Figure 5c reports forward attention throughput on A100 80GB SXM4; PyTorch is marked OOM at sequence length 16k. The panel covers causal attention with head dimension 64.

Figure 5c reports forward attention throughput on A100 80GB SXM4; PyTorch is marked OOM at sequence length 16k. The panel covers causal attention with head dimension 64. · CiteArk

测量结果

已评估 0/29 项

获得支持 0有争议或冲突 0尚未评估 29

保留同一结论下的完整结果组。已有评估排在前面；单项判断不替代整组结论。展开行查看条件、来源与历次证据。

每行是一项测量，分别显示论文报告、最近可用实测和累计证据状态。
指标与条件	论文报告	最近可用实测	证据判断
method: PyTorch · sequenceLength: 512	10 TFLOPs/s	—	尚未评估
method: FlashAttention · sequenceLength: 512	56 TFLOPs/s	—	尚未评估
method: xFormers · sequenceLength: 512	71 TFLOPs/s	—	尚未评估
method: FlashAttention Triton · sequenceLength: 512	78 TFLOPs/s	—	尚未评估
method: FlashAttention-2 · sequenceLength: 512	115 TFLOPs/s	—	尚未评估
method: PyTorch · sequenceLength: 1k	10 TFLOPs/s	—	尚未评估
method: FlashAttention · sequenceLength: 1k	70 TFLOPs/s	—	尚未评估
method: xFormers · sequenceLength: 1k	82 TFLOPs/s	—	尚未评估
method: FlashAttention Triton · sequenceLength: 1k	99 TFLOPs/s	—	尚未评估
method: FlashAttention-2 · sequenceLength: 1k	146 TFLOPs/s	—	尚未评估
method: PyTorch · sequenceLength: 2k	10 TFLOPs/s	—	尚未评估
method: FlashAttention · sequenceLength: 2k	81 TFLOPs/s	—	尚未评估

Figure 5c reports forward attention throughput on A100 80GB SXM4; PyTorch is marked OOM at sequence length 16k. The panel covers causal attention with head dimension 64.

报告指标与观测值

Assessment（0）

关联运行（0）

测量结果