Alexey Tumanov (@alsched) profile picture

Associate Professor @gatech_scs @gtcomputing | PostDoc @Berkeley_EECS @ucbrise | ML/LLM Systems: cloud & edge.

Atlanta, GA
Joined December 2012
Photo by Alexey Tumanov (@alsched)Photo by Alexey Tumanov (@alsched)Photo by Alexey Tumanov (@alsched)
I am presenting this work, "No Request Left Behind: Tackling Heterogeneity in Long-Context LLM Inference with Medha" at @ACMSoCC today, 14:00 local time. It has already been adopted by OSS from our earlier preprint. Key contributions: * chunked prefill unrolls the prefill computation in time abandoning the SOTA-standard spatial context parallelism (CP) * demonstrating that Sarathi-style chunked prefill read amplification cost can be fully hidden, because arithmetic intensity depends only on chunk size (the paper proves this) * contributing a 3D parallelism mechanism (TP+SPP+KVP) that fixes both TTFT and TBT (TPOT) latency challenges for long context (> 1M tokens) * LARS scheduler : prioritizing least remaining slack per unit total work -- a novel scheduling technique that outperforms much stronger baselines than the typical FCFS scapegoat. paper: dl.acm.org/doi/10.1145/38157… arxiv: arxiv.org/abs/2409.17264v4 This work is pure, approximation-free play. It's becoming increasingly pertinent in contexts extending beyond LLM inference serving alone. Perhaps the most foundational aspect of this work for me, from scheduling perspective, is that it achieves liquidity in scheduler decision space. It doesn't get more fundamental than this. That enables mixed batching, preemption, dropping the convoy effect, and LARS.
Thank you for pointing out the inadequacy of current SOTA implementations of long context support. Our team has demonstrated the superiority of chunked prefill mechanism for enabling preemptive scheduling in support of long context requests (10M tokens): arxiv.org/abs/2409.17264
1
3
9
446
The future of dev and design is voice -- done right with @AvarthaAI's Ava agent.
Voice is poised to be the dominant interaction surface for all of computing in the future. So far voice ai have been held back by chunky models, but when you plug in models like @Zai_org's GLM 5.3 in the loop, possibilities become endless. Here is a snippet of what our team has been cooking with Ava, our voice agent. Happy Halloween! 🎃
4
148
Alexey Tumanov retweeted
It is good to finally see the open-source community finally catching up ~1.5 years after we originally published the recipe for long context serving. Still missing several critical components - good for microbenchmarking not for end-to-end deployments. arxiv.org/pdf/2409.17264
2
7
418
Alexey Tumanov retweeted
Five months ago, we started Avartha with the belief that better systems can unlock the next generation of real-time applications. Today, we’re announcing Ava Zero, our voice agent built entirely on open models. Ava ranks #1 on @SierraPlatform voice agent benchmark, ahead of OpenAI. Incredibly proud of our team. We have some of the best systems builders in the world, and we’re just getting started.
Introducing Avartha. We're building the infrastructure to bring frontier-grade intelligence to real-time applications. Our first release, Ava Zero, is a voice agent entirely running open models using our inference infrastructure and ranks #1 on Sierra's tau-voice benchmark with highest task completion rate and generation quality and just costs $2.5 per hour. Open models are now at the frontier of voice AI.
5
6
50
47,490
Incredible feat of research and engineering. The @AvarthaAI Team managed to achieve frontier intelligence at a fraction of the cost, while serving open models on the real-time voice AI platform they built from scratch. Not a wrapper, not a derivative product. Built from first principles, grounded in years of research, dedication, and singular focus on latency-first systems design coupled with end-to-end design principles -- the holistic approach that helped Avartha's Ava Zero voice agent outperform @OpenAI's GPT Live 1 on tau-Voice bench.
Introducing Avartha. We're building the infrastructure to bring frontier-grade intelligence to real-time applications. Our first release, Ava Zero, is a voice agent entirely running open models using our inference infrastructure and ranks #1 on Sierra's tau-voice benchmark with highest task completion rate and generation quality and just costs $2.5 per hour. Open models are now at the frontier of voice AI.
1
4
18
1,004
Alexey Tumanov retweeted
LLM inference is quickly becoming a global infrastructure problem. Leveraging AI agents to optimize these systems is the natural way to accelerate their development. But AI driven optimization requires a fast, cheap, and accurate evaluation mechanism. 🧵
Systems research is changing: AI is enabling white-box algorithm discovery, not just black-box tuning. Our new ADRS paper evaluates three frameworks and shows consistent gains over human designs—up to 13× speedup in load balancing for MoE experts and 35% lower cloud costs across different cloud regions.
1
2
6
641
Thank you for pointing out the inadequacy of current SOTA implementations of long context support. Our team has demonstrated the superiority of chunked prefill mechanism for enabling preemptive scheduling in support of long context requests (10M tokens): arxiv.org/abs/2409.17264
Chunked pipeline parallelism is arguably the most general and scalable system technique for accelerating super-long-context inference. It remains underrated today, largely because there still isn’t a strong, high-quality open-source implementation. The SGLang team recently fully optimized it and published a detailed blog post explaining all the key details.
1
851
Academic credit for chunk prefill keeps getting missed by those reimplementing it in their own systems. arxiv.org/abs/2409.17264 demonstrated the superiority of chunked prefill mechanism for enabling preemptive scheduling in long context requests (up to 10M tokens).
