BTC $64,366.13 +0.03%
ETH $1,812.62 +0.80%
BNB $590.14 +0.17%
XRP $0.9559 +1.22%
SOL $72.44 +2.33%
TRX $0.3057 +1.15%
DOGE $0.0585 +0.49%
ADA $0.1527 +0.73%
BCH $208.97 +1.76%
LINK $9.14 +0.58%
HYPE $68.73 +2.98%
AAVE $111.85 +2.85%
SUI $0.7208 +0.40%
XLM $0.1373 -0.00%
ZEC $916.11 +3.13%
AAPL $352.45 -1.54%
AMZN $256.60 -3.14%
GOOGL $282.02 -4.29%
MSFT $516.52 -2.02%
META $621.51 -2.31%
NVDA $155.55 -2.81%
TSLA $376.41 -2.26%
SNDK $792.94 -5.28%
INTC $53.74 -5.49%
SPCX $136.47 -2.50%
MU $713.80 -3.74%
AMD $502.51 -1.56%
BTC $64,366.13 +0.03%
ETH $1,812.62 +0.80%
BNB $590.14 +0.17%
XRP $0.9559 +1.22%
SOL $72.44 +2.33%
TRX $0.3057 +1.15%
DOGE $0.0585 +0.49%
ADA $0.1527 +0.73%
BCH $208.97 +1.76%
LINK $9.14 +0.58%
HYPE $68.73 +2.98%
AAVE $111.85 +2.85%
SUI $0.7208 +0.40%
XLM $0.1373 -0.00%
ZEC $916.11 +3.13%
AAPL $352.45 -1.54%
AMZN $256.60 -3.14%
GOOGL $282.02 -4.29%
MSFT $516.52 -2.02%
META $621.51 -2.31%
NVDA $155.55 -2.81%
TSLA $376.41 -2.26%
SNDK $792.94 -5.28%
INTC $53.74 -5.49%
SPCX $136.47 -2.50%
MU $713.80 -3.74%
AMD $502.51 -1.56%

DeepSeek 推出 NSA,用于超快速的长上下文训练和推理

2025-02-18 16:37:45

ChainCatcher 消息,据金十报道,DeepSeek 推出 NSA。

DeepSeek 称,NSA 是一种与硬件一致且本机可训练的稀疏注意力机制,用于超快速的长上下文训练和推理。通过针对现代硬件的优化设计,NSA 加快了推理速度,同时降低了预训练成本,而不会影响性能。

在一般基准测试、长上下文任务和基于指令的推理上,它的表现与完全注意力模型相当甚至更好。

链捕手ChainCatcher提醒,请广大读者理性看待区块链,切实提高风险意识,警惕各类虚拟代币发行与炒作, 站内所有内容仅系市场信息或相关方观点,不构成任何形式投资建议。如发现站内内容含敏感信息,可点击,我们会及时处理。

关联标签
关联标签
app_icon
ChainCatcher 与创新者共建Web3世界