Deep Dives

How to Choose an LLM Inference Engine — A 2026 Map from Local Single-GPU to PD Disaggregation

2026-07-19·3568 words· 17 min

Deep Dives LLM Inference Engine VLLM SGLang Model Serving Inference Optimization Selection Guide Research

Aliyun’s CAP has a piece on picking an inference engine that narrows the field to four: Ollama, vLLM, SGLang, and Hugging Face Pipeline. In 2024, that framing was fine. By 2026, it’s missing half the map. NVIDIA’s TensorRT-LLM has completed its “PyTorch-ification,” SGLang became famous as the first open-source project to reproduce DeepSeek’s large-scale deployment, Hugging Face slapped a “maintenance mode” banner on TGI and told you to switch to vLLM — and the real throughline of the entire 2025 inference landscape can be summed up in one word: disaggregate.

Why LLMs Can't Remember You — Cursor / Claude Code / Codex Memory Mechanisms Dissected

2026-05-04·更新于: 2026-07-09·4692 words· 23 min

Deep Dives AI Agent LLM Memory Research Context Engineering

Cursor has Rules + Memories. Claude Code has CLAUDE.md + MEMORY.md. Codex has AGENTS.md + Memories. These Agent tools all have their own “memory” systems. But not a single one actually modifies model weights — every “memory” feature works by injecting structured text back into the system prompt. This report cross-validates that conclusion with 67+ primary sources, tearing down the architecture of Agent memory systems from first principles to product implementation.

↑