https://wanli-jiang.github.io/Wanli JiangA blog to record my learning and research. 2026-08-18T13:11:33+00:00 Wanli Jiang https://wanli-jiang.github.io/ Jekyll © 2026 Wanli Jiang /assets/img/favicons/favicon.ico /assets/img/favicons/favicon-96x96.png 廉价生成之后：AI Agent 与再生式软件工程2026-08-18T00:00:00+00:00 2026-08-18T00:00:00+00:00 https://wanli-jiang.github.io/posts/AI-agent-Cheap-generation-and-regenerative-software-engineering/ Wanli Jiang 当代码、实验和方案的生成成本快速下降，工程的稀缺资源会转向反馈、验证、责任与长期判断力。本文提出“再生式软件工程”：一次自动化不仅要完成任务，还应积累验证资本、人类能力资本与架构选择权，让下一次变化更容易理解、验证、恢复和重新选择。 Removing the Legacy TensorRT Backend from TensorRT-LLM: A Retrospective2026-07-23T00:00:00+00:00 2026-07-23T00:00:00+00:00 https://wanli-jiang.github.io/posts/trtllm-maintain-Removing-legacy-TensorRT-backend-from-TensorRT-LLM/ Wanli Jiang A retrospective on removing the legacy TensorRT engine backend from TensorRT-LLM through small, reviewable, always-compilable pull requests. Piecewise CUDA Graph and Breakable CUDA Graph for LLM Inference2026-07-06T00:00:00+00:00 2026-07-25T00:52:58+00:00 https://wanli-jiang.github.io/posts/LLM-optimization-Piecewise-CUDA-Graph-and-Breakable-CUDA-Graph-for-LLM-Inference/ Wanli Jiang A technical note on full, piecewise, and breakable CUDA Graphs in modern LLM and VLM inference systems. When NVLS Is "Supported" on H100, Why Does AllReduce Take Another Path?2026-06-23T00:00:00+00:00 2026-06-23T00:00:00+00:00 https://wanli-jiang.github.io/posts/LLM-optimization-NVLS-supported-on-H100-why-AllReduce-takes-another-path/ Wanli Jiang A TensorRT-LLM debugging story about separating static NVLS capability from fabric handle usability and single-node POSIX-FD allocation. GDPval Bench introduction2026-06-01T00:00:00+00:00 2026-06-01T00:00:00+00:00 https://wanli-jiang.github.io/posts/LLM-bench-GDPval-Bench-introduction/ Wanli Jiang An exploration of GDPval, OpenAI's benchmark for evaluating AI models on realistic, economically valuable knowledge-work tasks.