Peiye Liu

I work with the TikTok Core Recommendation team on the Large Recommendation Model (LRM) direction.

I lead low-level communication and compute kernel development for large-scale LLM training and inference acceleration. LLM Acceleration & Hardware Co-Design Built a high-performance kernel library for a novel dataflow computing architecture, significantly improving LLM training and inference throughput. Designed a CUDA-like interface with ATen-style APIs, ensuring seamless integration with compilers such as ONNX, TVM, and MLIR. Analyzed QWEN-1/2 using Nsight Systems and a custom roofline model, delivering 30+ fused kernels and achieving 80% computational efficiency. Contributed system-level design insights for LLM-oriented chip development, ensuring compatibility with DP, MP, and ZeRO parallelism. Developed a layer-adaptive early-exit strategy for QWEN, tripling inference efficiency without accuracy loss. Designed a dynamic scheduling framework for on-chip PEs to support runtime-adaptive execution in large models. Proposed a hardware-friendly, block-wise sparse compression method to align memory layouts with optimal utilization patterns.

Neural Architecture Search (NAS) & Model Optimization Developed a scalable evolutionary NAS framework that reduced CNN memory usage by 50% while improving ImageNet top-1 accuracy by 1.1%. Designed a peer-based structure correlation controller, accelerating search convergence and improving candidate ranking accuracy by 20%. Built a DARTS-based architecture tailored for homomorphic encryption, reducing memory usage by 60%. Integrated a chip-level cost roofline model as a reward signal to discover an optimized QWEN sub-block, reducing memory by 30% without performance loss. Low-Level Vision & Diffusion Models Developed a structure-preserving tone mapping pipeline for HDR-to-SDR conversion, maintaining visual fidelity across dynamic ranges. Designed a dual-control stable diffusion model for enhanced luminance retention and detailed restoration in low-level vision tasks. Proposed a decoupled zero-shot training strategy within the diffusion process, achieving a 30% gain in unsupervised evaluation metrics.