---
title: "schieh710 (@swchieh) | Hashnode"
description: "AI/ML Infrastructure | System Design | MLOps. Read the latest articles by schieh710 on Hashnode."
url: "https://hashnode.com/@swchieh"
---

@swchieh

# schieh710

@swchieh · San Jose, USA · Joined May 2026

AI/ML Infrastructure | System Design | MLOps

### [I Built a 9× Faster LLM Memory Allocator in C++ — and Worked Out When It Pays Off](https://seng-wei-chieh.hashnode.dev/i-built-a-9-faster-llm-memory-allocator-in-c-and-worked-out-when-it-pays-off?utm_source=hashnode&utm_medium=feed)

A paged KV-cache allocator for Python, with every result checked against predictions I wrote before coding

18 min read · [seng-wei-chieh.hashnode.dev](https://seng-wei-chieh.hashnode.dev/i-built-a-9-faster-llm-memory-allocator-in-c-and-worked-out-when-it-pays-off?utm_source=hashnode&utm_medium=feed)

[Sep 25](https://hashnode.com/posts/i-built-a-9-faster-llm-memory-allocator-in-c-and-worked-out-when-it-pays-off/6ab6065e7b4232acdc17b980)

### [Same GPU, Same Impossible Job: One Engine Crashed, the Other Slowed Down 8%](https://seng-wei-chieh.hashnode.dev/same-gpu-same-impossible-job-one-engine-crashed-the-other-slowed-down-8?utm_source=hashnode&utm_medium=feed)

Measuring stock HuggingFace against vLLM on one 8 GB laptop GPU — including two results that contradicted what I expected

9 min read · [seng-wei-chieh.hashnode.dev](https://seng-wei-chieh.hashnode.dev/same-gpu-same-impossible-job-one-engine-crashed-the-other-slowed-down-8?utm_source=hashnode&utm_medium=feed)

[Sep 5](https://hashnode.com/posts/same-gpu-same-impossible-job-one-engine-crashed-the-other-slowed-down-8/6a9c5bb0a7774286a604288e)

### [5.1× From Overlapping, 11.6× From Processes, 0× From Fewer Bytes](https://seng-wei-chieh.hashnode.dev/5-1-from-overlapping-11-6-from-processes-0-from-fewer-bytes?utm_source=hashnode&utm_medium=feed)

Four ways to read 10,000 files out of object storage, and which one did nothing

14 min read · [seng-wei-chieh.hashnode.dev](https://seng-wei-chieh.hashnode.dev/5-1-from-overlapping-11-6-from-processes-0-from-fewer-bytes?utm_source=hashnode&utm_medium=feed)

[Aug 1](https://hashnode.com/posts/5-1-from-overlapping-11-6-from-processes-0-from-fewer-bytes/6a6e683be34c4ebf1434fe95)

### [I Replaced a Flink Cluster With One Python Process](https://seng-wei-chieh.hashnode.dev/i-replaced-a-flink-cluster-with-one-python-process?utm_source=hashnode&utm_medium=feed)

Every dependency is a liability you pay whether or not you collect its benefit.

9 min read · [seng-wei-chieh.hashnode.dev](https://seng-wei-chieh.hashnode.dev/i-replaced-a-flink-cluster-with-one-python-process?utm_source=hashnode&utm_medium=feed)

[Jun 8](https://hashnode.com/posts/i-replaced-a-flink-cluster-with-one-python-process/6a26173f130f85eb073b5cda)

### [The GIL Fingerprint: Why My Kafka Producer Stalled at 14k events/sec](https://seng-wei-chieh.hashnode.dev/the-gil-fingerprint-why-my-kafka-producer-stalled-at-14k-events-sec?utm_source=hashnode&utm_medium=feed)

One CPython fact explained three mysteries: an unmovable plateau, five failed fixes, and the thread count that finally worked.

9 min read · [seng-wei-chieh.hashnode.dev](https://seng-wei-chieh.hashnode.dev/the-gil-fingerprint-why-my-kafka-producer-stalled-at-14k-events-sec?utm_source=hashnode&utm_medium=feed)

[May 16](https://hashnode.com/posts/the-gil-fingerprint-why-my-kafka-producer-stalled-at-14k-events-sec/6a07f4efea8975272b798a1e)