Anxhelo Xhebraj (@0xA95) profile picture
Joined October 2015
Photo by Anxhelo Xhebraj (@0xA95)Photo by Anxhelo Xhebraj (@0xA95)Photo by Anxhelo Xhebraj (@0xA95)Photo by Anxhelo Xhebraj (@0xA95)
Anxhelo Xhebraj retweeted
When Carmack says “functions that return with the stack extended,” that’s precisely what we did in “What If We Don’t Pop the Stack?” (ECOOP’22). Return a stack-allocated value → leave the frame. Reclaim only when a 1st-class value comes back. 10–25% faster in our benchmarks from less heap traffic + better locality. drops.dagstuhl.de/entities/d…
I always thought it was a shame that modest default stack setups made alloca() a generally-frowned-upon technique in C/C++. With 64 bit architectures, give every stack a virtual address space equal to the physical memory by default! You can do a surprisingly good job of memory management with just the stack, and I contend that it has both correctness and performance advantages. Allocations are a trivial pointer adjustment, and frees are literally free. Keeping everything linearly packed is friendlier for all layers of the cache hierarchy, and it makes keeping everything in huge / gigantic pages easy. Leaving temporary intermediate blocks on the stack can seem wasteful, but unless you are dealing with algorithms that make > O(n) temporary allocations, it is usually not a big deal. Static analysis can check if you store a stack pointer into a structure deeper in the stack, which is a use-after-free, and then memory management correctness is guaranteed. The discipline of caller-allocates-return is awkward in some dynamic cases. It would take a change in low level calling conventions, but you could plausibly have functions that return with the stack extended from where it was when they were called. I’m not holding my breath for that. It isn’t obvious if having AI coding agents makes discussions of software architecture more or less important!
3
10
104
6,662
Anxhelo Xhebraj retweeted
(competitive kernel authoring with humans and AI systems was such a good idea that we decided to do it again for MLSys '26, partnering with @ye_combinator and the FlashInfer team at Nvidia) Join us! Free B200s. mlsys26.flashinfer.ai/
1
2
12
625
Anxhelo Xhebraj retweeted
🚀 MLSys 2026 Contest - @nvidia Track is LIVE! Registration is now open for the FlashInfer-Bench Challenge! Submit high-performance GPU kernels for cutting-edge LLM architectures on NVIDIA Blackwell GPUs. Three Tracks * MoE (Mixture of Experts) * DSA (Deepseek Sparse Attention) * GDN (Gated Delta Net) Human experts AND AI agents welcome — evaluated separately. Let's see who builds the best kernels! 🤖 🎁 Prizes: Winners take home NVIDIA GPUs and are invited for presentation at MLSys 2026. ⚡ First 50 teams to register get free GPU credits from @modal - huge thanks for the sponsorship @charles_irl ! Whether you're a kernel wizard or building autonomous coding agents, we want to see what you've got. 🔗 Contest details: mlsys26.flashinfer.ai See you at MLSys 2026! 🔥
4
56
12
294
76,288
This is how I found out about Jax for the first time. @s_scardapane gave a few presentations on it some time in early 2019 to later publish the corresponding blogposts web.archive.org/web/20210814…
i'm sorry but how did that happen
1
287
Anxhelo Xhebraj retweeted
ARRAY 2025 (co-located with @PLDI) CFP is open! pldi25.sigplan.org/home/ARRA…
1
6
880
If you would like to share your work on array programming please consider submitting your paper to ARRAY '25 (co-located with @PLDI ). pldi25.sigplan.org/home/ARRA…
3
13
2,433
Anxhelo Xhebraj retweeted
We are excite to announce FlashInfer v0.2! Core contributions of this release include: - Block/Vector Sparse (Paged) Attention on FlashAttention-3 - JIT compilation for customized attention variants - Fused Multi-head Latent Attention (MLA) decoding kernel - Lots of bugfix and improvements involving CUDAGraph compatibility, RMSNorm/RoPE numerical issue, etc. blog post: flashinfer.ai/2024/12/16/fla…
6
41
6
158
33,865
Pretty excited to see this work landing in OCaml! Second-class (nonescaping) values (shameless plug dx.doi.org/10.4230/LIPIcs.EC…) are going mainstream! Great for mem management but also interacting with unsafe apis as shown in @jckarter's proposal for Swift forums.swift.org/t/a-roadmap…
Apropos of nothing in particular...modal types for OCaml is going to be amazing. We built it for stack allocation, but it's unlocking so many other possibilities...
1
16
33
In yesterday's #VCOOP22 meeting, Anxhelo Xhebraj asked (and answered) the question "What If We Don't Pop the Stack?" #ECOOP22 Thanks for the fascinating talk! 🧑‍💻💬 📄: drops.dagstuhl.de/opus/vollt…
4
7
Compilers: Dex/Futhark/SPIRAL/Lift do rewrites/optimisations based on high-level algebratic properties of the code (not sure any of them do scan-reduce fusion though). The problem is that even after you have all the information, exploiting it well is still an open problem.
2
1
4
Anxhelo Xhebraj retweeted
The Episode #2 of the typetheoryforall.com is out! In this episode I interview Dr. John Sarracino & Rajan Walia and we talk all about grad school life! @ your favorite podcast provider
1
6
23
For this semester I have decided to take some course notes in @obsdmd and I am really liking the aesthetics of it so far.
1
4
Still not as cool and feature rich as @tiarkrompf 's Octopus Notes though. tiarkrompf.github.io/notes/
1
2
💫Farewell to our #PurdueCS #GoBoiler2019 Interns - Timon, Marcel, Beatrice, Manuel, Christian, Anxhelo, Sara & Lala. Safe travels and thank you for your contributions! #computerscience
3
16
Meet Anxhelo Xhebraj (@0xA95) - one of our #GoBoiler2019 Interns here: ow.ly/Lnq750wit6e He is an MS student at @SapienzaRoma He works w/ #PurdueCS Prof. @tiarkrompf, on advantages Pyctr can provide to overcome the current limits of TorchScript. @camildemetrescu@dcdelia
2
10