Hi, I'm Beichen Huang 👋

MSCS @ UIUC · Working on LLM inference systems — quantization, efficient serving, and system–algorithm co-design.

Open Source

vLLM

• Improving LoRA correctness and robustness for RL training workflows
• Fixing LoRA kernel numerical issues on Hopper (SM90) GPUs
• Extending LoRA support to prompt logprobs and multimodal models

LLM Compressor

• Built block-wise quantization-aware distillation (QAD) to recover accuracy in low-bit models
• Working on low-precision quantization schemes such as W4A16 and NVFP4
• Evaluating quantized models on downstream and agentic benchmarks

Contact

📫 polarishuang0.0@gmail.com