https://caiyf03.github.io/tags/lgtorch/LGTorchHugoBlox Kit (https://hugoblox.com)en-usFri, 15 May 2026 00:00:00 +0000https://caiyf03.github.io/media/icon_hu_702a800cd775dbac.pnghttps://caiyf03.github.io/tags/lgtorch/https://caiyf03.github.io/internships/zijie/Fri, 15 May 2026 00:00:00 +0000https://caiyf03.github.io/internships/zijie/<h2 id="overview">Overview</h2> <p><strong>Model Engineering Intern</strong> <em>May 2026 – Present</em></p> <h3 id="recommendation-model-migration">Recommendation Model Migration</h3> <ul> <li>Reconstructed and upgraded ByteDance&rsquo;s internal RL-based retrieval recommendation models from TensorFlow to LGTorch, including dual-tower architectures, vector quantization (VQ), and multi-runstep graph execution pipelines.</li> <li>Designed sequence encoding tower decomposition, weight normalization alignment strategies, and write-back mechanisms for frequency estimation modules.</li> <li>Completed full-stack migration across sparse feature processing, dense computation, and online serving pipelines.</li> </ul> <h3 id="precision-alignment-infrastructure">Precision Alignment Infrastructure</h3> <ul> <li>Built a bidirectional forward/backward diff toolchain for end-to-end TensorFlow–LGTorch numerical verification.</li> <li>Resolved feature-slot registration mismatches, VQ codebook numerical deviations, and framework-specific execution inconsistencies.</li> <li>Achieved operator-level precision alignment between both frameworks; the associated technical documentation was classified as internal L3 confidential.</li> </ul> <h3 id="production-model-upgrade">Production Model Upgrade</h3> <ul> <li>Upgraded Douyin’s RL-based retrieval pipeline with logQ debiasing and improved sampled-softmax estimation over sharded candidate repositories.</li> <li>Achieved online performance comparable to the production baseline and successfully supported full-scale deployment.</li> </ul> <hr>