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...
Recent Posts