Telegram Group & Telegram Channel
✔️ Учёные Яндекса, НИУ ВШЭ,MIT, ISTA и KAUST разработали новый метод сжатия LLM без использования данных

Недавно был представлен HIGGS (Hadamard Incoherence with Gaussian MSE-optimal GridS) — data-free метод квантизации, который позволяет запускать большие языковые модели локально, за минуты, без GPU.

🔥 Особенности:
🟢Работает без обучающих данных (data-free)
🟢Квантизует даже модели масштаба DeepSeek R1 (671B) и Llama 4 Maverick (400B)
🟢Полностью open-source

📈 Результаты:
🟠Лучшее соотношение качество / размер среди всех data-free методов (NF4, HQQ и др.)
🟠Проверено на Llama 3, Qwen2.5
🟠Статья принята на NAACL 2025

Применение:
▶️Прототипирование без серверов и долгих калибровок
▶️Демократизация доступа к LLM
▶️Подходит для стартапов, исследователей, независимых лабораторий, образовательных и ограниченных сред

🛠 Установка:
pip install flute-kernel

🌟 Пример:
python 
from transformers import AutoModelForCausalLM, AutoTokenizer, HiggsConfig

model = AutoModelForCausalLM.from_pretrained(
"google/gemma-2-9b-it",
quantization_config=HiggsConfig(bits=4),
device_map="auto",
)


🟡Paper
🟡Hugging Face
🟡GitHub

@ai_machinelearning_big_data

#quantization #LLM #opensource #HIGGS #ai
Please open Telegram to view this post
VIEW IN TELEGRAM



group-telegram.com/ai_machinelearning_big_data/7319
Create:
Last Update:

✔️ Учёные Яндекса, НИУ ВШЭ,MIT, ISTA и KAUST разработали новый метод сжатия LLM без использования данных

Недавно был представлен HIGGS (Hadamard Incoherence with Gaussian MSE-optimal GridS) — data-free метод квантизации, который позволяет запускать большие языковые модели локально, за минуты, без GPU.

🔥 Особенности:
🟢Работает без обучающих данных (data-free)
🟢Квантизует даже модели масштаба DeepSeek R1 (671B) и Llama 4 Maverick (400B)
🟢Полностью open-source

📈 Результаты:
🟠Лучшее соотношение качество / размер среди всех data-free методов (NF4, HQQ и др.)
🟠Проверено на Llama 3, Qwen2.5
🟠Статья принята на NAACL 2025

Применение:
▶️Прототипирование без серверов и долгих калибровок
▶️Демократизация доступа к LLM
▶️Подходит для стартапов, исследователей, независимых лабораторий, образовательных и ограниченных сред

🛠 Установка:
pip install flute-kernel

🌟 Пример:

python 
from transformers import AutoModelForCausalLM, AutoTokenizer, HiggsConfig

model = AutoModelForCausalLM.from_pretrained(
"google/gemma-2-9b-it",
quantization_config=HiggsConfig(bits=4),
device_map="auto",
)


🟡Paper
🟡Hugging Face
🟡GitHub

@ai_machinelearning_big_data

#quantization #LLM #opensource #HIGGS #ai

BY Machinelearning







Share with your friend now:
group-telegram.com/ai_machinelearning_big_data/7319

View MORE
Open in Telegram


Telegram | DID YOU KNOW?

Date: |

On February 27th, Durov posted that Channels were becoming a source of unverified information and that the company lacks the ability to check on their veracity. He urged users to be mistrustful of the things shared on Channels, and initially threatened to block the feature in the countries involved for the length of the war, saying that he didn’t want Telegram to be used to aggravate conflict or incite ethnic hatred. He did, however, walk back this plan when it became clear that they had also become a vital communications tool for Ukrainian officials and citizens to help coordinate their resistance and evacuations. Telegram does offer end-to-end encrypted communications through Secret Chats, but this is not the default setting. Standard conversations use the MTProto method, enabling server-client encryption but with them stored on the server for ease-of-access. This makes using Telegram across multiple devices simple, but also means that the regular Telegram chats you’re having with folks are not as secure as you may believe. The last couple days have exemplified that uncertainty. On Thursday, news emerged that talks in Turkey between the Russia and Ukraine yielded no positive result. But on Friday, Reuters reported that Russian President Vladimir Putin said there had been some “positive shifts” in talks between the two sides. Ukrainian forces successfully attacked Russian vehicles in the capital city of Kyiv thanks to a public tip made through the encrypted messaging app Telegram, Ukraine's top law-enforcement agency said on Tuesday. The account, "War on Fakes," was created on February 24, the same day Russian President Vladimir Putin announced a "special military operation" and troops began invading Ukraine. The page is rife with disinformation, according to The Atlantic Council's Digital Forensic Research Lab, which studies digital extremism and published a report examining the channel.
from us


Telegram Machinelearning
FROM American