Skip to main content
mshr-h.com

Curriculum Vitae

Masahiro Hiramori (平森 将裕)

E-mail: [email protected]
Location: Greater Tokyo Area (Kanagawa), Japan
GitHub: mshr-h
LinkedIn: masahiro-hiramori

📄 Download 2-page CV as PDF

Languages: Japanese (Native), English (Professional working proficiency)

Professional Summary

AI systems engineer with 9+ years of experience translating ML model representations into compiler and inference implementations. Apache TVM Committer and PyTorch Ambassador, with work spanning PyTorch model import, ONNX operator definitions, custom LLM support in llama.cpp, and a C++ runtime built from scratch for an FPGA accelerator. I combine hands-on engineering with technical leadership at Mitsubishi Electric.

Selected Impact

  • Model representations and upstream development: Introduced and maintain Apache TVM’s PyTorch ExportedProgram frontend, authored 140+ merged TVM pull requests as of July 2026, and added ONNX’s FlexAttention operator in the ai.onnx.preview domain.
  • Custom LLM inference: Implemented LatentLLM support in an internal llama.cpp fork, from weight conversion and model loading to computation graphs and latent KV caches, using the existing CUDA backend, and evaluated it on NVIDIA DGX Spark.
  • C++ runtime engineering: Built a runtime from scratch and developed the driver for a proprietary FPGA accelerator, achieving 100+ FPS with an internal YOLO v2-based object detector on Ultra96-V2.

Engineering Profile & Skills

  • Programming Languages: C++, C, Python
  • Model Representations & Compilers: PyTorch torch.export / ExportedProgram, PyTorch FX, Apache TVM Relax and MetaSchedule, ONNX operator definitions, ONNX-MLIR
  • Runtime Integration & Edge AI: llama.cpp (existing CUDA backend), weight conversion, model loading, computation graphs, latent KV caches, C++ drivers and runtimes, model compilation, and edge deployment
  • Open-Source Engineering: Upstream implementation, design and code review, maintainer workflows, CI/CD, test automation, and reproducible builds
  • Technical Leadership: Technical roadmapping, architecture review, cross-functional collaboration, mentoring, and OSPO

Experience

Head Engineer, Open-Source & InnerSource Program Office, Mitsubishi Electric Corporation, Japan - Apr 2025 to Present

  • Implemented LatentLLM support in an internal llama.cpp fork: weight conversion, model loading, computation graph definitions, and latent KV cache support, using the existing CUDA backend.
  • Internal benchmarks on NVIDIA DGX Spark used CUDA, batch size 1, and default llama.cpp benchmark settings, with llm-jp-4-8b in BF16 as the baseline.
  • LatentLLM in BF16 achieved 50% fewer parameters and 2x decode throughput; adding Q4_K_M achieved 80% smaller model files and 5x decode throughput, all relative to the baseline.
  • Set technical direction for ML compiler and edge-AI open-source adoption across multiple R&D teams, connecting industrial deployment requirements to upstream PyTorch and Apache TVM development.
  • Led deployment validation of a 1.8B-parameter manufacturing-domain language model on Jetson Orin Nano and RK3588, measuring latency, memory use, and portability across constrained edge environments.
  • Shape company-wide adoption of open-source and InnerSource engineering practices for AI/ML software, with emphasis on ML compilers, edge inference, and on-device LLMs.
  • Provide architecture review and mentoring across compiler integration, edge deployment, and open-source engineering.

Head Researcher, Information Technology R&D Center, Mitsubishi Electric Corporation, Japan - Apr 2024 to Mar 2025

  • Expanded upstream ownership in Apache TVM and became a Committer in September 2024.
  • Led R&D on ML compiler frontends and edge deployment pipelines connecting PyTorch and ONNX models to Apache TVM and heterogeneous hardware.
  • Evaluated Apache TVM MetaSchedule on edge devices and co-authored a peer-reviewed paper presented at IEEE GCCE 2024.
  • Developed an on-device deployment pipeline for a manufacturing-domain language model on NVIDIA Jetson Orin Nano and Rockchip RK3588, including model conversion, optimization, and runtime integration.

