Avik Chaudhuri
@__avik
Followers
1K
Following
2K
Media
96
Statuses
1K
Creator of @flowtype. Hacker on @PyTorch compiler. ML explorer. Rusty PL researcher. Amateur chef. Soccer dad. Blogging at https://t.co/KRv2Q5Adsd
San Francisco Bay Area
Joined March 2014
Episode 1 of a new video series on #PyTorch Compiler topics is now out!. In future episodes I'll be hosting other people on the team to share tips and tricks that help you get the max out of torch.compile, torch.export, and related technologies, while enjoying a glimpse into all.
We're kicking off a new PyTorch Compiler video series 🎬 In the first episode, Avik Chaudhuri discusses a few simple principles that drive the programming model for export & illustrates them via a series of examples. 🔍 Export creates ahead-of-time IR representations of PyTorch.
1
1
9
Episode 6 of the #PyTorch Compiler series (playlist: . In which @Boyuan_Feng tells us all about FlexAttention! What a slick piece of tech. Fun little origin story: most of this work was done as part of a summer intern project at Meta! (@cHHillee was.
FlexAttention is a novel compiler-driven programming model that allows implementing the majority of attention variants in a few lines of idiomatic PyTorch code @Boyuan_Feng & @__avik show how many existing attention variants can be implemented via FlexAttention & that we achieve
0
0
2
This was a pretty cool paper to read when it came out:
Earlier this year, the computer scientist Binghui Peng and his team proved mathematically that there may be a hard limit to LLMs’ compositional task-solving abilities.
0
0
0
RT @leanprover: Incredibly grateful to @TheOfficialACM SIGPLAN for awarding #LeanLang the Programming Languages Software Award 2025 at #PLD….
0
38
0
RT @leanprover: @fede_intern Check out these excellent resources:. - Functional Programming in Lean: - Hitchhiker'….
0
2
0
Can’t think of a better example of how long it can take for a significant piece of technology done really well to truly take off at the right time. I’m not a big fan of “thought leadership.” To me Lean was probably driven less by a grand vision to justify its existence but.
🎂#LeanLang 0.1 was released 11 years ago today! . While it's true that Lean had been in development for almost a year prior (earliest commit July 15, 2013!) the release of Lean 0.1 was a major milestone. The screenshot below is from a June 26th, 2014 snapshot of the
1
0
7
Episode 5 of the #PyTorch Compiler Series (bookmark: . Guards are key to how torch.compile and torch.export guarantee soundness. (In torch.export, we focus on tensor-shape guards, but because torch.compile does not make any assumptions about code, it also.
Inside torch.compile Guards: How They Work, What They Cost, and Ways to Optimize with Animesh Jain and @__avik. Watch the next episode of the PyTorch Compiler Video Series:
0
0
1
RT @lexfridman: Here's my conversation with Terence Tao, one of the greatest mathematicians in history. We talk about the hardest problems….
0
736
0
@3blue1brown would love to get your eyes on this. In any case would love to see a video on this topic!.
0
0
0
Everything then comes together beautifully! . 1/ By Birkhoff-von Neumann, we decompose the transition matrix into a convex combination of permutation matrices. 2/ By symmetry, all permutations of a probability distribution have the same information entropy. 3/ By applying.
1
0
0
Our plan is to compute and compare the information entropy of an observable state before and after a transition. To do that, we use the notion of Kullback-Leibler divergence that measures how close a given probability distribution is to some baseline. By using the uniform
1
0
0
We further assume that the transition matrix is doubly stochastic. This assumption turns out to be a crucial physical constraint, ensuring time-reversibility of transitions between physical states. In particular, this means that by the Birkhoff-von Neumann theorem, we can
1
0
0
We model the dynamics of the system as a Markov process where observable states are probability distributions over physical states, and transitions between them are defined by some matrix of probabilities of transitioning from one physical state to another.
1
0
0
We begin by reviewing Shannon's definition of information entropy, and a quick-and-dirty way to "derive" it by counting the number of ways a physical system may be in a certain observable state. This is closely related to how Boltzmann defines thermodynamic entropy in their
1
0
0
How is entropy in #InformationTheory related to entropy in #ThermoDynamics? . In particular, can we prove that information entropy must increase in a simple, purely mathematical model of states and transitions, in a way that provides intuition on the second law of thermodynamics.
1
0
0