Hi Pablo and Beam Dev Team,

  My name is Sai Shashank Mudliar, and I am a full-time Computer
Engineering student at Purdue University. I am writing to express my
interest in the GSoC 2025 project "A learning path to using accelerators
with Beam."

  I believe I bring a strong fit for this project because I have direct
production experience with the exact stack involved. I served as a Staff
Machine Learning Engineer at CVS Health, where I:

  - Deployed multiple ML models for inference on Google Cloud Dataflow at
scale, several of which are running in production today
  - Authored an internal ML Inference Cookbook used across CVS Health's ML
teams for deploying models on Dataflow with accelerators
  - Led a Proof-of-Concept integrating NVIDIA Triton Inference Server with
Beam by building a custom ModelHandler -- a framework not natively
supported at the time
  - Built GPU-accelerated Dataflow pipelines with custom Docker containers,
Flex Templates, and T4/A100 GPU workers

  I have also identified and am actively working on a concrete
contribution: the existing TensorRTEngineHandlerNumPy in
tensorrt_inference.py uses deprecated TensorRT 8.x/9.x APIs (num_bindings,
get_binding_name, execute_async_v2, trt.volume) that are incompatible with
TensorRT 10.x. I plan to submit a PR fixing this before the GSoC period
begins.

https://github.com/apache/beam/pull/36309/changes/c260d81b3f9dda4bfff8e069b4c30fee31bac710..caa66d7f135cd1a0fa6bcd67d26b28101707a0b9

  My proposal covers the five deliverables outlined in the project idea --
a CPU baseline script, GPU-accelerated inference, TPU inference with LLMs,
a parallel multi-model training pipeline, and a blog post -- plus
continuous testing infrastructure. Each deliverable is provided
  as both a standalone Python script and a Jupyter notebook, addressing the
gap that all 44 existing beam-ml examples are notebooks only. I have
attached my full proposal for your review.

  I would welcome any feedback on the proposal or guidance on how to best
align with the project's priorities. I am also happy to hop on a call at
your convenience.

  Thank you for your time and for mentoring this project.

  Best regards,
  Sai Shashank Mudliar
  Purdue University, Computer Engineering

Attachment: GSoC_Proposal_HW_Accelerators_Learning_Path.docx
Description: MS-Word 2007 document

Reply via email to