Vima Gupta (@vima_gupta) profile picture

PhD Student at Georgia Tech @gt_computing Research Intern previously @MSFTResearch , @CerebrasSystem, @arm

Joined November 2022
1/7 🧵 MoEs: A tale of expectation vs reality Marketing: "Only compute the expert parameters you need!" Reality: Batch 16 requests → ALL experts activate At serving time (vLLM/TGI), arithmetic intensity: AI ≈ (num_tokens * top_k) / total_experts In simpler terms: Your decode arithmetic intensity scales inversely with expert count 🤔 #MoE #LLMs #ChatGPT #Claude #vllm #AI #ML
4
7
1
31
3,239
MoE models promise sparse compute. But once you batch requests, that promise starts to break. Open-sourcing Lynx - a system for efficient MoE inference through workload-agnostic expert reduction. Code: github.com/gt-nexs/lynx Paper: vimagupta.github.io/docs/pap… #MLSys2026 #YPS
1
2
4
315
The repo includes a quickstart so you can try Lynx on your own models and workloads. If you work on MoE serving, LLM inference, or ML systems, please try it, break it, open issues, and send PRs 🐈‍⬛ github.com/gt-nexs/lynx
1
135
I’m also presenting Lynx at the #MLSys2026 Young Professionals Symposium today at 5 PM PST. Thanks to my awesome collaborators: Jae Hyung Ju, @Karts6876, Ada Gavrilovska, @anandpiyer; and labmates at @gtnexs !! If you’re at MLSys, come find me — I’d love to talk all things MoE.
147
As a Chinese-born professor teaching hundreds of US and international students each semester, I’m saddened by remarks stereotyping Chinese students. Talent and integrity cross borders. Let’s uphold respect and fairness that will enrich future generations and our broader society.
11
11
4
300
23,044
Probably I'm a little late to this, but for all those applying for PhD positions during this cycle: I have seen PhD applicants stressing a lot on getting into a "great" school rather than getting an opportunity to work with a supportive and empathetic advisor. Believe me, working with someone who keeps on supporting and believing in you is really underrated.
2
7
1
78
7,615
6/7 We present Lynx 😼 Dynamically picks experts during decode = 1.5x speedup. Preserves critical routing decisions while optimizing for latency.
1
1
9
3,008
7/7 Paper: arxiv.org/abs/2411.08982 Code dropping soon! Work done at @gtcomputing with awesome collaborators Kartik Sinha, @anandpiyer and Ada Gavrilovska!!
6
369
🚀 Introducing Metron: Redefining LLM Serving Benchmarks! 📊 Tired of misleading metrics for LLM performance? Our new paper introduces a holistic framework that captures what really matters - the user experience! 🧠💬 github.com/project-metron/me… #LLM #AI #Benchmark
2
15
5
33
6,710
1/ LLM inference systems are like high-performance engines ⚙️—complex, powerful, and full of intricate settings. Efficiently deploying them to maximize GPU performance is a challenge typically tackled by experts at orgs like @OpenAI and @AIatMeta 🚀. 🧵
1
13
39
3,247