Gate News 消息，3月25日，谷歌研究院发布量化压缩算法TurboQuant，可将大语言模型的KV缓存压缩至3 bit，内存占用缩减至少6倍，无需训练或微调，不损失模型精度。在4 bit模式下，于英伟达H100 GPU上计算注意力的速度较32 bit未量化基线提升最高8倍。研究团队在LongBench、Needle In A Haystack、ZeroSCROLLS等长上下文基准上使用Gemma和Mistral模型进行验证，TurboQuant在所有测试中均达到最优表现。该算法由两个子算法组成：PolarQuant通过极坐标变换消除传统量化方法的内存开销，QJL仅用1 bit校正残余误差。该研究由谷歌研究院Amir Zandieh和副总裁兼Google Fellow Vahab Mirrokni主导，与韩国KAIST和纽约大学合作完成，将在ICLR 2026上发表。谷歌表示该技术的主要应用之一是解决Gemini等模型的KV缓存瓶颈。
Disclaimer: The information on this page may come from third parties and does not represent the views or opinions of Gate. The content displayed on this page is for reference only and does not constitute any financial, investment, or legal advice. Gate does not guarantee the accuracy or completeness of the information and shall not be liable for any losses arising from the use of this information. Virtual asset investments carry high risks and are subject to significant price volatility. You may lose all of your invested principal. Please fully understand the relevant risks and make prudent decisions based on your own financial situation and risk tolerance. For details, please refer to
Disclaimer.
Articoli correlati
CEXs Process $19.17T in Spot Crypto Trading in 2025, TradFi Expands with $37B M&A Activity
Gate News message, cryptocurrency exchanges processed $19.17 trillion in spot crypto trading in 2025, while equities reached $155 trillion and foreign exchange markets conducted $9.6 trillion in daily trading. The market has witnessed $37 billion deployed in TradFi M&A by major players, alongside th
GateNews2h fa
Hyperscale Data Reports $5M in Crypto Mining Revenue for Q1 2026
Hyperscale Data (NYSE American: GPUS) disclosed first-quarter 2026 preliminary revenue, with its cryptocurrency mining business generating approximately $5 million, contributing to total company revenue of around $44 million, up 76% year-over-year.
The company plans to divest its diversified
GateNews3h fa
SoFi Reports $1.1B Q1 Revenue, Up 41%, Launches SoFiUSD Stablecoin
According to Businesswire, SoFi Technologies reported record Q1 net revenue of $1.1 billion, up 41% year-over-year, with net income of $167 million, marking its tenth consecutive quarter of GAAP profitability. The fintech company officially launched its full-reserve U.S. dollar stablecoin SoFiUSD
GateNews3h fa
Tokenized RWA Market Reaches $193.2B by End of Q1 2026, Up 256% in 15 Months
According to CoinGecko, the Tokenized Real-World Assets (RWA) market reached $193.2 billion by the end of Q1 2026, up 256% from $54.2 billion at the start of 2025. Tokenized Treasuries led growth, accounting for 67.2% of the market at approximately $130 billion, while Tokenized Commodities rose to $
GateNews5h fa
DeFi Hacks Hit $624.58M in April 2026, Sixth-Largest Loss on Record With Most Incidents
According to DefiLlama, DeFi and on-chain infrastructure hacks caused $624.58 million in losses in April 2026, marking the sixth-largest monthly loss on record. The 23 incidents recorded that month also represent the highest number of attacks in a single month since tracking began in
GateNews8h fa
Galaxy Digital Reports $216M Q1 Loss, Stock Rallies 5% on AI Infrastructure Progress
According to BlockBeats, Galaxy Digital reported a net loss of $216 million in Q1 2026 on April 30, primarily due to a 20% decline in total cryptocurrency market capitalization. The company's crypto asset holdings fell from $1.67 billion at the end of Q4 2025 to $1.36 billion in early
GateNews9h fa
