Exciting Breakthrough in Extreme Quantization - HF Cracks Code of BitNet - HF1BitLLM
ฝัง
- เผยแพร่เมื่อ 21 ก.ย. 2024
- This video shows how to locally install Llama3-8B-1.58-100B-tokens in which Hugging Face has finally cracked the code for BitNet - no pretraining needed with warmup quantization! By fine-tuning a Llama3 8B model, they’ve achieved results close to Llama 1 & 2 7B on key downstream tasks.
🔥 Buy Me a Coffee to support the channel: ko-fi.com/fahd...
🔥 Get 50% Discount on any A6000 or A5000 GPU rental, use following link and coupon:
bit.ly/fahd-mirza
Coupon code: FahdMirza
▶ Become a Patron 🔥 - / fahdmirza
#bitnet #warmupquantization #hf1bitllm
PLEASE FOLLOW ME:
▶ LinkedIn: / fahdmirza
▶ TH-cam: / @fahdmirza
▶ Blog: www.fahdmirza.com
RELATED VIDEOS:
▶ Resource huggingface.co...
All rights reserved © Fahd Mirza
A great solution for those with lighter weight computer systems. Aesome research!
It indeed is.
Great video, as always! My brain is doing somersaults upon hearing about the concept of 1.58 bits! 😄 Since I always imagined 'bits' to be binary by definition. Whatever they are doing sounds fascinating. If my math is correct, a 70B model in 1.58Q should be around 16 gigabytes in size, which means it will be viable for some commercially available GPUs. I'm really curious how a Q1.58 version of such large models will actually perform in practice.