LoRA explained (and a bit about precision and quantization)

▬▬ Papers / Resources ▬▬▬
LoRA Paper: arxiv.org/abs/2106.09685
QLoRA Paper: arxiv.org/abs/2305.14314
Huggingface 8bit intro: huggingface.co/blog/hf-bitsan...
PEFT / LoRA Tutorial: www.philschmid.de/fine-tune-f...
Adapter Layers: arxiv.org/pdf/1902.00751.pdf
Prefix Tuning: arxiv.org/abs/2101.00190
▬▬ Support me if you like 🌟
►Link to this channel: bit.ly/3zEqL1W
►Support me on Patreon: bit.ly/2Wed242
►Buy me a coffee on Ko-Fi: bit.ly/3kJYEdl
►E-Mail: [e-mail]
▬▬ Used Music ▬▬▬▬▬▬▬▬▬▬▬
Music from #Uppbeat (free for Creators!):
uppbeat.io/t/danger-lion-x/fl...
License code: M4FRIPCTVNOO4S8F
▬▬ Used Icons ▬▬▬▬▬▬▬▬▬▬
All Icons are from flaticon: www.flaticon.com/authors/freepik
▬▬ Timestamps ▬▬▬▬▬▬▬▬▬▬▬
00:00 Introduction
00:20 Model scaling vs. fine-tuning
00:58 Precision & Quantization
01:30 Representation of floating point numbers
02:15 Model size
02:57 16 bit networks
03:15 Quantization
04:20 FLOPS
05:23 Parameter-efficient fine tuning
07:18 LoRA
08:10 Intrinsic Dimension
09:20 Rank decomposition
11:24 LoRA forward pass
11:49 Scaling factor alpha
13:40 Optimal rank
14:16 Benefits of LoRA
15:20 Implementation
16:25 QLoRA
▬▬ My equipment 💻
- Microphone: amzn.to/3DVqB8H
- Microphone mount: amzn.to/3BWUcOJ
- Monitors: amzn.to/3G2Jjgr
- Monitor mount: amzn.to/3AWGIAY
- Height-adjustable table: amzn.to/3aUysXC
- Ergonomic chair: amzn.to/3phQg7r
- PC case: amzn.to/3jdlI2Y
- GPU: amzn.to/3AWyzwy
- Keyboard: amzn.to/2XskWHP
- Bluelight filter glasses: amzn.to/3pj0fK2

[@khangvutien2538]

This is one of the easiest to follow explanations of LoRA that I’ve seen. Thanks a lot.

[@DeepFindr]

Glad you found it useful!

[@InturnetHaetMachine]

Another great video. I appreciate that you don't skip on giving context and lay a good foundation. Makes understanding a lot easier. Thanks!

[@teleprint-me]

I've been scouring for a video like this. You're the best explanation so far!

[@chrisschrumm6467]

Nice job with summarizing transfer learning and LoRA!

[@aurkom]

Awesome! Waiting for a video on implementing LoRA from scratch in pytorch.

[@k_1_1_2_3_5]

What an excellent video!! Congrats!!

[@mohamedezzat5048]

Thanks a lot Amazing explanation, very clear and straightforward

[@omgwenxx]

Amazing video, feel like I finally understood every aspect of LoRA, thank you!

[@DeepFindr]

Glad it was helpful :)

[@marjanshahi979]

Amazing explanation! Thanks a lot!

[@aron2922]

Another great video, keep it up!

[@beyond_infinity16]

Explained quite well !

[@binfos7434]

Really Helpful!

[@user-in2dd6by9q]

great video to explain lora！ thanks

[@dennislinnert5476]

Amazing!

[@moonly3781]

I'm interested in fine-tuning a Large Language Model to specialize in specific knowledge, for example about fish species, such as which fish can be found in certain seas or which are prohibited from fishing. Could you guide me on how to prepare a dataset for this purpose? Should I structure it as simple input-output pairs (e.g., 'What fish are in the Mediterranean Sea?' -> 'XX fish can be found in the Mediterranean Sea'), or is it better to create a more complex dataset with multiple columns containing various details about each fish species? Any advice on dataset preparation for fine-tuning an LLM in this context would be greatly appreciated.
Thanks in advance!"

[@henrywang4010]

Great video! Liked and subscribed

[@unclecode]

Yes, indeed was hrlpful! Do you have a video on quantization?

[@nomad_3d]

Good summary! Next time it would be great if you add headings to the tables that you show on the video. Sometimes it is hard to follow. For example, what is computational efficiency? is it inference time or inference time increase over the increase in performance (e.g. accuracy, recall, etc.)? Thanks.

[@flecart]

good job!

[@Canbay12]

Thank you very much for this amazing vide. However, although this was probably only for demo purposes of a forward pass after LoRA finetuning; the modified forward pass method you`ve shown might be mislieading; since the forward pass of the function is assumed to be entirely linear. So, does the addition of the LoRA finetuned weights to the base model weights happen directly within model weights file (like .safetensors) or can it be done on a higher level on pytorch or tensorflow?

[@sougatabhattacharya6703]

Good explanation

[@msfasha]

Brilliant

[@ahmadalis1517]

XAI techniques on LLMs is really interesting topic! When you would consider it?

[@ibongamtrang7247]

Thanks

[@darshandv10]

What softwares do you use to make videos?

[@ArunkumarMTamil]

how is Lora fine-tuning track changes from creating two decomposition matrix? How the ΔW is determined?

[@xugefu]

Thanks!

[@DeepFindr]

Thank you!

[@Menor55672]

How do you make the illustrations ?

[@ad_academy]

Good video

[@prashantlawhatre7007]

please make video on QLoRA

[@alkodjdjd]

As clear as mud

[@truck.-kun.]

Sounds like AI

[@anudeepk7390]

Is it a compliment or no? Cause mud is not clear.

[@iloos7457]

​@@anudeepk7390😂😂😂😂😂😂

[@kutilkol]

Ideot read paper. Lol

[@susdoge3767]

gold