Repository navigation

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

Merged

inureyes merged 21 commits into

mainfrom

feature/issue-866-qwen25-vl-xla

Oct 15, 2026

Merged

feat: add Qwen2.5-VL OpenXLA vision path#920
inureyes merged 21 commits into
mainfrom
feature/issue-866-qwen25-vl-xla

Commits on Oct 15, 2026
- feat: add Qwen2.5-VL OpenXLA vision path
  inureyes
  committed
  20f52ad
- fix: honor unquantized Qwen2.5 vision towers
  inureyes
  committed
  03f428a
- test(xla): add Qwen2.5-VL strict oracle
  inureyes
  committed
  a60b042
- fix(xla): align Qwen vision GPU precision
  inureyes
  committed
  053f072
- test(xla): probe Qwen layer 29 substages
  inureyes
  committed
  f3c46d8
- test(xla): materialize Qwen eager probe
  inureyes
  committed
  f8b4e9a
- fix(xla): snapshot Qwen diagnostic substages
  inureyes
  committed
  74d4f99
- fix(xla): own Qwen diagnostic snapshots
  inureyes
  committed
  c206b8a
- test(xla): expose Qwen F16 norm overflow
  inureyes
  committed
  d00ecfd
- fix(xla): stabilize Qwen vision RMSNorm
  inureyes
  committed
  541702c
- test(xla): localize Qwen full attention drift
  inureyes
  committed
  dd3ae20
- fix(xla): preserve Qwen diagnostic output ranks
  inureyes
  committed
  9436475
- test(xla): trace Qwen pre-probe layer outputs
  inureyes
  committed
  d7f64f9
- test(xla): probe Qwen layer 17 substages
  inureyes
  committed
  78294fa
- test(xla): trace Qwen layer 17 SwiGLU stages
  inureyes
  committed
  6d61402
- test(xla): distinguish Qwen MLP precision paths
  inureyes
  committed
  c666f44
- test(xla): replay Qwen MLP from IREE norm
  inureyes
  committed
  c246989
- fix(xla): preserve Qwen vision F16 state boundaries
  inureyes
  committed
  7497d02
- fix(xla): revert unsafe Qwen F16 state rounding
  inureyes
  committed
  453d213
- fix(xla): mark Qwen diagnostics unqualified
  inureyes
  committed
  a4bdeef
- fix(xla): qualify Qwen2.5-VL vision and oracle on ROCm
  inureyes
  committed
  6a68bb9

Commits

Commits on Oct 15, 2026