# Sequence Packing & Variable-Length Attention: Throughput'u %40 Artıran Trick

> Source: https://sukruyusufkaya.com/learn/fine-tuning-cookbook/ftc-sequence-packing-varlen-attention
> Updated: 2026-08-16T14:41:11.349Z
> Category: Fine-Tuning Cookbook (Model-by-Model)
> Module: Part II — Tokenizer & Data Engineering
**TLDR:** Padding token'lar boşa giden compute. Packing: birden fazla kısa örneği tek sequence'a doldur. Variable-length attention (flash_attn_varlen_func) ile block-diagonal mask. TRL SFTTrainer packing=True internals, cu_seqlens tensor anatomisi, throughput bench (Llama 3.1 8B RTX 4090: 7290 → 10200 tokens/s).