Hygon DCU Long-Context Attention: From page784 to page832
AbstractIn large-model inference, the Attention formula is usually not the hardest part; what really determines performance is often how data lands in memory and at what granularity the hardware can fetch that data at once. This article uses Qwen3.5’s official prefix-off long-context path as an e...
Qwen3.5 Inference Performance Engineering on Hygon DCU: From Prefill/Decode Hotspots to End-to-End Gains
AbstractIn large-model inference optimization, one of the easiest things to misinterpret is treating a local speedup of a certain kernel directly as an improvement in service throughput. Even if a kernel is 1.08x faster in isolated tests, the complete service may be only a few percent faster, or ...
Hygon DCU Profile in Practice: How to Reconstruct a Token from Sixty Thousand GPU Kernels
OpeningWhen profiling large models on Hygon DCU, the hardest part is usually not whether a trace can be generated, but how to read it after it has been generated. A local sampling run of Qwen3.5-27B BF16 produced: 1234Total events: 235,964GPU events: 61,804kernel events: 60,124GPU event total duration: 5773.493...
Build Success Does Not Mean It Takes Effect — Verifying Incremental vLLM Compilation on Hygon DCU
The biggest waste of time I encountered while modifying vLLM kernels on Hygon DCU was not a compilation error. That time, compilation ended normally, the service could start, the endpoint could return results, and the performance test even produced a change that looked explainable. Only after rec...
THU-BDC2026 Experiment Log - Those Optimizations That Made the Results Worse and Worse
A public version experiment record covering single-period high return, data leakage fix, feature pruning, money flow factors, and several failed optimizations.
THU-BDC2026 Technical Route: From Transformer to LightGBM Ranking
Why Write a Separate Technical Route PostThe previous post was more like a competition retrospective; this one is specifically to sort out the technical route for THU-BDC2026. For this project, from the earliest deep model to the later LightGBM LambdaRank, the transition was not simply “switching...
THU-BDC2026 Retrospective: I Almost Got Fooled by a Beautiful Backtest
OriginFor this THU-BDC2026, I actually kept three project directories: the original, archive, and v2. They look like three projects, but after going through them, a more accurate description is: this is the three-stage growth history of the same competition project. The problem itself is very dir...
Xiaomi MIMO Token Application: My Real Experience of Using AI/Agent to Write Projects
BackgroundRecently, while organizing my projects, I also dug out the application materials for the Xiaomi MIMO developer token. These materials were originally meant to explain what I had built with AI/Agent, but after writing them, I found they were also quite suitable as a retrospective: W...
Building an E-commerce Platform with Servlet and JSP: A Bit Old-School, but It Really Trains the Fundamentals
BackgroundWriting Servlet + JSP again nowadays feels a bit retro. The frontend world has long been full of React and Vue, and on the backend most projects start with Spring Boot. But I still think that writing an e-commerce demo by hand with Servlet and JSP is a great way to shore up Web fundamen...
Campus Dormitory Management System: Even a CRUD Project Can Be Full of Pitfalls
How It StartedA campus dormitory management system sounds a lot like a course design project. Student information, dorm numbers, repair tickets, announcement management—these features are all very common. At first I also thought it was just a CRUD project: write a few APIs, connect a database, an...
FIELD NOTES01 / 02
This page—we will stop here for now.
Turn to the next page and keep picking up fragments of code and the journey.
Recent Posts