writing

A series on GPT-2 inference in NumPy — batching, padding, KV cache, sampling — plus notes on evaluation and fine-tuning on Azure.

Three threads.

GPT-2 inference extends Jay Mody’s picoGPT with the machinery real serving needs — batching, padding, a KV cache, sampling, then scheduling and quantization — proving at each step that the faster version still agrees with the slower one. Code lives here - github.com/firozgit/picoGPT.

Evaluation is the day job: benchmarks, judge models, assertion design, and what failure analysis actually looks like at scale.

Fine-tuning on Azure is a set of walkthroughs from 2025, comparing the low-code path through AI Foundry against full control in Azure ML — code for those lives in github.com/firozgit/LLM-Finetuning.

Posts tagged medium are hosted on Medium and open in a new tab. Everything else lives here.