BTC $63,722.38 -1.94%
ETH $1,791.19 -1.63%
BNB $587.03 -0.83%
XRP $0.9624 +0.44%
SOL $72.73 +1.89%
TRX $0.3071 +0.43%
DOGE $0.0588 -1.09%
ADA $0.1521 -1.07%
BCH $207.59 -0.50%
LINK $9.09 -1.07%
HYPE $67.84 -1.11%
AAVE $109.98 +2.22%
SUI $0.7174 -1.04%
XLM $0.1379 +0.00%
ZEC $914.90 +1.87%
AAPL $351.04 -0.32%
AMZN $257.58 -0.26%
GOOGL $280.83 -1.40%
MSFT $506.37 -2.43%
META $620.41 -0.98%
NVDA $153.35 -1.51%
TSLA $379.59 +0.49%
SNDK $787.92 -2.95%
INTC $53.69 -3.03%
SPCX $141.15 +1.80%
MU $697.76 -3.18%
AMD $489.66 -2.73%
BTC $63,722.38 -1.94%
ETH $1,791.19 -1.63%
BNB $587.03 -0.83%
XRP $0.9624 +0.44%
SOL $72.73 +1.89%
TRX $0.3071 +0.43%
DOGE $0.0588 -1.09%
ADA $0.1521 -1.07%
BCH $207.59 -0.50%
LINK $9.09 -1.07%
HYPE $67.84 -1.11%
AAVE $109.98 +2.22%
SUI $0.7174 -1.04%
XLM $0.1379 +0.00%
ZEC $914.90 +1.87%
AAPL $351.04 -0.32%
AMZN $257.58 -0.26%
GOOGL $280.83 -1.40%
MSFT $506.37 -2.43%
META $620.41 -0.98%
NVDA $153.35 -1.51%
TSLA $379.59 +0.49%
SNDK $787.92 -2.95%
INTC $53.69 -3.03%
SPCX $141.15 +1.80%
MU $697.76 -3.18%
AMD $489.66 -2.73%

DeepSeek 推出 NSA,用於超快速的長上下文訓練和推理

2025-02-18 16:37:45

ChainCatcher 消息,据金十報導,DeepSeek 推出 NSA。

DeepSeek 稱,NSA 是一種與硬體一致且本機可訓練的稀疏注意力機制,用於超快速的長上下文訓練和推理。通過針對現代硬體的優化設計,NSA 加快了推理速度,同時降低了預訓練成本,而不會影響性能。

在一般基準測試、長上下文任務和基於指令的推理上,它的表現與完全注意力模型相當甚至更好。

鏈捕手ChainCatcher提醒,請廣大讀者理性看待區塊鏈,切實提高風險意識,警惕各類虛擬代幣發行與炒作,站內所有內容僅係市場信息或相關方觀點,不構成任何形式投資建議。如發現站內內容含敏感信息,可點擊,我們會及時處理。

關聯標籤
關聯標籤
app_icon
ChainCatcher 與創新者共建Web3世界