Daya Khudia
Engineering Lead And Manager at Databricks
San Francisco Bay Area United States
Summary
🤩
Rockstar🎓
Top SchoolDaya Khudia is an engineering lead and manager in the San Francisco Bay Area with 11 years of experience building high-performance ML training and inference systems. Currently leading a Databricks team focused on PyTorch, vLLM, and TRT-LLM optimizations, he combines deep expertise in CUDA, compiler tech (including Triton), and distributed systems to remove performance bottlenecks at scale. His background spans research and production roles at MosaicML, Meta, and Intel, with hands-on work in model export (ONNX, FasterTransformer), quantization for transformers, and compiler-level optimizations in projects like PyTorch/Glow and MosaicML’s Composer. An active open-source contributor, he has improved LLM training pipelines and quantization tooling for widely used repos, and has a PhD in Computer Engineering from the University of Michigan—bringing a rare mix of kernel-level performance engineering and practical ML deployment experience.
11 years of coding experience
16 years of employment as a software developer
Indian Institute of Technology Delhi (IIT Delhi)Doctor of Philosophy (Ph.D.) Computer Engineering, Doctor of Philosophy (Ph.D.) Computer Engineering at University of Michigan
Github Skills (32)
pytorch10
quants10
c-language10
python10
operation10
tensorrt10
machine-learning10
data-export10
inference10
onnx10
exporter10
ml10
llm10
transformer-models10
deep-learning10
Programming languages (5)
Github contributions (5)
mosaicml/composer
Feb 2022 - Oct 2022
Supercharge Your Model Training
Role in this project:
ML Engineer
ML EngineerContributions:303 reviews, 44 commits, 74 PRs in 8 months
Contributions summary:Daya contributed to the documentation and implementation of various callback functionalities within the `composer` repository, which focuses on supercharging model training. Their work included adding docstrings to callback hyperparameters, implementing changes to the `RunDirectoryUploader` for uploading the run directory to a blob store, and improving docstrings for core modules such as algorithms, callbacks, and states. They also integrated FFVC (Fast Forward Computer Vision) for faster data loading for various datasets. In addition, they added a few new model export features and also fixed a few issues related to models.
pytorchml-systemsdeep-learningneural-networksmachine-learning
mosaicml/llm-foundry
May 2023 - Feb 2024
LLM training code for Databricks foundation models
Role in this project:ML Engineer
ML EngineerContributions:28 reviews, 9 PRs, 3 pushes in 8 months
Contributions summary:Daya primarily contributes to the development and maintenance of LLM training and inference code within the repository. They added a script for exporting a model to ONNX format, enabling model deployment and inference optimization. They also focused on preparing the model for inference serving, modifying the inference pipeline and integrating Hugging Face's generate function. Further contributions include converting model checkpoints to FasterTransformer format and improving error messages.
deep-learningllmneural-networksnlppytorch
Find and Hire Top DevelopersWe’ve analyzed the programming source code of over 60 million software developers on GitHub and scored them by 50,000 skills. Sign-up on Prog,AI to search for software developers.