Repository navigation

jev_vl: add native inference and prefix caching - #96

Merged

hsliuustc0106 merged 11 commits into

InferMatrixLab:mainfrom

linear3735:codex/jev-vl-native-mm-cache

Oct 9, 2026

Merged

jev_vl: add native inference and prefix caching#96
hsliuustc0106 merged 11 commits into
InferMatrixLab:mainfrom
linear3735:codex/jev-vl-native-mm-cache

Commits on Oct 6, 2026
- cuda: add prefix-continuation kernels for hybrid qwen3_5 models
  
  linear3735
  and
  claude
  committed
  8a37fe0
- qwen3_5/native: multimodal inputs and prefix-state capture/continue
  
  linear3735
  and
  claude
  committed
  c4f3474
- jev_vl: add native multimodal worker with multi-level decision caches
  
  linear3735
  and
  claude
  committed
  127b20b
- jev_vl: integrate current backend and validate prefix caching
  linear3735
  committed
  522f225
- jev_vl: keep replay fixtures and archive generated run output
  hsliuustc0106
  committed
  4927d2a
- jev_vl: move validation artifacts out of the source tree
  linear3735
  committed
  e9d1321
- jev_vl: reject non-string decision controls
  linear3735
  committed
  90bdeb3
- jev_vl: reject extra image placeholders on cache hits
  linear3735
  committed
  e8a3ab5
- Fix JEV image preparation and prefix bounds
  linear3735
  committed
  f8f55a4
- Merge upstream Open-Jev prefill batching
  linear3735
  committed
  37525bc
Commits on Oct 8, 2026
- Merge Open-Jev-9B support and restore JEV replay fixtures
  linear3735
  committed
  18dfd11