chenk.top
中文

LLM Engineering

End-to-end modern LLM stack: architectures, post-training, inference, RAG, evaluation, safety, and production.

12 articles

  1. 01

    LLM Engineering (1): Architectures from Transformer to MoE

    MHA → GQA → MQA, sparse MoE routing in Mixtral and Qwen3-MoE, sliding-window attention, and the state-space alternatives …

    56 min
  2. 02

    LLM Engineering (2): Tokenization Deep Dive

    BPE vs SentencePiece vs WordPiece, byte-level fallback, the CJK token-bloat problem, vocabulary expansion costs, and the …

    18 min
  3. 03

    LLM Engineering (3): Pretraining at Scale

    Data mixing, deduplication, contamination, μP, FSDP vs ZeRO-3 vs pipeline parallel, the practical 200B-token cliff, and …

    42 min
  4. 04

    LLM Engineering (4): Post-training — SFT, DPO, RLHF, RLAIF

    What SFT, DPO, RLHF, and RLAIF each actually optimize, when reward models fail, KL constraints, the LoRA-vs-full-FT …

    48 min
  5. 05

    LLM Engineering (5): Inference Optimization

    KV cache mechanics, paged attention, continuous batching, speculative decoding, INT8/INT4/AWQ/GPTQ quantization, and the …

    42 min
  6. 06

    LLM Engineering (6): Long Context — RoPE, YaRN, Sinks

    How RoPE encodes position, why naive extension breaks, NTK-aware and YaRN scaling, ALiBi vs RoPE, attention sinks for …

    34 min
  7. 07

    LLM Engineering (7): Function Calling and Tool Use

    JSON-mode vs function-mode vs free-form, parallel tool calls, structured-output guarantees with grammars, error recovery …

    34 min
  8. 08

    LLM Engineering (8): Retrieval-Augmented Generation

    Chunking strategies, dense vs sparse vs hybrid retrieval, reranker selection, the long-context-vs-RAG tradeoff in 2026, …

    30 min
  9. 09

    LLM Engineering (9): Prompting at Production Scale

    Chain-of-thought when it actually helps, self-consistency, prompt-caching economics, jailbreak taxonomy, …

    40 min
  10. 10

    LLM Engineering (10): Evaluation

    Why MMLU is broken, the contamination problem, LLM-as-judge biases, position-bias mitigation, calibration, and the A/B …

    36 min
  11. 11

    LLM Engineering (11): Safety and Alignment

    What alignment means engineering-wise, refusal calibration, the red-team taxonomy, hallucination metrics, sleeper …

    36 min
  12. 12

    LLM Engineering (12): Production — Deployment, Monitoring, Cost

    Serving stack choices in detail, autoscaling LLMs, latency budgets, prompt+completion cost tracking, multi-model …

    36 min