Chunk pipeline parallelism provides two critical advantages: 1. It scales ridiculously well. you can get 85%+ efficiency even at pp=32, 128 H100! 2. It supports preemption, so you avoid the terrible convoy effect that happens with long context. 10x+ improvement over CP 1/2
3
233
Alexey Tumanov retweeted
After hitting evaluation puzzles like this in our own work, we analyzed patterns across LLM inference papers and identified 8 systematic evaluation issues that can make performance comparisons misleading. We have compiled a practical evaluation checklist to help avoid these pitfalls. 📄 arxiv.org/abs/2507.09019 We're also releasing Veeksha, our comprehensive LLM inference evaluation framework, later this month to help the community design more robust benchmarks! 🛠️ What evaluation issues have you discovered in your systems work? Let's learn from each other's mistakes! @nitinkedi @jayashree2912 @kwatra @thisissouvikk @ramaramjee @alsched @gtcomputing @MSFTResearch @intel
3
5
622
Alexey Tumanov retweeted
Interesting work on long context inference from @nvidia, where they scale KV parallelism on gb200-nvl72 systems! To learn more about accelerating long context inference and trade-offs between different parallelism dimensions checkout out our paper, Medha: arxiv.org/abs/2409.17264
What if you could ask a chatbot a question the size of an entire encyclopedia—and get an answer in real time? Multi-million token queries with 32x more users are now possible with Helix Parallelism, an innovation by #NVIDIAResearch that drives inference at huge scale. 🔗 nvda.ws/4eCXxqh
5
14
1,081
Congratulations 👏 to our faculty who were recognized on the Spring 2025 CIOS Honor Roll for their outstanding teaching and educational impact: Assoc. Prof. Alexey Tumanov and Asst. Prof. Jan Van Den Brand!
1
11
417
Alexey Tumanov retweeted
Super excited to share another incredible systems that we have built over the past two years! Training giant foundation models (like Llama-3 405B) costs a FORTUNE 💰 (millions of dollars)! Optimizing the training "recipe" (parallelism, memory tricks, etc.) is critical but incredibly complex. The wrong choices can waste millions. How do we find the best setup without burning GPUs? This is the problem we tackle with Maya, GPU cluster emulation tool -- have a 1000 GPU job, want to know how it would will perform, all you need is a one cpu and Maya virtual GPU runtime ✨ Arxiv: arxiv.org/abs/2503.20191 Code: Coming soon! 🧵
1
12
21
1,944
Alexey Tumanov retweeted
Super long-context models with context window spanning millions of tokens are becoming commonplace (@GoogleDeepMind Gemini, @xai Grok 3, @Alibaba_Qwen Qwen2.5). But efficiently serving these models is tough, especially alongside short requests. Head-of-Line (HOL) blocking becomes a major issue, hurting latency for everyone. We present Medha, a system designed to handle this mix efficiently. Achieving 30x lower latency, and 5x higher throughput compared to the state-of-the-art. Full paper: arxiv.org/pdf/2409.17264. 🧵
1
13
2
31
3,721
Alexey Tumanov retweeted
Maya offers a transparent, accurate, and efficient way to model and optimize large-scale DL training without needing expensive hardware clusters for exploration. A crucial step towards sustainable AI! Read the paper: arxiv.org/abs/2503.20191 Work done with @Y_Srihas , @1ntEgr8 , Hakesh Darapaneni, Mitali Meratwal, Shivam Mittal, Pranavi Bajjuri, Srinivas Sridharan, @alsched at @GeorgiaTech @NVIDIA @NVIDIAAI
1
2
311
Alexey Tumanov retweeted
Sequence pipeline parallelism being rapidly adopted for extreme long context inference in the industry! Checkout our paper on system design for long context inference for more details arxiv.org/abs/2409.17264
We're leveling up the game with our latest open-source models, Qwen2.5-1M ! 💥 Now supporting a 1 MILLION TOKEN CONTEXT LENGTH 🔥 Here's what’s new: 1️⃣ Open Models: Meet Qwen2.5-7B-Instruct-1M & Qwen2.5-14B-Instruct-1M —our first-ever models handling 1M-token contexts! 🤯 2️⃣ Lightning-Fast Inference Framework: We’ve fully open-sourced our inference framework based on vLLM , integrated with sparse attention methods. Experience 3x to 7x faster processing for 1M-token inputs! ⚡⚡ 3️⃣ Tech Deep Dive: Check out our detailed Technical Report for all the juicy details behind the Qwen2.5-1M series! 📊 📖 Technical Report: qianwen-res.oss-cn-beijing.a… 📄 Blog: qwenlm.github.io/blog/qwen2.… Experience Qwen2.5-1M live: 👉 Play with Qwen2.5-Turbo supporting 1M tokens in Qwen Chat (chat.qwenlm.ai) 👉 Try it on Huggingface (huggingface.co/collections/Q…) 👉 Or head over to Modelscope (modelscope.cn/collections/Qw…)
4
19
1,158
Alexey Tumanov retweeted
At SoCC’24, Anastasia Ailamaki from EPFL will give a keynote on how disaggregated memory resources are becoming the norm and how this “new memory wall” affects database system design. This talk will be amazing, make sure to be there!! acmsocc.org/2024/keynotes.ht…
1
10
556
Super-charged technical program this year at @ACMSoCC: acmsocc.org/2024/schedule.ht… Looking forward! Hope to see you there! #socc24
We are just under a month away from SoCC’24! This year’s conference will be from Nov 20-22 at the Microsoft Campus in Redmond, WA . Early bird registration is now open until Nov 6. Make sure to register! acmsocc.org/2024/register.ht…
4
321
Alexey Tumanov retweeted
⚡ Speed Meets Accuracy: Unlike approximation-based methods, Mnemosyne achieves exact inference—ensuring that the generated output remains precise, even when processing 10 million tokens by effectively combining all these parallelization techniques to scale up to hundred of GPUS
1
2
6
557