Repository navigation

Qwen3-Omni: a faster Thinker decode step - #1111

Merged

merceod merged 18 commits into

mainfrom

users/atindra/omni-decode-step

Nov 3, 2026

Merged

Qwen3-Omni: a faster Thinker decode step#1111
merceod merged 18 commits into
mainfrom
users/atindra/omni-decode-step

Commits on Nov 1, 2026
- qwen3-omni: text-only decode on its own walk
  merceod
  committed
  ea10b4f
- fused moe: one-kernel softmax route and B300 tiles
  merceod
  committed
  4b1ad57
- fused moe: route any top-k
  merceod
  committed
  33c2440
- docs: the Thinker's text decode walk
  merceod
  committed
  35200a9
- configs: the pd split places the text decode walk
  merceod
  committed
  356d3f8
- fused moe: tiles a symbolic token count can launch
  merceod
  committed
  169532b
Commits on Nov 2, 2026
- Merge main into omni-decode-step
  merceod
  committed
  101f9fe
- Merge main into omni-decode-step
  merceod
  committed
  a9e8221
- token stop: a set of walks can end the loop
  merceod
  committed
  cea7d03
- qwen3-omni: in-graph embed and rope on thinker_decode too
  merceod
  committed
  9fe581f
- fused moe: one flag for both fused routes
  merceod
  committed
  f469d63
- qwen3-omni: deepstack docstring keeps up
  merceod
  committed
  67513bb
- Merge main into omni-decode-step
  merceod
  committed
  f4cbbeb
Commits on Nov 3, 2026
- Merge main into omni-decode-step
  merceod
  committed
  f7fceb0
- fused moe: no route alignment where the experts run per pair
  merceod
  committed
  1efd566
- fused moe test: reset the tuned tables the way the fixture does
  merceod
  committed
  d117982
- Merge main into omni-decode-step
  merceod
  committed
  ad1b4e8
- Merge branch 'main' into users/atindra/omni-decode-step
  merceod
  authored
  24f3684