Structural Conventions: Why Claude Loves XML and GPT Loves JSON
Different models respond better to different formatting conventions. Here's what actually works across Claude, GPT, Llama, and others.
Writing
View allChat Templates: The Hidden Interface Layer
The same prompt behaves differently across models. The reason isn't the model weights - it's the chat template. Here's what you need to know.
Continuous Batching: Why LLM Serving Is Not Like Traditional Inference
Static batching wastes GPU cycles waiting for the slowest request. Continuous batching fixes this by adding and removing requests at every iteration. Here's how it works.
Quantization for LLM Inference: From FP16 to INT4
Quantization cuts memory and speeds up inference. But naive 8-bit quantization breaks at 6.7B+ parameters. Here's why, and how modern methods fix it.
Speculative Decoding: Draft Once, Verify in Parallel
LLM decoding is sequential and memory-bound. Speculative decoding breaks this by guessing multiple tokens and verifying in parallel. Here's how it works and when to use it.
Tensor Parallelism: How Large Models Fit Across GPUs
Data parallelism hits a wall when your model doesn't fit on one GPU. Tensor parallelism solves this by sharding the model itself.
Why Your GPU Utilization Is Lower Than You Think
A batch size of 1 on an A100 typically achieves 10-20% utilization. Understanding why—and how to fix it—is key to cost-effective inference.
The Two Phases of LLM Inference: Prefill and Decode
Time-to-first-token is compute-bound. Token generation is memory-bound. Understanding this split is key to optimizing inference.
Flash Attention: Why Memory Access Patterns Matter More Than FLOPs
Flash Attention doesn't reduce computation—it reduces memory traffic. Understanding why that matters is key to optimizing transformers.
MHA vs GQA vs MLA: A Visual Guide to Attention Mechanisms
Multi-Head, Grouped-Query, and Multi-Head Latent Attention explained with memory calculations you can verify yourself.
Mixture of Experts: The Scaling Strategy Behind Modern LLMs
DeepSeek uses 256 experts, Llama 4 uses 8,192. Why expert count matters, and what the tradeoffs actually are.
Pre-Norm vs Post-Norm: Why Normalization Placement Matters
GPT-2 popularized Pre-Norm. Now models like OLMo 2 and Gemma 3 are switching back to Post-Norm. What changed?
Linear Attention: Is O(n) Worth the Accuracy Tradeoff?
Standard attention is O(n²). Linear variants promise O(n). Qwen3 and Kimi K2 use hybrid approaches. Here's what you give up.