Charles Xu

I made HuggingFace tokenizers 3.3x faster by parallelizing single-input tokenization with overlapping chunks, zero-copy offset operations, SIMD-accelerated boundary detection, and cache-hierarchy-aware chunking.

Many inference speedup techniques mirror the classic systems regime—such as caching, paging, tiling, pipelining, and speculative execution (e.g. branch prediction and cache prefetch).

In the multi-head attention mechanism, why after reshaping the projection matrices for Q/K/V from 3 dimensions to 4, we need to transpose the tokens dimension with the heads dimension?

Careful planning and execution work for general management but not for startups.

Before Fundraise: Allow minimum three to six months to raise money.

In a large, busy cluster, how do you scale Istio to address Istio-proxy Container being OOM-Killed and Istiod crashes if too many connected istio-proxies?

AWS EKS on VPC networks need AWS Security Group Rules (SG) to receipt ingress traffic.

Your Kubernetes cluster probably has a shared ingress for north-south traffic, coming from a cloud load balancer and lands on your favorite proxies like Envoy, or Istio gateways, or Nginx.

How to do product-led growth and hands-on outbound sales at the same time?

Make sure to check out the previous post: Advices I wish I got at the start of my career.