feat(xla): support Molmo sparse additive visual embeddings - #918

Draft

inureyes wants to merge 16 commits into

mainfrom

feature/issue-870-molmo-xla

Commits on Jul 27, 2026

feat(xla): define Molmo sparse additive merge contract
inureyes
committed
ecc57fc
fix(xla): make Molmo sparse merge atomic
inureyes
committed
e46bc5d
feat(xla): map Molmo decoder into the dense emitter
inureyes
committed
2a50f81
feat(xla): wire Molmo sparse prefill producer
inureyes
committed
30c8493
feat(xla): emit Molmo native vision graph
inureyes
committed
e4cc57d
feat(xla): execute Molmo vision through resident IREE
inureyes
committed
20dfcfd
fix(xla): slice Molmo fused attention bias
inureyes
committed
cefddd6
fix(xla): preserve Molmo decoder and merge semantics
inureyes
committed
234b6c2
test(xla): add Molmo prepared-prefill parity gate
inureyes
committed
2939481
test(xla): add strict Molmo boundary diagnostics
inureyes
committed
691d28d
test(xla): stage Molmo eager vision diagnostics
inureyes
committed
3344113
test(xla): add CUDA Molmo boundary runner
inureyes
committed
2c5a5a1
style(xla): order Molmo exports
inureyes
committed
6803b17
fix(xla): mark Molmo artifact unqualified
inureyes
committed
f1c0a18
fix(xla): restore feature-neutral Molmo rounding
inureyes
committed
a4047b2
fix(xla): configure local-task threads before Molmo IREE creation
inureyes
committed
a954d21