Skip to content
View HGinkgo's full-sized avatar
🎯
Focusing
🎯
Focusing
11 followers · 22 following
  • University of Electronic Science and Technology of China
  • Chengdu,China
  • 14:19 (UTC +08:00)

Achievements

Achievement: Quickdraw

Achievements

Achievement: Quickdraw

Highlights

  • Pro

Block or report HGinkgo

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
HGinkgo/README.md

Hi, I'm Pengfei He 👋

I am an M.S. student at the University of Electronic Science and Technology of China (UESTC), with a background in communication systems, systems programming, and applied AI. I work with C/C++, Python, Linux, CUDA, and C++ inference runtimes. I am currently an LLM Inference Engine Intern at Enflame Technology Enflame Technology logo.

Research Focus

My research focuses on collaborative inference and scheduling optimization for small language models of different scales in resource-constrained edge environments. With a communication-systems background, I am particularly interested in the system-level trade-offs among inference latency, resource utilization, communication cost, and service quality.

In parallel, my engineering work focuses on LLM inference infrastructure: runtime adaptation, serving and scheduling, KV cache management, CUDA/Triton kernels, profiling, and C++ edge deployment. I am currently applying these skills to inference optimization on domestic AI accelerators.

Selected Projects

Project Description
LLM-Serve A single-node LLM inference runtime exploring Paged KV Cache, continuous batching, chunked prefill, admission control, CUDA Graph, speculative decoding, AWQ quantization, and experimental dual-GPU Prefill/Decode separation with shared-KV transport.
HunyuanOCR-ncnn A pure C++17 runtime for HunyuanOCR 1.5 using pnnx and ncnn, with dynamic vision input, KV-cache decoding, JSONL batching, CPU/Vulkan backends, and dimension-aware GEMM/GEMV routing. See the public technical report.
SeedVR2-ncnn Completed C++/ncnn/Vulkan implementation of SeedVR2 for image and video enhancement, including custom operators, staged graph execution, and fixed-shape GPU validation. See the public technical report.
LLM-Kernels CUDA and Triton implementations of LLM inference operators with correctness tests, benchmarks, profiling, and optimization notes.
riscv-memcpy-memset-opt RV64GC memcpy and memset optimizations using scalar unrolling, RVV 1.0, and Zicboz.

Open Source Contributions

Community Contributions
Mooncake logo Mooncake Contributed Mooncake Store allocator benchmarking and size-class-aware allocation strategy work, including a merged fragmentation benchmark and a size-class-aware strategy contribution.
Tencent logo ncnn Developed HunyuanOCR-ncnn and SeedVR2-ncnn, community implementations for deploying Tencent models through the pnnx/ncnn toolchain. See the HunyuanOCR report and SeedVR2 report.

Pinned Loading

  1. SeedVR2-ncnn SeedVR2-ncnn Public

    SeedVR2 image and video enhancement on Linux with C++/ncnn/Vulkan and BF16 inference.

    C++ 4

  2. HunyuanOCR-ncnn HunyuanOCR-ncnn Public

    HunyuanOCR 1.5 的 ncnn C++17 运行时，支持 Linux/Windows、JSONL 批处理、DFlash 与 Vulkan 视觉编码

    C++ 2 1

  3. LLM-Serve LLM-Serve Public

    A single-GPU LLM inference runtime for studying scheduling, KV cache management, serving benchmarks, and speculative decoding.

    Python 2

  4. riscv-memcpy-memset-opt riscv-memcpy-memset-opt Public

    Benchmark-driven RV64GC memcpy and memset optimizations using scalar unrolling, RVV 1.0, and Zicboz, validated with GLIBC 2.43 benchtests.

    Assembly 2

  5. LLM-Kernels LLM-Kernels Public

    CUDA and Triton kernels for LLM inference experiments

    Cuda 2

  6. ScaleFlow ScaleFlow Public

    Research framework for collaborative multi-scale language model inference and scheduling

    Python 1