LoRA (Low-Rank Adaptation) freezes the original weights and trains a few small new matrices alongside them — usually under 1 % of the parameters.
It needs much less memory, trains faster, and the result is an adapter of a few megabytes that is easy to share.
Example
A LoRA fine-tune of Llama 3.1 8B fits on a 24 GB L4; a full fine-tune would need well over 100 GB.
বাংলায়: LoRA — পুরো মডেল না বদলে ছোট “অ্যাডাপ্টার” শিখিয়ে ফাইন-টিউন করার কৌশল — কম মেমরি, কম খরচ।