Repository navigation

B70 attention opt - #2

Merged

ShengchenJ merged 8 commits into

mainfrom

b70_attention_opt

Oct 19, 2026

Merged

B70 attention opt#2
ShengchenJ merged 8 commits into
mainfrom
b70_attention_opt

Commits on Oct 13, 2026
- build: detect the local GPU target, 256-GRF mode on Battlemage
  
  ShengchenJ
  and
  claude
  committed
  921063c
- base_types: bf16_2 conversion reuses the scalar converter
  
  ShengchenJ
  and
  claude
  committed
  f94e4b8
- reductions: transposing butterfly for register-tile row reductions
  
  ShengchenJ
  and
  claude
  committed
  2550b22
- reductions: register-region transpose-reduce for float max/sum
  
  ShengchenJ
  and
  claude
  committed
  4fdd0b8
- gqa_forward: use wg256/seq_q8 on Battlemage
  
  ShengchenJ
  and
  claude
  committed
  b2cf57e
- B70 attention: hoisted 2D block-load payloads for K/V
  
  ShengchenJ
  and
  claude
  committed
  5e921a4
- B70 attention: cooperative K/V prefetch across subgroups
  
  ShengchenJ
  and
  claude
  committed
  27b206a
Commits on Oct 19, 2026
- B70 attention: derive K/V block-load layout from gl strides
  
  ShengchenJ
  and
  claude
  committed
  c5c93ba