Yuan Li
ML Acceleration @ Annapurna Labs (AWS) | CMU Alumi
- Role
- Software Engineer - Ai Ml, Aws Neuron Distributed Training at Amazon Web Services (AWS)
- Location
- Seattle, WA, US
- LinkedIn followers
- 500 followers
About Yuan Li
Focus on Deep Learning and Machine Learning Systems
Experience
Software Engineer - Ai Ml, Aws Neuron Distributed Training
Feb 2025 — Present · Seattle, WA, US
Accelerating large-scale foundation model training on AWS EC2 Trn2 instances via graph-compiler awared optimizations, design of distributed sharding strategies, and kernel/framework perf tuning.Designed experimental distributed training benchmarks on PyTorch top-of-tree wheels to evaluate module/kernel level performance, enabling compilation and runtime level analysis for ops lowering and custom kernel fusion strategiesConducted roofline analysis for 2-D sharded distributed training workloads under Tensor Parallelism and Fully Sharded Data Parallelism (FSDP) using PyTorch Distributed (DTensor) on Qwen3-8B; optimized tensor sharding and kernel fusion pipelines, achieving 1.8× MFU utilization on AWS EC2 Trn2 instances.Investigated memory inefficiencies in selective gradient checkpointing and activation recomputation by analyzing PyTorch XLA compilation graph dumps; inserted compiler optimization barriers to prevent gradient delay propagation, reducing peak memory during the backward pass by 40%.Diagnosed and resolved performance regressions caused by mis-lowered mixed-precision matmuls via PyTorch Profiler; implemented custom lowering hints for torch.neuron device backend, improving end-to-end training step time by 25%.
Education
Zhejiang University
Bachelor of Science - BS
2019 — 2023
Carnegie Mellon University
Master of Science - MS
2023 — 2024
Find verified contacts for anyone on LinkedIn
Unifers gives sales teams verified emails and direct dials, enriched profiles, and outreach that lands in the inbox.
Free plan included · No credit card required
This profile is compiled from publicly available professional sources. Unifers is not affiliated with or endorsed by LinkedIn. Request removal of this profile.