← Back to feed

Marble: Exact sequence parallelism for packed multimodal masks

AI · 2026-10-20 07:08

Luma Labs提出Marble方案,针对打包多模态序列的注意力掩码差异,组合环形注意力、在线FlashAttention-4内核与稀疏all-to-all。

Source: lumalabs.ai


Auto-generated summary for reference only. RSS