Akash Dasarraju

Looking for a summer 2027 internship in ML research or ML systems.

I write low-level GPU kernels and ML systems for LLM inference. My engine, hipinfer, has every kernel written from scratch and beats llama.cpp at batch-1 decode on every model I tested, on three GPUs. In UT Dallas’s CARAML Lab I work on two submodular ML papers in preparation, on learning from less and agent memory. I study computer science at UT Dallas and help organize HackUTD, a 1,300-person hackathon.

Orange links take you into the room where that work happens, each with a small demo you can play. Hover one first for a look at its screen.

CARAML Research LabData selection research Lecture hallAI workshops HackUTDOrganizing
UT Dallas, Richardson, Texas

Projects

hipinfer

LLM inference engine in HIP, every kernel written from scratch. Beats llama.cpp at batch-1 decode on every model I tested, on three GPUs: 4.14 ms per token against 4.77 on a 2080 Ti.

Play: tune TinyLlama to beat the memory wall

Details on hipinfer

C++, HIP/CUDA. GitHub, site, Colab

  • Built an LLM inference engine in HIP with every kernel written from scratch (no cuBLAS or vendor libraries), supporting dense and MoE models, multi-GPU tensor parallelism, int8 weights, speculative decoding and KV-cache eviction. Tested on NVIDIA; compiles for AMD.
  • Verified against Hugging Face: WikiText-2 perplexity within 0.01% on every model, 99.4 to 99.9% top-1 token agreement, a greedy token-match gate, and kernel tests against PyTorch.
  • Beat llama.cpp at single-request decode on every model tested on 3 GPUs (5 to 16% on 1B to 3B models on a 2080 Ti, at 81 to 87% of memory bandwidth); 23 to 31% faster than vLLM on an A100; GEMM at 92 to 95% of cuBLAS on Turing.
  • 1.5 to 1.7x faster decode on two RTX 2080 Tis with tensor parallelism and a custom all-reduce 1.4x faster than NCCL on small messages; runs 7B to 8B models (Llama-3.1-8B, Mistral-7B) too large for one card.
  • 1.6 to 1.8x faster decode than fp16 with int8 weights (16 to 26% faster than llama.cpp’s Q8_0, for 0.04 to 0.23% more perplexity).
  • Up to 1.39x faster decode with speculative decoding (1B draft, 3B target), its output token-for-token identical to plain greedy decoding in all 105 test runs on each of 2 GPUs.

SubCache

KV-cache compression benchmark. Keeping a quarter of the cache fit 3× the requests and gave 2.8× throughput.

Play: keep 32 of 86 tokens and see if it remembers

Details on SubCache

Python, PyTorch. GitHub, site

  • 2.8x higher throughput from keeping 25% of the KV cache on long coding-agent sessions (3x more requests per GPU, for a 5% rise in perplexity), measured in my open-source compression benchmark.
  • Benchmarked 9 eviction policies against H2O and R-KV on agent sessions, LongBench, needle-in-a-haystack retrieval and MATH-500.
  • My submodular eviction method beat H2O by 1.1 to 2.5% on repetitive agent sessions. Profiling showed selection took 71% of decode time, so I wrote GPU kernels for it in hipinfer, 6 to 25x faster than PyTorch.

mnk-bench

Ten game-playing methods on one bench. My 39k-parameter self-play network never lost to a perfect tic-tac-toe player.

Play: play tic-tac-toe against matchboxes and an LLM

Details on mnk-bench

Python, PyTorch. GitHub, live demo

  • Built an open-source benchmark of 10 game-playing AI methods, from classic search to self-play networks and LLMs, on boards from tic-tac-toe to Gomoku, with 118 tests and CI. My 39k-parameter AlphaZero-style network never lost to a perfect tic-tac-toe player.
  • Applied my lab's targeted data-selection research to games, choosing which positions to label to fix an agent's tactical blind spots. It found 2 to 3x more relevant examples than random selection.

Research

CARAML Lab, Prof. Rishabh Iyer
2026 to present

A submodular ML lab at UT Dallas. I implement and run the experiments for two papers in preparation, with PhD mentors.

Submodular learning from less

A submodular set loss that learns from several positives and hard negatives at once beat C2LIP, a published state-of-the-art CLIP, by 3% on fine-grained reasoning. Paper targeting CVPR.

Play: teach a model with 4 sentences

Details on submodular learning from less

Research, CARAML Lab

  • Beat C2LIP, a published state-of-the-art fine-tuned CLIP model, by 3% on fine-grained object–attribute reasoning (SugarCrepe, SugarCrepe++) with only 81k COCO image–caption sets, by implementing a submodular set loss that learns from several positives and hard negatives at once. Targeting CVPR.

Memory for LLM agents

Choosing which past turns an agent keeps under a fixed token budget, against top-k and MMR retrieval. Paper targeting ICML.

Play: pack a conversation into 48 tokens

Details on memory for LLM agents

Research, CARAML Lab

  • Developing a submodular method that chooses which past conversation turns an LLM agent keeps under a fixed token budget, compared against top-k and MMR retrieval, self-hosting Qwen2.5-32B with vLLM on an A6000. Targeting ICML.

Leadership

Organizer, HackUTD leadership team

Play: prompt-inject the model to leak a flag

UT Dallas

  • Help organize a 1,300-person hackathon, leading 10+ volunteers.
  • Built a RAG and MCP Discord bot that answers near 100% of questions and cut pre-event questions to organizers by 80% against prior years, and a 3D campus map (three.js) used by hundreds.
  • Led 5 AI workshops, tripling turnout from 30 to over 100.

Education

B.S. Computer Science, CS2 Honors Program

The University of Texas at Dallas

  • GPA 3.62. National Merit Finalist. ACT 36.
  • Coursework: Data Structures, Probability and Statistics, Computer Architecture (honors); Linear Algebra; Mathematics of Computing (senior level).

Skills

Languages and tools
Python, C++, Java; CUDA, HIP/ROCm, NCCL, PyTorch, vLLM, Hugging Face
GPUs
NVIDIA Turing and Ampere (RTX 2080 Ti, RTX A6000, A100)

Hiring for summer 2027? Let’s talk.

dasarrajuakash@gmail.com