╔══════════════════════════════════════════════════╗
║ almaz@amsterdam:~$ whoami ║
║ ║
║ Senior data engineer, transitioning to MLOps ║
║ and ML infrastructure engineering. ║
║ ║
║ status: training | measuring | writing it down ║
╚══════════════════════════════════════════════════╝
- Language models - the whole path from raw text to a trained model: data pipeline, tokenizer, training loop, benchmarking, evaluation, and fancy RLVRs.
- PyTorch exercises for LLM developers - rustlings-style drills, from tensor basics to a KV cache.
- PyTorch Distributed exercises for LLM developers - in progress: the same format for distributed training and serving.
focus = {
"now": ["pretraining", "MFU and FLOP accounting", "distributed PyTorch"],
"next": ["inference as a storage hierarchy", "KV cache tiering"],
"day job": ["CICD", "Databricks", "dbt", "Terraform", "GCP"],
}- almaz.murzabekov.net - notes on what broke, with the numbers behind it.
- nookat - a lightweight container manager in Tauri, Rust and React.
If you are working on the same seam between data engineering and ML infrastructure, open an issue on any of these repos.