feat(xla): support Molmo2 indexed attention pooling - #916

Merged

inureyes merged 20 commits into

mainfrom

feature/issue-871-molmo2-xla

Aug 21, 2026

Commits on Aug 21, 2026

feat(xla): add Molmo2 vision runtime
inureyes
committed
5378d9e
fix(xla): satisfy Molmo2 runtime lints
inureyes
committed
482d56a
fix(xla): harden Molmo2 pooling preflight
inureyes
committed
b9c4d09
fix(xla): resolve Molmo2 layers against declared depth
inureyes
committed
0e4c2a8
test(xla): add Molmo2 vision parity gate
inureyes
committed
0457293
test(xla): expose Molmo2 CPU parity diagnostics
inureyes
committed
4a8724d
fix(xla): trace Molmo2 vision first divergence
inureyes
committed
813a41f
fix(molmo2): prefer declared vision sections
inureyes
committed
4f73c50
fix(xla): preserve Molmo2 projector SiLU order
inureyes
committed
64a19c3
test(xla): report Molmo2 gate progress
inureyes
committed
9053754
test(xla): localize Molmo2 projector divergence
inureyes
committed
162c22f
test(xla): probe Molmo2 layer 24 row
inureyes
committed
af3ad9f
test(xla): probe Molmo2 selected layer 18
inureyes
committed
aae31e7
fix(xla): mark Molmo2 artifact unqualified
inureyes
committed
0e846fc
refactor(xla): share Molmo2 numeric decompositions
inureyes
committed
ce1891f
fix(xla): configure local-task threads before Molmo2 IREE creation
inureyes
committed
2df457f
fix(xla): gate the diagnostics-only weight-loader re-export
inureyes
committed
ed881bd
fix(xla): reject a context capacity that can never admit one image
inureyes
committed
272fbc2
test(xla): gate Molmo2 vision parity on a scale-aware tolerance
inureyes
committed
22a753c
test(xla): check the tolerance shape in every build of the gate
inureyes
committed
ba25807

Navigation Menu

Uh oh!

feat(xla): support Molmo2 indexed attention pooling - #916

feat(xla): support Molmo2 indexed attention pooling#916
inureyes merged 20 commits into
mainfrom
feature/issue-871-molmo2-xla

Commits on Aug 21, 2026

feat(xla): add Molmo2 vision runtime

fix(xla): satisfy Molmo2 runtime lints

fix(xla): harden Molmo2 pooling preflight

fix(xla): resolve Molmo2 layers against declared depth

test(xla): add Molmo2 vision parity gate

test(xla): expose Molmo2 CPU parity diagnostics

fix(xla): trace Molmo2 vision first divergence

fix(molmo2): prefer declared vision sections

fix(xla): preserve Molmo2 projector SiLU order

test(xla): report Molmo2 gate progress

test(xla): localize Molmo2 projector divergence

test(xla): probe Molmo2 layer 24 row

test(xla): probe Molmo2 selected layer 18

fix(xla): mark Molmo2 artifact unqualified

refactor(xla): share Molmo2 numeric decompositions

fix(xla): configure local-task threads before Molmo2 IREE creation

fix(xla): gate the diagnostics-only weight-loader re-export

fix(xla): reject a context capacity that can never admit one image

test(xla): gate Molmo2 vision parity on a scale-aware tolerance

test(xla): check the tolerance shape in every build of the gate