first author
ZenFlow: Stall-Free Offloading Training via Asynchronous Updates
preprint · 2025
Async updates kill the offloading stall — train bigger models on the same GPUs.
♪ tingfeng (felix) lan presents —
Isn't it so fun?
PhD candidate at DS² Lab, University of Virginia, with Yue Cheng; also working closely with Juncheng Yang at Harvard. Building systems across the LLM stack — training, inference, storage — with a soft spot for the I/O bottleneck. When I'm not chasing throughput, I'm chasing bebop.
✦ selected papers
co-author
mLoRA: Fine-Tuning LoRA Adapters via Pipeline Parallelism
VLDB '25
Pipeline-parallel LoRA training across multiple GPUs.
co-first
DLRover-RM: Resource Optimization for Deep Recommendation Models in the Cloud
VLDB '24
Resource autoscaling that understands recommendation training workloads.
first author
TStore: Rethinking AI Model Hub with Tensor-Centric Compression
preprint · 2026
A tensor-centric storage layer for AI model hubs — compressing checkpoints by exploiting their internal structure.
co-author
ZipLLM: Efficient LLM Storage via Model-Aware Deduplication and Compression
NSDI '26
Synergistic dedup + compression tuned to how LLM weights actually look on disk.
co-author
MorphServe: Workload-Aware LLM Serving via Runtime Layer Swapping and KV Cache Resizing
MLSys '26
Adapt the serving stack at runtime — swap layers, resize KV cache, ride the workload.
co-author
λScale: Fast Scaling for Serverless LLM Inference
MLSys '26
Cold start is no longer a death sentence for serverless LLMs.
co-author
Scorpio: Serving the Right Requests at the Right Time for Heterogeneous SLOs in LLM Inference
preprint · 2025
SLO-aware LLM serving — TTFT/TPOT guards with credit-based batching for workloads with heterogeneous deadlines.
co-author
IGenBench: Benchmarking the Reliability of Text-to-Infographic Generation
preprint · 2026
First benchmark for text-to-infographic generation — 600 tests across 30 infographic types, automated reliability checks via atomic yes/no questions.
co-author
Demonstrating ViviDoc: Generating Interactive Documents through Human-Agent Collaboration
preprint · 2026
Human-agent system for interactive educational documents — multi-agent pipeline (Planner / Executor / Evaluator) plus a human-readable DocSpec IR.
✦ projects / systems
training · first author
ZenFlow
Stall-free async offloading for LLM training. Integrated into DeepSpeed via official PR.
storage · first author
TStore
Tensor-centric storage layer for AI model hubs.
storage · NSDI '26
ZipLLM
Model-aware dedup + compression tuned to how LLM weights actually look on disk.
inference · MLSys '26
MorphServe
Adapt the serving stack at runtime — swap layers, resize KV cache, ride the workload.
inference · MLSys '26
λScale
Fast scaling for serverless LLM inference.
vis · preprint
IGenBench
First benchmark for text-to-infographic generation reliability.
✦ news
2026.05
🎉🎉 Our papers “IGenBench” and “Demonstrating ViviDoc” are accepted by ACL and ACL Demo!
2026.04
🎉🎉 Honored to receive the EuroSys Distinguished AEC Award!
2026.04
💡💡 Honored to serve on the OSDI’26 Artifact Evaluation Committee (AEC)!
2026.01
2025.09
💡💡 Thrilled to receive a grant from Modal for Academics — big thanks to Modal!