (like AVX2) and GPUs. This results in bloated, slow-to-install binaries and complex, user-hostile installation processes—especially for data science and machine learning libraries like PyTorch. The solution lies in a new set of PEPs (817 and 825) that introduce 'wheel variants,' allowing packages to declare hardware-specific requirements (e.g., CUDA version, CPU instruction sets) in metadata. Installers like UV can then automatically select the optimal build for the user’s machine, eliminating the need for manual index configuration and enabling smaller, faster wheels. The episode features interviews with key contributors Jonathan Decker (NVIDIA), Ralph Gommers (QuantSight, NumPy/SciPy), and Charlie Marsh (Astral, UV), who discuss the technical challenges, the massive cross-industry collaboration, the year-long prototyping process, and the realistic timeline for adoption—likely 1-2 years for full ecosystem integration. The conversation also touches on related projects like the PyPackaging Native Guide and the experimental PYX registry, which is already offering variant-enabled builds as a stopgap. The episode concludes with a call to action: developers and maintainers are encouraged to engage with the community on Discourse, try the experimental variant-enabled UV build, and contribute to shaping the future of Python packaging. The overarching message is one of optimism and pragmatism: while the full rollout will take time, the foundational work is solid, and early adopters—especially in the data science and ML space—stand to gain immediate benefits. The long-term vision is a seamless, high-performance Python ecosystem where users don’t need to worry about hardware compatibility; the tools handle it automatically.
Wheel Next introduces PEPs that allow packages to declare hardware-specific requirements (like CUDA version or CPU instruction sets), enabling installers to automatically pick the optimal build.
This solves the 'fat binary' problem, reducing wheel sizes from ~900MB to ~250MB and dramatically improving install speed and bandwidth efficiency.
The initiative is a massive cross-industry collaboration involving NVIDIA, Meta, Intel, AMD, Red Hat, and open source maintainers, making it one of the most diverse Python packaging efforts ever.
The core innovation is 'wheel variants'—a standardized way to express build differences in metadata, allowing tools like UV to resolve them automatically without user intervention.
Adoption is expected to start with major data science libraries (like PyTorch, JAX, VLLM), which will trigger a rapid cascade of adoption across the ecosystem.
…and 2 more takeaways available in PodZeus
The Problem: Fat Binaries and the 2009 CPU Limit
“When you pip install a package with compiled code, the wheel you get is built for CPU features from 2009. Want newer optimizations like AVX2? Your installer has no way to ask for them.”
Introducing Wheel Next: A Coalition for Change
The episode introduces the 'Wheel Next' initiative, a collaborative effort by NVIDIA, QuantSight, and Astral to create a new set of PEPs that allow packages to declare hardware requirements. This enables installers like UV to automatically select the right build.
The Technical Challenge: Why Current Packaging Falls Short
Jonathan Decker and Ralph Gommers explain the limitations of current platform tags, which only cover OS, architecture, and Python version, but not CPU instruction sets or GPU capabilities. This forces maintainers to create 'fat binaries' that work everywhere but are huge and slow.
The Solution: Wheel Variants and Metadata
“It's not going to be this year. There's a very long tale of how the implementation rolls through the ecosystem, and then you have to wait until users get newer tools, and only then can you start uploading wheels.”
Real-World Impact: Performance and Size Gains
“If we have variants, we can just slim it down to one CUDA architecture per wheel so you can go down to like 200 megabytes or so, 250 maybe but it's way better for both for index servers, it's better for users.”
“When you pip install a package with compiled code, the wheel you get is built for CPU features from 2009. Want newer optimizations like AVX2? Your installer has no way to ask for them.”
“A problem well stated is a problem half solved. So this is exactly what we are trying to say.”
“It's not going to be this year. There's a very long tale of how the implementation rolls through the ecosystem, and then you have to wait until users get newer tools, and only then can you start uploading wheels.”
Host
Guests
PyTorch
product
Charlie Marsh
person
Jonathan Decker
person
UV
product
Ralph Gommers
person
CUDA
other
NVIDIA
organization
Astral
organization
NumPy
product
QuantSight
organization
#543: Deep Agents: LangChain's SDK for Agents That Plan and Delegate
Talk Python To Me • 1h 3m • 4/1/2026
#545: OWASP Top 10 (2025 List) for Python Devs
Talk Python To Me • 1h 6m • 4/16/2026
#546: Self hosting apps for Python people
Talk Python To Me • 1h 3m • 4/27/2026
#547: Parallel Python at Anyscale with Ray
Talk Python To Me • 59m • 5/6/2026
#548: Event Sourcing Design Pattern
Talk Python To Me • 1h 8m • 5/11/2026
Get the full intelligence
Search transcripts, export clips, track mentions, and explore all topics from “#544: Wheel Next + Packaging PEPs” inside PodZeus.
Start discovering podcast insights today
Start with a 7-day trial and explore a growing catalog of popular podcasts. No credit card required.
No credit card required • 7-day trial • Cancel anytime
