PeopleofSuperintelligence

Builder

Deyu Fu

Deyu Fu is an AI engineer whose work spans large-scale recommender training, multimodal language models, and model-training systems. He co-authored NVIDIA's 2022 work on terabyte-scale recommender training and appeared on Reka's credited teams for Flash, Core, and Edge in 2024. His 2024 work also includes Vibe-Eval and the use of NVIDIA Megatron-Core to train Reka's multimodal language models. In 2026, he co-authored research on mixture-of-experts training and LLM optimizers, and was named as an inventor on an NVIDIA graph-execution patent application. As of October 2026, he is a senior developer technology engineer on NVIDIA's Deep Learning Frameworks team.

Now at
NVIDIA
Works
6

Publication numbers from OpenAlex, as of 19 October 2026. Citations and h-index are hidden because this OpenAlex record may be incomplete or mix in other authors.

Timeline

  1. August 2022

    Terabyte-scale recommender training

    In August 2022, Fu co-authors an NVIDIA post about Merlin Distributed-Embeddings, which distributes recommender-model embeddings across multiple GPUs. It reports a 683× speedup over CPU-only training with an eight-GPU NVIDIA DGX A100.

  2. February 2024–April 2024

    Reka Flash, Core, and Edge

    In February 2024, Reka credits Fu on its Flash announcement and describes Flash as a 21B model trained from scratch. The April 2024 report lists him among the authors of Core, Flash, and Edge and says the models process and reason with text, images, video, and audio.

  3. May 2024

    A multimodal chat benchmark

    In May 2024, he co-authors Vibe-Eval, an open benchmark for multimodal chat. It contains 269 visual-understanding prompts, including 100 difficult prompts with gold responses written by experts.

  4. July 2024

    Megatron-Core at Reka

    In July 2024, Fu is a technical staff member at Reka AI and says Megatron-Core integrates into the company's multimodal LLM architecture. He says optimized GPU kernels and parallelism help train very large models and extensive contexts across clusters.

  5. March 2026

    Scaling MoE with Megatron Core

    In March 2026, Fu co-authors a Megatron Core report on mixture-of-experts training. The report describes integrated optimizations for memory, communication, and computation, with Parallel Folding and support for FP8 and NVFP4.

  6. March 2026

    An NVIDIA graph-execution patent

    An NVIDIA patent application published in March 2026 lists Fu as an inventor. It covers techniques for modifying an executable graph to perform a workload associated with a new task graph.

  7. July 2026

    LLM pretraining optimizers

    In July 2026, he co-authors a study of SOAP, Muon, and AdamW for large-scale LLM pretraining. It introduces a layer-wise distributed optimizer compatible with Megatron-LM.

  8. As of October 2026

    NVIDIA Deep Learning Frameworks

    Fu is a senior developer technology engineer on NVIDIA's Deep Learning Frameworks team. His work spans training across CUDA kernels, math, communication, frameworks, and models.

Career

  1. NowNVIDIASenior Developer Technology Engineer
  2. From 2024Reka AITechnical Staff Member

From the biography.

  1. Google alumni founded or led 87 other organizations, the most among the labs we compareIncludes NVIDIA. We trace 212 organizations founded or led by AI lab alumni, and 223 career moves between the labs.
  2. 44% of AI career moves go from one university to anotherIncludes NVIDIA. We estimate how 14,272 AI career moves link universities, companies, labs and government research centers, with net flows and the most common paths.
  3. Google has the highest combined h-index among AI organizationsIncludes NVIDIA. Our estimates of combined and median h-index at AI organizations, using current affiliations and OpenAlex records.
  4. Google had the most people shaping AI every year from 2014 to 2025Includes NVIDIA. Our estimates of which organizations employed the most people shaping AI from 1961 to 2025, with a year-by-year chart.

Colleagues

People with dated positions at the same organization in the same years.

Sources and credits

Photo: We have no published photo of Deyu Fu yet. Send us a free-licensed photo.

  1. Fast, Terabyte-Scale Recommender Training Made Easy with NVIDIA Merlin Distributed-Embeddings developer.nvidia.com. Checked 19 October 2026.
  2. Reka Flash: Efficient and Capable Multimodal Language Models reka.ai. Checked 19 October 2026.
  3. [2404.12387] Reka Core, Flash, and Edge: A Series of Powerful Multimodal Language Models arxiv.org. Checked 19 October 2026.
  4. [2405.02287] Vibe-Eval: A hard evaluation suite for measuring progress of multimodal language models arxiv.org. Checked 19 October 2026.
  5. Train Generative AI Models More Efficiently with New NVIDIA Megatron-Core Functionalities developer.nvidia.com. Checked 19 October 2026.
  6. Scalable Training of Mixture-of-Experts Models with Megatron Core arxiv.org. Checked 19 October 2026.
  7. US20260079755A1 - Techniques for modifying an executable graph to perform a workload associated with a new task graph patents.google.com. Checked 19 October 2026.
  8. [2607.20548] SOAP, Muon, and Beyond: Pushing LLM Pretraining Scales arxiv.org. Checked 19 October 2026.
  9. Author: Deyu Fu | NVIDIA Technical Blog developer.nvidia.com. Checked 19 October 2026.