Skip to content
View shreyansh26's full-sized avatar
👨‍🎓
Always Learning
👨‍🎓
Always Learning
375 followers · 243 following

Achievements

Achievement: Pair ExtraordinaireAchievement: Starstruckx2Achievement: Pull Sharkx2Achievement: Arctic Code Vault Contributor

Achievements

Achievement: Pair ExtraordinaireAchievement: Starstruckx2Achievement: Pull Sharkx2Achievement: Arctic Code Vault Contributor

Highlights

Organizations

@COPS-IITBHU

Block or report shreyansh26

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Content in all repositories owned by your account will be closed.
Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. Markdown is supported. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
shreyansh26/README.md

Hi, I’m Shreyansh 👋

I’m the Director of AI Research at Level AI, working on LLMs for conversational AI. My current focus is inference optimization, post-training, and agent harnesses.

I like understanding how models work—from the ideas in a paper to the kernels and systems that make them run efficiently. This GitHub is where I implement papers, experiment with CUDA and Triton, and build tools I want to use.

What I work on

  • LLM inference: attention, speculative decoding, sampling, KV caches, and GPU kernels.
  • Post-training: supervised fine-tuning and reinforcement learning.
  • Agents: harnesses and tools for useful, reliable workflows.

Previously at Mastercard AI Garage and Samsung Research. Computer Science, IIT (BHU) Varanasi, 2020.

📫 How to reach me

📕 Latest Blog Posts

Pinned Loading

  1. Annotated-ML-Papers Annotated-ML-Papers Public

    Annotations of the interesting ML papers I read

    287 28

  2. FlashAttention-PyTorch FlashAttention-PyTorch Public

    Implementation of FlashAttention (FA1-FA4) in PyTorch for educational and algorithmic clarity

    Python 227 26

  3. LLM-Sampling LLM-Sampling Public

    A collection of various LLM sampling methods implemented in pure Pytorch

    Python 30 3

  4. Speculative-Decoding Speculative-Decoding Public

    Speculative Decoding Implementations: MTP, EAGLE-3, Medusa-1, PARD, Draft Models, N-gram and Suffix Decoding from scratch

    Python 15 2

  5. multilora-llm-inference multilora-llm-inference Public

    An implementation of CUDA and Triton kernels for BGMV and SGMV for multi-lora LLM inference

    Python

  6. optimizing-qwen3-asr-inference optimizing-qwen3-asr-inference Public

    Optimizing Qwen3-ASR serving with vLLM through FP8 calibration, fused Triton kernels, and CUDA graphs - with latency, throughput, and accuracy benchmarks.

    Python