- Serving LLM
- FlashAttention → tối ưu kernel attention ở mức CUDA/GPU
- PagedAttention → quản lý KV cache hiệu quả
- vLLM → inference engine dùng PagedAttention
- TensorRT-LLM → NVIDIA stack tối ưu latency/throughput
- SGLang → orchestration + structured generation/agent runtime
- llama.cpp → inference local trên CPU/mobile
- Transformer in LLM old and new
- Topics transformer old and new: qweb mistral, gpt,
- Dataset trick
- Training trick
- Evaluation
- Tokernizer Topic
- Attention topic
- Before Transformer topic
- Transformers
- After Transformers
May 12, 2026
Topics LLM to do
A practical explanation of Incremental Static Regeneration (ISR) and when to use it.