feat(vision): native mmproj multimodal chat for Qwen35 - #571

Open

davidmroth wants to merge 10 commits into

Luce-Org:mainfrom

davidmroth:feat/vision-native-mmproj

Commits on Aug 1, 2026

feat(vision): native mmproj prefill with per-request DFlash routing.
davidmroth
committed
1f635e0
fix(vision): chunk multimodal text prefill to avoid CUDA OOM
davidmroth
committed
d451dc2
Add layer-split multimodal prefill for Qwen35 vision.
davidmroth
committed
071fbaf
Fix layer-split vision build: include attn opts type, dedupe mask helper.
davidmroth
committed
e30d26d
fix: expose layer-split mmproj via supports_multimodal()
davidmroth
committed
32abfd1
fix(sync): gate Qwen35 supports_multimodal on DFLASH_HAVE_MMPROJ
davidmroth
committed
abbe46a
fix(vision): declare ModelBackend::supports_multimodal
davidmroth
committed
27b9a05
docs(vision): add mmproj guide with end-to-end example screenshot
davidmroth
committed
82d9a7d
docs(vision): store example screenshot outside LFS for PR/docs embeds
davidmroth
committed
317a1d7
docs(vision): add Quick start for testing mmproj builds
davidmroth
committed
1223936