Birchlabs

pytorch 2.7 is out!
- Mega Cache looks nice, unclear whether those modules get cached by legacy mechanisms too
- foreach map looks good for optimizers
- trainable biases means you can now train T5 on flex
- prologue fusion is hype
- context parallel brings ring attention
pytorch.org/blog/pytorch...

pytorch 2.6 is out!
highlights:
- flex attention: better compilation of blockmask creation, better support for dynamic shapes
- cuDNN SDPA: fixes for memory layout
- CUDA 12.6
- python 3.13
- MaskedTensor memory leak fix

yeah but how do I type two hyphens

fish, dairy, beyond meat burgers

pytorch 2.6 final RC is out, promoting to stable in a couple of days!
mostly I'm looking forward to better compilation of flex block mask creation, and better support for flex attention on dynamic shapes.
there's also fixes for memory layout in cuDNN SDPA.
dev-discuss.pytorch.org/t/pytorch-re...

Claude does SVG memes

drink cups should put the hole in the bottom.
heat rises. "the top is cool enough to drink" implies "everything below it is colder".
drinking from the bottom lets us access safe temperatures earlier and before the whole cup cools.

running npm version from a subdirectory of a git repository is literally an unsolved problem in 2024
github.com/npm/cli/issu...

I should just get this tattooed, I never remember how to find it

pip install huggingface_hub[hf_transfer]
HF_HUB_ENABLE_HF_TRANSFER=1 huggingface-cli download

when the standard library comments out std::experimental::observer_ptr just to stop you having fun

NovelAI v4 makes dreams come true

when you're measuring torch compile warmup "oh 11 secs that's not so bad"
then you realize it was 111 secs

Claude's alright

if you care about multiprocess debugging in VSCode please upvote this issue
so we don't have to click terminate a hundred times
github.com/microsoft/vs...

GitHub - facebookresearch/flow_matching: A PyTorch library for implementing flow matching algorithms, featuring continuous and discrete flow matching implementations. It includes practical examples fo... A PyTorch library for implementing flow matching algorithms, featuring continuous and discrete flow matching implementations. It includes practical examples for both text and image modalities. - fa...

Meta releases flow-matching code
github.com/facebookrese...

GitHub - NVlabs/edm2: EDM2 and Autoguidance -- Official PyTorch implementation EDM2 and Autoguidance -- Official PyTorch implementation - NVlabs/edm2

the EDM2 repository had an Autoguidance update 4 days ago demonstrating their NeurIPS Oral paper "Guiding a Diffusion Model with a Bad Version of Itself"
github.com/NVlabs/edm2

torch profiler record_function spans are not free, even when you're not profiling a model. the model I'm benchmarking trained 2.5% faster when I commented them all out.

I thought this was about Box2D v3

it's here
main thing it's missing is bindings. lots of bindings
and I don't know how to bind to raw pointers and callbacks (it's not working)
github.com/Birch-san/bo...

Box2D 3.0.0 in WebAssembly+TypeScript starting to work

🤔 Why do we extract diffusion features from noisy images? Isn’t that destroying information?

Yes, it is - but we found a way to do better. 🚀

Here’s how we unlock better features, no noise, no hassle.

📝 Project Page: compvis.github.io/cleandift
💻 Code: github.com/CompVis/clea...

🧵👇

I worked on Angular apps where people would type their arrays as `any`.
not even `any[]`, just `any`.
not a shred of pride

torch.compile is hard for dynamic shapes / large number of static shapes, and non-transformer architectures.
I measure suites of shapes, log recompiles, check which require warmup.
operation compile competitive with whole-model compile.
some operations prefer compiler disabled.
dynamic often slow.

“that makes two of us!”

come to ML, where there are no tests, and dependencies are the wild west

A common question nowadays: Which is better, diffusion or flow matching? 🤔

Our answer: They’re two sides of the same coin. We wrote a blog post to show how diffusion models and Gaussian flow matching are equivalent. That’s great: It means you can use them interchangeably.

incidentally the inductor problem is here.
it could be fixed by omitting the guard altogether, or perhaps eliding the function grid wrapper entirely
github.com/pytorch/pyto...

if you don't do this, then inductor codegen emits invalid python