Deep Dives 11
- vLLM Deep Dive Part 3: Architectures — 60+ Models, What Actually Makes Them Different, and the 2026 Frontier
- vLLM Deep Dive Part 2: Scaling — Speculative Decoding, Parallelism, and Disaggregated Serving
- vLLM Deep Dive: The Engine — How vLLM Turns a Single GPU into a Serving Machine
- LLM Serving in Depth: Batching, Scheduling, and Parallelism
- Speculative Decoding: Generating Multiple Tokens Per Step
- CUDA Kernels and FlashAttention: Why Memory Bandwidth Is the Bottleneck
- A Quantization Primer: Formats, Architecture Sensitivity, and a Gemma 4 Case Study
- Training vs Inference: Why the Same Model Costs 10× More to Train
- The Memory Math: What Fits on a GPU?
- Attention Mechanisms and KV Cache: From First Principles to Gemma 4's Architecture
- Inside LLM Inference: Every Calculation from Text to Token using Gemma 4 12B