Technical reports with stated environments, methods, measured results, and scope. English and Chinese editions are linked from the featured work where available.
Compares the MinerU 3.4.4 production baseline with the 4.0.7 evaluation on gfx906, including measured throughput, deployment constraints, and batch-processing results.
PaddleOCR 3.7.0 and PaddleX 3.7.2 complete end-to-end validation on one gfx906 card; the 23-page PDF runs at 0.49 pages/s in steady state.
Full paper covering the limiter formulation, mesh tests, error analysis, figures, and references in a typeset online edition.
A reproducible three-page XeLaTeX validation covering bilingual text, mathematics, figures, code, cross-references, and the Biber bibliography workflow.
A technical account of the vLLM 0.27.x port for DeepSeek V4 Flash FP4 on gfx936, covering the root cause, patch history, validation evidence, performance, and batch-processing limits.
Build and correctness results, multi-GPU scaling, hotspot optimization, and DTK compatibility for Kokkos/HIP on gfx936 and gfx906/Z100.
ABI adaptation, build debugging, algorithm validation, and HIP/OpenCL comparisons across two Hygon DCU generations and DTK releases.
Build fixes, correctness checks, and controlled performance comparisons between native HIP and the CUDA translation path on Hygon Z100.