·40 min read·AI & LLMs
LLM Architectures · Part 1
Python Internals for LLM Engineers
A walkthrough of what I learnt during Phase 1 of learning about LLM architectures.
AILLMsAI EngineeringLLM Architectures
All posts
Series · 6 posts, in reading order
LLM Architectures
Learning LLM internals from Python to vLLM, following Mentiforce's roadmap and ending in a queue-scheduler project.
·47 min read·AI & LLMs
LLM Architectures · Part 2
Neural Network Fundamentals
A walkthrough of what I learnt on Neural network fundamentals and hardware execution with a broader scope on learning about LLM architectures.
AILLMsAI EngineeringLLM Architectures
·41 min read·AI & LLMs
LLM Architectures · Part 3
Transformer Architecture Principles
A walkthrough of what I learnt on Transformer architecture principles with emphasis on transformer models, attention - multi-head and cross attention, softmax functions and positional encoding methods with a broader scope on learning about LLM architectures.
AILLMsAI EngineeringLLM Architectures
·38 min read·AI & LLMs
LLM Architectures · Part 4
Transformer Inference Optimization Strategies
A walkthrough of what I learnt on Transformer Inference Optimization Strategies with emphasis on inference efficiency, KV cache, flash attention and quantization methods with a broader scope on learning about LLM architectures.
AILLMsAI EngineeringLLM Architectures
·35 min read·AI & LLMs
LLM Architectures · Part 5
vLLM Framework and Inference Engine Design
A walkthrough of what I learnt about vLLM: why it handles high concurrency so well, how PagedAttention manages KV cache memory, how its scheduler mixes prompt and decode work, and how users and applications interact with the engine.
AILLMsAI EngineeringLLM Architectures
·40 min read·AI & LLMs
LLM Architectures · Part 6
Batching LLM Requests: Foundations for a Queue Scheduler
The foundations of my next project, a minimal LLM queue scheduler: why batching matters and what it costs, how requests arrive, how naive, dynamic and continuous batching differ, how to measure TTFT, TPOT and throughput, and the plan for building and benchmarking it.
AILLMsAI EngineeringLLM ArchitecturesProjects