@Is36E
Joined May 2017
  • Tweets623
  • Following88
  • Followers215
68Media
Lots of exciting changes coming to MPS in @PyTorch 2.15 release: - Faster LLM decode - FP8 dtype on MPS - Better Linear Algebra support - Less memory usage - And plans for 2.16! Take a look at the blog post for more details! huggingface.co/blog/mps-pyto…
🤖 Made with AI
1
4
2
13
498
MPS memory usage has always haunted users running variable-length workloads. This PR brings CUDA-style splitting/coalescing to MPS which leads 5 times less memory usage!
1
2
17
1,338
Apple just announced lower bit dtypes(FP8/FP4/2-bit Int) support coming to macOS 27 on WWDC! developer.apple.com/videos/p…
2
10
95
6,177
Flex attention is now available on @PyTorch MPS nightlies! github.com/pytorch/pytorch/p…
6
1
30
2,145
🚀 open sourced metalBLAS, hand-tuned Metal matmul kernels for Apple Silicon, callable from PyTorch on mps. Matches/beats MPS Graph (torch) matmuls on bf16/fp16, 2-3x faster on fp32 (TF32-relaxed) across the bench suite on M5 Pro. Next step is to upstream this to PyTorch! github.com/Isalia20/metalBLA…
1
12
1
78
9,039
Shipped specialized SDPA kernels for PyTorch MPS, up to 16x faster than the previous MPSGraph path 🚀 Metal kernels for both decode (q_len=1) and prefill (long causal) - Decode, 16k ctx, D=128: **1.42 → 0.087 ms (16.3x) - Prefill, 4k seq, D=96: **99.6 → 18.8 ms (5.3x)
2
6
2
80
5,364
Pesky bug killing performance on @PyTorch MPS. Can you spot it?
2
255
Hopefully soon distributed on @PyTorch MPS will come to life
1
166
They did it again. I am once again asking you report pytorch bugs instead of just vagueposting about them @tenderizzation
Lucas Beyer (bl16)
@giffmana6 Nov 2025
Hello? PyTorch? Yes, please, I would like to have under 9000% sparsity. Thank you! And a coke with that, please.
2
22
5,500
🎉 Congrats to the 2025 #PyTorch Contributor Award winners & nominees! 🏅 Winners: Vincent Moens, Mingfei Ma, Eikan Wang, Irakli Salia, Jiawei Li, Aaron Gokaslan, Justin Silver, Jonathan Dekhtiar, Xuehai Pan, Zingo Andersen, Zesheng Zong. Nominees: Abdulsalam Yazid Bande, ghostspiders, Muhammed Tolga Cangöz, Yu Guangye, Masahiro Hiramori, Xu Han, Kaichao You, Simon Mo, Ralf Gommers, Logan Thomas, Sayak Paul, Hussain. 🔗 Read the announcement for more: hubs.la/Q03R8CTf0 #PyTorchFoundation #PyTorch #OpenSourceAI
2
6
44
11,208
another @PyTorch PR reverted
1
216
Finally escaped the permanent underclass
197
Trying to find where the kernel is being dispatched from in torch
117
Get different metrics when using multi gpu training. I'm so cooked
1
127
Second moment in RMSProp refers to just an average of squared grads? Holy shit people like to complicate words
422
🧵 (1/5) How to bind CUDA kernels to PyTorch Got ultra-fast CUDA kernels but stuck on using them? This thread shows how to make them work with torch. Perfect for when you need that custom op that torch can't handle natively.
1
2
2
789
more replies
🧵 (4/5) For the final step you need to define your CUDAExtension with the utilities torch provides for easily integrating the kernel with the torch tensors. "custom_vector_multiplicator" will be the "library" name here
1
87
🧵 (5/5) After you have done each above step simply call: python setup.py install This will install the extension in your virtual environment and then you will be able to import your extension and use it with torch tensors as if it was defined already.
66
Load more