omkaar
@omkizzy
Followers
4K
Following
7K
Media
196
Statuses
3K
3e-4 | ex- avg lazeez enjoyer @uwaterloo
Joined July 2015
I built software to cut docker image sizes by more than 80%. It uses dynamic analysis and and has a LOT more room for improvement. The demo here shows an ETL workflow docker image, cut down from 1.28GB to 193MB (-85%)! If large docker image sizes are becoming a production
75
238
3K
very useful!
I built https://t.co/VJ7oHdvf7s, a new interface for arXiv As we enter an era of accelerated scientific discovery, we need better tools that augment human cognition to help us keep up. Try it: visit papiers ai or swap arxiv -> papiers on any paper URL
1
1
1
wait andrew tulloch is back at meta?! @tbpn, insane news for meta
Leaving Meta and PyTorch I'm stepping down from PyTorch and leaving Meta on November 17th. tl;dr: Didn't want to be doing PyTorch forever, seemed like the perfect time to transition right after I got back from a long leave and the project built itself around me. Eleven years
0
0
0
an agent's function calls are like it's ISA, there's so much you can do with this knowledge of this analogy
0
0
1
the lord assigned use cases for both bf16 and fp16. we were just oblivious...
can't believe that in the year of our lord 2025 I'm about to try training in fp16
0
0
1
docker-trimmer should work well on pytorch images. ops are device + dtype specefic, lots of overhead for a given build.
0
0
0
india needs so much more of this, i'm glad people like Hardeep exist
I moved from San Francisco to Mumbai to host India's Biggest AI Film Festival And I have never seen such hypocrisy Bollywood "A-listers" don't want to associate with AI because of public opinion, but then ask us if they could get intros to our AI filmmakers to hire them Many
0
0
2
common waterloo w
i've been working on a tool for running background coding agents inside of their own workspace VMs. introducing https://t.co/y7xdMxEhSo :)
1
0
8
another trick spawned by this is being able to tell if the magnitude of ANY float is > 1 from binary. if the "sign bit" or left-most bit of the exponent portion is 1, its magnitude > 1, else it's between 0 and 1.
a lot of people don't understand floating point numbers. i added this diagram to an old article I wrote to turn any fp repr to decimal. helps me intuit through fp math.
1
0
4
for example, deepseek uses a custom E5M6 fp type. with the formula below, its easy to make quick calculations. e.g. the max value is 1 x 2^15 x (1 + 0.984375) = 65,024.
a lot of people don't understand floating point numbers. i added this diagram to an old article I wrote to turn any fp repr to decimal. helps me intuit through fp math.
0
0
5
triton is here to stay => swarth of papers and repos (like TLX) extending it's ecosystem for different applications. also new DSLs like helion are getting built on top of triton
0
0
1
there's so much to learn from this PR, sick contrib
as of 5 minutes ago, our gpt-oss implementation is merged into torchtitan! thanks to all the work by @jianiw_wang @__tianyu at @pytorch making it clean & scalable for the community ❤️ i hope y'all play around with training gpt-oss, it's great for its sparsity & reasoning
0
0
4
a lot of people don't understand floating point numbers. i added this diagram to an old article I wrote to turn any fp repr to decimal. helps me intuit through fp math.
1
0
7
Alibaba used FPGAs to build a performant and stable storage engine in 2018... you really can just do things
0
0
5
the ml compiler frontend framework boom has parallels to the web frontend framework boom of the 2010s
0
0
2