Transformer GPU Kernel Optimization
A notebook-based investigation of inference performance: quantization, KV caching, tiled attention, speculative decoding, and fused Triton softmax kernels.
AKARSH MALIK / SOFTWARE ENGINEER · ML SYSTEMS & AGENTIC AI
Open to opportunitiesML Systems, GPU Kernels & LLM Inference
GPU KernelsLLM InferenceSystems
I’m Akarsh, a software engineer focused on GPU kernels, LLM inference, and performance-oriented ML systems, with experience building agentic AI systems and production software.
MS Computer Science @ NYU / New York
FastInfer
Triton / PyTorch / GPU Kernels
PROJECTION MICROBENCHMARKMeasured
Qwen2.5-1.5B · first-layer up_proj
1.174–1.176×isolated projection speedup
Latency / μs Lower is better ↓
Measured across 3 captured decode inputs.
01 / GPU KERNELS & INFERENCE
Featured Project
GPU KERNELS / INFERENCE RUNTIME
FastInfer
FastInfer is an experimental LLM inference optimization project exploring how custom GPU kernels translate from microbenchmark improvements to real model-level performance.
It implements Triton kernels for common transformer operations and evaluates them against PyTorch references for correctness, latency, and integration into cached Qwen decoding.
01
Kernel optimization. Implemented and benchmarked a Triton GEMV kernel for single-token inference, alongside RMSNorm, softmax, and SwiGLU kernels.
02
Correctness. Validated GPU outputs against PyTorch and higher-precision references before performance measurement.
03
Model integration. Integrated opt-in kernel replacements into cached Qwen decoding and measured whether isolated kernel improvements translated into lower end-to-end latency.
MODEL-LEVEL CACHED DECODEMeasured
Qwen2.5-1.5B · RMSNorm candidate
4.8%lower median decode latency
Latency / ms/step Lower is better ↓
−1.588 ms/stepsaved versus baseline
02 / PROJECTS
Selected Projects
Inference optimization, interactive data visualization, and real-time auctions.
Weave
Turn a CSV file and a plain-English prompt into an interactive D3.js chart - no code required. Supports 16 chart types with interactive sliders and filters.
Snatch
A real-time, multi-phase auction platform supporting descending Dutch auctions, ascending English auctions, and round-based elimination.
03 / ENGINEERING BACKGROUND
Experience
Professional experience and selected technical leadership.
At a glance
- Professional experience
- 2 years
- Industry roles
- 1 full-time · 2 internships
- Primary areas
- ML Systems · Agentic AI · Backend
Professional Experience
AB InBev
FULL-TIMEData Scientist
Aug 2023 - Aug 2025
AB InBev
INTERNSHIPData Analyst Intern
Jun - Jul 2022
Quantale
INTERNSHIPSoftware Developer Intern
Sep - Nov 2021
Student Leadership
Delta Force · NIT Trichy
- Head - Delta Winter of Code
- Manager - Pragyan CTF
- Web Operations - Pragyan & Sensors
Led open-source initiatives, managed cybersecurity events, and contributed to web and CTF infrastructure at NIT Trichy.
04 / ALGORITHMS & PROBLEM SOLVING
Algorithms and Problem Solving
LeetCode Knight with a peak contest rating of 1963. My algorithm practice includes graphs, dynamic programming, trees, and binary search.
Graphs / Dynamic programming / Trees / Binary search
View LeetCode profile
LeetCode KnightPEAK CONTEST RATING
AKARSH MALIK / NEW YORK
05 / BACKGROUND
About
I’m pursuing my master’s in Computer Science at NYU after two years as a Data Scientist at AB InBev. My work includes agentic AI projects using LangGraph and LangChain, alongside interests in distributed systems and LLM orchestration.
I’m currently working on Triton kernels and studying how kernel performance translates to LLM inference latency.
OFF THE CLOCK
Let’s talkRunning (4:50 min/km) · Hiking · Systems rabbit holes
06 / EDUCATION
Education
New York University
M.S. Computer Science
Sep 2025 – May 2027 · GPA 3.88/4
Deep Learning / Machine Learning / Operating Systems / Design and Analysis of Algorithms
National Institute of Technology, Tiruchirappalli
B.Tech in Instrumentation and Control Engineering
Jul 2019 - May 2023 · GPA 8.27/10
Distributed Systems / Machine Learning / Big Data Analytics
TECHNICAL SKILLS
Programming
Python / SQL / C++ / Java / JavaScript / Shell
ML / AI Systems
PyTorch / Transformers / LLMs / RAG / Triton / ONNX / FAISS / LangGraph / LangChain
Backend / Systems
Node.js / REST APIs / WebSockets / Docker / PostgreSQL / Redis / Elasticsearch / Kubernetes
Cloud / Data
Azure / AWS / Databricks / Delta Lake