Repository navigation

exl3 cuda: the 27B's EXL3 speed series on the native Dense module (#260), stacked on #1465 - #1467

Merged

ashhart merged 11 commits into

ashhart:mainfrom

BHCC2025:pr-qwen-exl3-speed

Nov 2, 2026

Merged

exl3 cuda: the 27B's EXL3 speed series on the native Dense module (#260), stacked on #1465#1467
ashhart merged 11 commits into
ashhart:mainfrom
BHCC2025:pr-qwen-exl3-speed

Commits on Nov 2, 2026
- exl3 cuda: per-weight plans, mid-M kernels and linear_wc for 17-128 rows, projection groups, PDL (same bits)
  
  BHCC2025
  authored and
  ashhart
  committed
  f40662e
- qwen3_5 cuda: EXL3 decode groups and fixed plans for the 27B's shapes (same bits)
  
  BHCC2025
  authored and
  ashhart
  committed
  d17c6a0
- exl3 cuda: builds for sm_80-89, the group kernels named in source, contract records for the new launches
  
  BHCC2025
  authored and
  ashhart
  committed
  d1ba34c
- exl3 cuda tests: the 27B's shapes against float64 under plan()'s launch and the fixed one (exl3-plan-f64)
  
  BHCC2025
  authored and
  ashhart
  committed
  c4a0085
- exl3 cuda: review fixes - plan()'s launches, group reserves after the MLP, group counters through grownCounters, the replacements in the kernel contract
  
  BHCC2025
  authored and
  ashhart
  committed
  6ce2e26
- exl3 cuda: every mid-M, linear_wc and group form the dispatch reaches in the kernel contract
  
  BHCC2025
  authored and
  ashhart
  committed
  81561fa
- exl3 cuda: the mid-M and linear_wc tables in exl3_midm.zig, so exl3.zig stays under 600 lines on main
  
  BHCC2025
  authored and
  ashhart
  committed
  ca113f4
- kernel contract: the EXL3 set joins the family sets, so contract_tests.zig stays under 600 lines
  
  BHCC2025
  authored and
  ashhart
  committed
  5fb3a36
- docs: the EXL3 recipe lists TF_EXL3_PDL=0, plain launches for the rotations and the linear
  
  BHCC2025
  authored and
  ashhart
  committed
  8bda27e
- kernel contract: the Flash Next and EXL3 sets side by side after the rebase
  ashhart
  committed
  0e57f4f
- changelog: #1467 under Unreleased
  ashhart
  committed
  fb16783

Commits

Commits on Nov 2, 2026