Minimax MSA

5 atoms · 跨 2 天 · 首见 2026-05-26 · 最近 2026-05-27

三色: 🟦 fact 2 · 🟥 take 3 · stance ▲0/▼0/◆0

来源: X 5

时态: fresh:5

标签: 好思考:3 · 好观点:2

叙事 (narrative) (1)

• 2026-05-26 · indexer speed improvement
  _my first read i thought they make the indexer block_size times faster, and doubled the number of heads in exchange_
  tags: 好思考 · → daily
  value: qty=block_size times faster · direction=first read thought

事实 (fact) (4)

• 2026-05-27 · bottleneck in Q_idx and K_idx projections
  _I found that bottleneck becomes these Q_idx and K_idx projections for which I probably need fused triton kernels._
  tags: 好思考 · → daily
• 2026-05-26 · indexer attention type
  wdym? for MSA you do block level indexer attention
  tags: 好思考 · → daily
• 2026-05-26 · indexer calculates full sequence before down sampling
  hmm the indexer still calculate the full sequence before down sampling
  tags: 好观点 · → daily
• 2026-05-26 · indexer operation
  _i mean you still do q_idx dot k_idx on the full sequence length_
  tags: 好观点 · → daily
  value: direction=q_idx dot k_idx on full sequence length

时间轴 (近 20)

• 2026-05-27 · fact · bottleneck in Q_idx and K_idx projections · → daily
• 2026-05-26 · fact · indexer calculates full sequence before down sampling · → daily
• 2026-05-26 · fact · indexer operation · → daily
• 2026-05-26 · narrative · indexer speed improvement · → daily
• 2026-05-26 · fact · indexer attention type · → daily

← 实体目录 · 系统日志