feat: add Qwen2.5-VL OpenXLA vision path - #920

Draft

inureyes wants to merge 20 commits into

mainfrom

feature/issue-866-qwen25-vl-xla

Commits on Jul 27, 2026

feat: add Qwen2.5-VL OpenXLA vision path
inureyes
committed
4b044cc
fix: honor unquantized Qwen2.5 vision towers
inureyes
committed
0a3be95
test(xla): add Qwen2.5-VL strict oracle
inureyes
committed
e89057a
fix(xla): align Qwen vision GPU precision
inureyes
committed
572ffc2
test(xla): probe Qwen layer 29 substages
inureyes
committed
ead807e
test(xla): materialize Qwen eager probe
inureyes
committed
5b3d16a
fix(xla): snapshot Qwen diagnostic substages
inureyes
committed
11b3313
fix(xla): own Qwen diagnostic snapshots
inureyes
committed
fb47646
test(xla): expose Qwen F16 norm overflow
inureyes
committed
2e7a19a
fix(xla): stabilize Qwen vision RMSNorm
inureyes
committed
fd32071
test(xla): localize Qwen full attention drift
inureyes
committed
015e59c
fix(xla): preserve Qwen diagnostic output ranks
inureyes
committed
0024800
test(xla): trace Qwen pre-probe layer outputs
inureyes
committed
28234ca
test(xla): probe Qwen layer 17 substages
inureyes
committed
9cf6622
test(xla): trace Qwen layer 17 SwiGLU stages
inureyes
committed
d7c67d9
test(xla): distinguish Qwen MLP precision paths
inureyes
committed
41fee35
test(xla): replay Qwen MLP from IREE norm
inureyes
committed
961b436
fix(xla): preserve Qwen vision F16 state boundaries
inureyes
committed
786f77c
fix(xla): revert unsafe Qwen F16 state rounding
inureyes
committed
3756d19
fix(xla): mark Qwen diagnostics unqualified
inureyes
committed
02d41a1