Commit 1d04de8
[TRANSFORMATIONS] Add Qwen 2.5 VL M-RoPE handling to SDPAToPA (#34365)
- Details:
The original Qwen 2.5 VL model has 3D position_ids input facilitating
the 3D M-RoPE. In order to handle this in the PA scenario, where data
for the input is provided as a flat tensor, the 3D coordinate semantics
also had to be preserved. That is why instead of setting the input shape
for position_ids as [-1], for the given case we set it as [3, -1]. This
allows for the correct model output result.
- Tickets:
[CVS-167316](https://jira.devtools.intel.com/browse/CVS-167316)
Signed-off-by: Andrii Staikov <andrii.staikov.intel.com>
---------
Co-authored-by: Denis Orlov <denis.orlov@intel.com>1 parent 682e12d commit 1d04de8
File tree
2 files changed
+429
-4
lines changed- src
- common/transformations/tests/op_conversions
- core/src/pass
2 files changed
+429
-4
lines changed
0 commit comments