ut-sysml

Systems and ML Research Group at UT Austin led by Neeraja Yadwadkar

[ut-sysml @UT_SysML]

@li_ruihao presenting our work on Oneiros at SoCC'25!

KV cache drives multi-tenant LLM performance. Offloading it to the CPU is costly. Oneiros flips the script—repurposing parameter memory from inactive models into KV cache and exploiting today’s fast CPU–GPU bandwidth.

[ut-sysml @UT_SysML]

Blunt truth: LLM deployments are quietly bleeding money. A wrong config can make the same model 8× slower and 25× pricier.

With MaverIQ, we fix it by auto-deploying LLMs to meet each user’s intent—at a fraction of today’s cost

@dimliak99
@prasoon_sinha25
Tianrui Hu
@NeerajaJY

[Neeraja Yadwadkar @NeerajaJY]

Dimitris presenting our work on Intent-based LLM inference serving, MaverIQ, at SC'25, earlier this week!
@UT_SysML

[ut-sysml @UT_SysML]

Congrats, @li_ruihao!

[ut-sysml @UT_SysML]

Congrats, @prasoon_sinha25 👏

[Neeraja Yadwadkar @NeerajaJY]

Looking forward to @ACMSoCC'25!

Our group, @UT_SysML, will present 2 papers.

@prasoon_sinha25 will present ALAP, our intent-based serverless computing platform.

@li_ruihao will present Oneiros, a parameter remapping engine that accelerates multi-tenant LLM inference serving.

[ut-sysml @UT_SysML]

Congrats, @li_ruihao!

[Neeraja Yadwadkar @NeerajaJY]

My second Ph.D. student, co-advised with Lizy John, Ruihao Li, defended successfully! Ruihao is headed to Meta Research soon.

Congrats @li_ruihao, really proud of you :)

@UT_SysML @UTAustin @utexasece

[ut-sysml @UT_SysML]

Congrats, Susy!! @esalcorta

[Neeraja Yadwadkar @NeerajaJY]

A proud moment as my first co-advised PhD student, Erika Alcorta, graduates 🎓

It was an honor to hood her with Prof. Gerstlauer (@AGerstlauer). She’s now advancing ML for Systems at Ampere Computing. Excited to see all that lies ahead! 🚀

Congrats, Susy! @esalcorta

@UT_SysML