If GPU optimization and systems problems excite you, why limit your impact to a single company or lab?
Working on PyTorch allows you to ship impact to the entire AI industry!
We're hiring across experiences — junior and senior engineers. Read more below
@soumithchintala
-
PyTorch Hiring: GPU Optimization Engineers Wanted
By
–
-
PyTorch Seeks Skilled Systems Engineers for Growth
By
–
https://
marksaroufim.com/2025/04/16/pyt
orch-needs-great-systems-engineers.html
… https://
workwithpytorchmeta.com -
PyTorch F.pad Tensor Padding API Design Discussion
By
–
lol, I think I helped design it, and I think it looks clowntown from the lens of Tensor padding. F.pad was written as a consolidation of all the vision padding — for padding images, videos, etc. It consolidated nn.ReflectionPad2d, etc. from torch7. But looking at it from the
-
FFmpeg Wrappers Often Hide Correctness Bugs
By
–
you'd be surprised at how many ffmpeg wrappers have correctness bugs 🙂
-
FFmpeg I-frame Approximate Mode for Faster ML Training
By
–
oh yea, so the last thing about i-frames, ffmpeg has an approximate mode which will fast-seek to whatever is the closest iframe and just return that. it works for some training workloads and is much faster. the approximate mode is exposed in torchcodec.
-
Efficient FFmpeg Wrapper for PyTorch Video Processing
By
–
it's a "proper" efficient wrapper around ffmpeg for PyTorch, using some of ffmpeg's fast seeking and read-ahead APIs correctly, as well as using memory buffers correctly, avoiding allocs and copies.
-
Efficient FFmpeg wrapper for PyTorch video processing
By
–
it's a "proper" efficient wrapper around ffmpeg for PyTorch, using some of ffmpeg's fast seeking and read-ahead APIs correctly, as well as using memory buffers correctly, avoiding allocs and copies.
-
PyTorch Releases torchcodec: Fast Open-Source Video Decoding Library
By
–
A few months ago we quietly open-sourced a @PyTorch video decoding library called torchcodec — small, nimble, fast.
The @LeRobotHF folks had some nice things to say about it. -
Julia JIT limitations for GPU accelerator support
By
–
Julia's JIT works really well for CPU overheads and CPU parallelism to an extent, but doesn't work as out-of-the-box for accelerators (like GPUs).