Repository navigation

qwen3_omni: prefix caching for the Thinker - #836

Merged

Gaurav-Shah05 merged 32 commits into

mainfrom

prefix-cache-omni

Oct 27, 2026

Merged

qwen3_omni: prefix caching for the Thinker#836
Gaurav-Shah05 merged 32 commits into
mainfrom
prefix-cache-omni

Commits on Oct 22, 2026
- engine: probe a walk whose submodule cuts its own tensor_inputs
  Gaurav-Shah05
  committed
  4419d97
- conductor: cache off when a keyed walk streams to a running partition
  Gaurav-Shah05
  committed
  199d9ea
- qwen3_omni: thinker cuts its inputs and places a cut walk's positions
  Gaurav-Shah05
  committed
  c8d15e9
- engine: an item encoder may opt out of its dynamic=True compile
  Gaurav-Shah05
  committed
  b8985ab
- qwen3_omni: run the audio encoder eager as an item encoder
  Gaurav-Shah05
  committed
  0970ba5
- qwen3_omni: the vision encoder compiles one graph for every image size
  Gaurav-Shah05
  committed
  7757d3e
- qwen3_omni: the thinker keys its prompt, images, audio and replies
  Gaurav-Shah05
  committed
  0c6cabd
- test: prefix caching for the thinker
  Gaurav-Shah05
  committed
  fd22735
- comment pass
  Gaurav-Shah05
  committed
  815b345
- test pass
  Gaurav-Shah05
  committed
  644b3da
Commits on Oct 24, 2026
Commits on Oct 27, 2026
- qwen3_omni: mark the vision encoder's cu_seqlens dynamic
  
  dghosef
  and
  claude
  committed
  1937be9
- qwen3_omni: size the vision encoder's max_seqlen to one frame
  
  dghosef
  and
  claude
  committed
  863e0da
- docs: Qwen3-Omni's Thinker counts prefix-cache hits too
  
  dghosef
  and
  claude
  committed
  00b6fe0
- prefix-caching skill: tensor inputs a split_inputs override cuts are eligible
  
  dghosef
  and
  claude
  committed
  02381ed
- engine: a submodule that cuts its tensor_inputs says so with a flag
  Gaurav-Shah05
  committed
  fada3f7
- qwen3_omni: a request with audio output turns the prefix cache off
  Gaurav-Shah05
  committed
  fdd6c43
- model: Qwen3.5 and Qwen3-Omni share their 3D position ids
  Gaurav-Shah05
  committed
  9f6c6b7
- test: the tensor_inputs flag, the audio-output opt-out and shared 3D positions
  Gaurav-Shah05
  committed
  de5fd16
- Merge remote-tracking branch 'origin/main' into prefix-cache-omni
  Gaurav-Shah05
  committed
  f38c386
- conductor: cache off for a request that runs a partition a keyed node streams into
  Gaurav-Shah05
  committed
  4722f7c
- test: a request whose talker runs is not cached
  Gaurav-Shah05
  committed
  65a1f31

Navigation Menu

Repository navigation

qwen3_omni: prefix caching for the Thinker - #836

qwen3_omni: prefix caching for the Thinker#836
Gaurav-Shah05 merged 32 commits into
mainfrom
prefix-cache-omni

Commits on Oct 22, 2026

engine: probe a walk whose submodule cuts its own tensor_inputs

conductor: cache off when a keyed walk streams to a running partition

qwen3_omni: thinker cuts its inputs and places a cut walk's positions

engine: an item encoder may opt out of its dynamic=True compile

qwen3_omni: run the audio encoder eager as an item encoder

qwen3_omni: the vision encoder compiles one graph for every image size

qwen3_omni: the thinker keys its prompt, images, audio and replies

test: prefix caching for the thinker

comment pass

test pass

Commits on Oct 24, 2026

qwen3_omni: build the audio encoder's chunking before the compiled call

engine: drop the item encoder's eager opt-out

test: Qwen3-Omni's encoders reuse their graphs for sizes in range

Merge remote-tracking branch 'origin/main' into prefix-cache-omni

worker: find streamed partitions from keyed labels

qwen3_omni: drop the keyed stream that kv_labels replaces

test: the prefix tests declare kv labels and layouts

qwen3_omni: a prompt with a video keys its leading text

qwen3_omni: the layout takes an image's span from image_span

qwen3_omni: a video prompt keys one span, with the reason

test: a Qwen3-Omni prompt with a video keys its leading text

Commits on Oct 27, 2026

qwen3_omni: mark the vision encoder's cu_seqlens dynamic

qwen3_omni: size the vision encoder's max_seqlen to one frame

docs: Qwen3-Omni's Thinker counts prefix-cache hits too

prefix-caching skill: tensor inputs a split_inputs override cuts are eligible

engine: a submodule that cuts its tensor_inputs says so with a flag

qwen3_omni: a request with audio output turns the prefix cache off

model: Qwen3.5 and Qwen3-Omni share their 3D position ids

test: the tensor_inputs flag, the audio-output opt-out and shared 3D positions

Merge remote-tracking branch 'origin/main' into prefix-cache-omni

conductor: cache off for a request that runs a partition a keyed node streams into

test: a request whose talker runs is not cached