Profile
I am an AI systems engineer and researcher working on distributed systems for large-scale foundation-model development. My work spans pretraining, long-context training, post-training, reinforcement learning, inference, and data generation, with a particular focus on correctness, diagnosability, and resilience.
I enjoy working across system boundaries: translating model and algorithm requirements into distributed execution, identifying failures that do not immediately crash a job, and turning one-off solutions into infrastructure that other researchers can rely on.
Previously, I was a Senior Research Engineer in the System Research Group at Microsoft Research Asia, where I worked on distributed training systems and training correctness. In 2026, I joined an early-stage AI startup, working on infrastructure for large-scale pretraining, post-training, and reinforcement learning.
Selected highlights
- Built and maintained infrastructure supporting internal dense and MoE model training on 2,048–4,096 GPUs, spanning BF16/FP8 pretraining, post-training, and reinforcement learning.
- Core contributor to nnScaler, published at OSDI 2024; contributed from the early development of AutoDist through later system development and model-research integrations.
- Identified and root-caused silent correctness failures in Megatron-LM involving gradient synchronization and MoE auxiliary-loss scaling.
- Acted as a hands-on technical lead for three engineers building RL, rollout, and agent infrastructure used by approximately 20 model researchers and engineers.
Current work
Large-scale model training
In 2026, our infrastructure supported the following internal training programs:
- 8B dense model, trained in BF16 on 2,048 GPUs through pretraining, post-training, and reinforcement learning.
- 30B-A3B MoE model, trained in BF16 on 2,048 GPUs through pretraining, post-training, and reinforcement learning.
- 392B-A15B MoE model, trained in FP8 on 4,096 GPUs.
These models and their technical reports were not publicly released; I describe only the aggregate system scope and my infrastructure responsibilities.
I worked hands-on across distributed parallelization, Megatron-LM maintenance, MoE and FP8 training, communication and computation kernels, correctness debugging, and performance analysis. We also supported long-context extension for the 8B and 30B-A3B models.
Reliability and diagnosis
The training stack supported asynchronous checkpoint saving and recovery in approximately 20 minutes. In one representative investigation, training became slow immediately after resuming and then gradually recovered. After ruling out model computation and communication, I traced the behavior to the data loader, worked with the storage team, and modified the Megatron-LM-based stack to resolve it.
Post-training, reinforcement learning, and agents
I have acted as a hands-on technical lead for three engineers working on reinforcement-learning infrastructure, rollout-engine deployment and optimization, and isolated execution environments for agents and data generation. I participated in technical decisions spanning the training framework, distributed parallelism, rollout engine, and sandbox architecture. The resulting infrastructure served approximately 20 model researchers and engineers.
I have also been developing an agent-assisted engineering loop for monitoring training performance and diagnosing regressions. My longer-term interest is to build agent-assisted, verifiable, and resilient distributed systems for large-scale model development.
Selected technical investigations
Silent gradient corruption caused by CUDA stream affinity
I identified and root-caused a race in Megatron-LM between gradient post-hook operations and reduce-scatter when overlapping gradient reduction with multiple distributed-optimizer instances. An autograd node created on a side CUDA stream caused its hook to inherit that stream affinity, allowing communication to begin before gradient checking and scaling had completed. I provided a minimal reproducer isolating the underlying stream behavior.
GitHub issue · Technical article in Chinese
Incorrect MoE auxiliary-loss scaling under distributed execution
I identified an inconsistency in Megatron-LM's global MoE auxiliary-loss normalization. The implementation averaged over the global token count, reducing the effective gradient contribution by the distributed group size and making the meaning of the auxiliary-loss coefficient depend on the parallel configuration.
GitHub issue · Technical article in Chinese
Together, these cases illustrate a broader concern in large-scale training: job stability and reasonable-looking loss curves are not sufficient evidence of correctness. Execution ordering and distributed normalization can silently change the effective optimization algorithm.
Selected systems work
nnScaler
nnScaler, published at OSDI 2024, compiles a model written for a single GPU into an efficient distributed execution plan.
I was a core contributor, beginning with the early development of AutoDist, which searches for feasible and efficient parallel execution plans while accounting for model memory consumption. I later contributed across the system and supported its use in model-research projects. After the original project lead departed in 2024, I became a technical point of contact for major design discussions and partner integrations, including work on YOCO and Differential Transformer.
TrainVerify
TrainVerify, published at SOSP 2025, verifies semantic equivalence between distributed training executions. I contributed real-world distributed-training bug cases, design feedback grounded in production requirements, and technical support for the nnScaler-based implementation.
Model-system co-design
I have provided distributed-training and system support for model-research projects, translating evolving algorithm requirements into training and inference systems rather than treating the model and system layers independently.
- YOCO — distributed and long-context training support using nnScaler.
- Differential Transformer — distributed-training support for long-sequence model development.
- rStar2-Agent — SFT, reinforcement-learning, rollout, and system support for agentic reasoning.
I have also contributed to rStar-Math, rStar-Coder, VibeVoice, and VibeVoice-ASR. Earlier, during my master's studies, I was an early contributor to OneFlow.
Research interests
Reliable and verifiable large-scale model training; distributed parallel execution; training correctness and silent-failure detection; fault diagnosis, recovery, and system resilience; reinforcement-learning systems; inference, rollout, and data-generation infrastructure; communication and computation kernels; agent-assisted infrastructure development; model-system co-design.
My work has contributed to research published at OSDI, SOSP, NeurIPS, ICLR, and ICML. A selection of publications is listed below.
Education
- Tsinghua University
- M.S. in Software Engineering, 2016–2019
- Beijing Institute of Technology
- B.Eng. in Software Engineering, 2012–2016. Ranked 1st in the Software Engineering program.
Selected publications
- nnScaler: Constraint-Guided Parallelization Plan Generation for Deep Learning Training. OSDI 2024. [code]
- TrainVerify: Equivalence-Based Verification for Distributed LLM Training. SOSP 2025.
- You Only Cache Once: Decoder-Decoder Architectures for Language Models. NeurIPS 2024.
- Differential Transformer. ICLR 2025.
- rStar-Math: Small LLMs Can Master Math Reasoning with Self-Evolved Deep Thinking. ICML 2025. [code]
- rStar-Coder: Scaling Competitive Code Reasoning with a Large-Scale Verified Dataset. 2025. [code]
- rStar2-Agent: Agentic Reasoning Technical Report. 2025. [code]
- VibeVoice: Expressive Podcast Generation with Next-Token Diffusion. ICLR 2026. [code]
- VIBEVOICE-ASR Technical Report. 2026.