Researcher, Information Technology R&D Center, Mitsubishi Electric Corporation, Japan - Apr 2017 to Mar 2024

  • Developed the user-mode driver and built the runtime from scratch in C++ for a proprietary FPGA-based neural-network accelerator, achieving 100+ FPS with an internal YOLO v2-based object detector on Ultra96-V2 (AMD/Xilinx Zynq UltraScale+ MPSoC).
  • Implemented debug tools for instruction memory and internal buffer dumps, memory loading, and memory load/store support to run the runtime on x86 development PCs.
  • Built image-anomaly-detection pipelines on Hailo-8 and Renesas RZ/V2H for visual inspection in manufacturing, optimizing models and runtimes for resource-constrained edge devices.
  • Prototyped end-to-end model conversion, compilation, and deployment pipelines using Apache TVM and related tooling across heterogeneous edge hardware.
  • Contributed upstream to Apache TVM, Microsoft Hummingbird, and ONNX-MLIR, focusing on PyTorch model import, CI/CD infrastructure, and industrial edge workloads.

Open-Source & Community Activities

Apache TVM (Committer)

  • Introduced and maintain the ExportedProgram frontend, enabling models exported with PyTorch 2 torch.export to be imported into TVM Relax; added dynamic-shape support, custom operators, and expanded operator coverage.
  • Authored 140+ merged pull requests as of July 2026 across Relax and PyTorch frontends, torch.export.ExportedProgram import, dynamic-shape support, target and runtime fixes, and CI modernization.
  • Review and maintain contributions related to Relax frontends, model import and export, compiler infrastructure, and CI pipelines.

ONNX (Contributor)

  • Added the FlexAttention operator to ONNX’s ai.onnx.preview domain in PR #7534, merged in April 2026, supporting customizable attention through user-defined subgraphs.
  • Implemented the operator schema, type and shape inference, function expansion, Python reference implementation, and tests for the operator and shape inference.

PyTorch (Contributor)

  • Contributed integration paths between PyTorch and Apache TVM, including TVM backend interfaces, MetaSchedule configuration, and platform compatibility fixes.
  • Represent industrial and edge deployment use cases in the wider PyTorch community as a PyTorch Ambassador.

Microsoft Hummingbird (Committer)

  • Authored 33 merged pull requests through 2024, including Apache TVM backend integration and CI improvements for model conversion and execution.

Selected Projects

vscode-verilog-hdl-support, Creator, Maintainer

  • Created and maintain a Verilog/SystemVerilog development extension for Visual Studio Code with 1.44M+ installs on the Visual Studio Marketplace. It provides syntax highlighting, snippets, linting, formatting, waveform viewing, and optional language-server integration for FPGA/ASIC development workflows.

onnx2fx, Creator, Maintainer

  • Created an Apache-2.0-licensed converter that turns ONNX models into inspectable, transformable, and trainable PyTorch FX GraphModule objects for optimization, analysis, and deployment.

Professional & Community Service

  • PyTorch Ambassador, PyTorch Community - Represent industrial AI, edge deployment, and ML compiler use cases in the PyTorch community.
  • Program Committee Member, PyTorch Conference 2026 - Application, Inference, and Compilers & Kernels tracks. Program committee
  • Program Committee Member, PyTorch Conference 2025 - Compilers & Kernels track. Program committee

Education

M.S. in Information Science and Technology
Osaka Institute of Technology, Japan - Apr 2015 to Mar 2017

B.S. in Information Science and Technology
Osaka Institute of Technology, Japan - Apr 2011 to Mar 2015

Talks and Publications

Patents

Media Coverage

Awards & Recognition

  • PyTorch Community Awards 2025 - Nominee
  • PyTorch Community Awards 2024 - PyTorch Innovator Nominee