Skip to content
View waynehacking8's full-sized avatar
:electron:
:electron:
6 followers · 12 following

Achievements

Achievement: Pair Extraordinairex3Achievement: Pull Sharkx2Achievement: QuickdrawAchievement: YOLO

Achievements

Achievement: Pair Extraordinairex3Achievement: Pull Sharkx2Achievement: QuickdrawAchievement: YOLO

Block or report waynehacking8

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.

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
waynehacking8/README.md

Wei Cheng (Wayne) Chiu

Field Application Engineer @ Taiwan AILabs

I am a Field Application Engineer targeting customer-facing AI Solutions Architect roles, with hands-on work across Generative AI (GenAI), LLM systems, and NVIDIA GPU computing. I translate customer requirements into production-ready, on-premises deployments on Linux, Kubernetes, PyTorch, CUDA, TensorRT-LLM, vLLM, and Triton—from proof of concept (PoC) through troubleshooting, acceptance, and handover.

Portfolio · PR wall · LinkedIn · CV

What I work on

  • AI solutions architecture: turn customer requirements and site constraints into GenAI/LLM deployment plans, acceptance criteria, and production handover.
  • GPU inference: benchmark and debug serving, communication, quantization, and kernel paths across vLLM, TensorRT-LLM, SGLang, Triton, Dynamo, and FlashInfer.
  • Upstream correctness: 18 merged/landed changes, with the complete live record on prs.wayne.is-a.dev.

Representative merged work

Project Change
FlashInfer Fixed an SM120/121 multi-CTA radix top-k stream hang and NVFP4 attention correction layout
vLLM Made tokenizers survive pickling and fixed streaming message metadata
PyTorch Restored the dropped root argument in nccl.broadcast
Dynamo Cancelled in-flight KV-router recovery when a worker is removed and resolved aggregate planner workers by DGD component type

Selected projects

Project Evidence
trtllm-triton-serving TensorRT-LLM vs vLLM on H100; 12 controlled studies
tensor-core-from-scratch 10 CUDA matmul kernels from naive to Tensor Cores
inference-kernel-cookbook Flash Attention, KV cache, and paged attention from scratch
nccl-collectives-bench NCCL bandwidth, latency, NVLS, and TP-decode limits
nim-agent-blueprint NVIDIA NIM agentic RAG with evaluation and observability
llm-security-lab Reproducible LLM attacks and defenses

Technical writing

Pinned Loading

  1. trtllm-triton-serving trtllm-triton-serving Public

    TensorRT-LLM vs vLLM controlled head-to-head on H100 — 12 studies including a knob-by-knob waterfall reproducing NVIDIA's published 27.7k tok/s (100.3%) and attributing the gap to real serving, plu…

    Python

  2. tensor-core-from-scratch tensor-core-from-scratch Public

    From naive matmul to tensor cores on NVIDIA Blackwell — step by step. 8 self-contained CUDA kernels, each benchmarked against cuBLAS.

    Cuda 1 1

  3. inference-kernel-cookbook inference-kernel-cookbook Public

    LLM inference techniques from scratch — Flash Attention, KV Cache, Paged Attention, each in one self-contained CUDA file. Benchmarked on Blackwell.

    Cuda

  4. nccl-collectives-bench nccl-collectives-bench Public

    NCCL collective benchmarks on an 8×H100 NVSwitch host — busbw vs link budget, NVLS/Ring/Tree, small-message latency floors (eager vs CUDA Graph vs symmetric memory), and the TP-decode comms ceiling…

    Python

  5. nim-agent-blueprint nim-agent-blueprint Public

    Agentic RAG hallucination evaluation on adversarial SQuAD 2.0 (N=200) — nine gate methods compared (self/cross-family/70B judges, PoLL, CoT, MiniCheck, semantic entropy): grounding beats capacity, …

    Python

  6. llm-security-lab llm-security-lab Public

    Learn LLM security by building attacks and defenses from first principles. System prompt extraction, prompt injection, model extraction — each in one runnable Python file.

    Python