Skip to content

26.1.1 hotfix#35170

Merged
peterchen-intel merged 2 commits intoopenvinotoolkit:releases/2026/1from
peterchen-intel:26.1.1_hotfix
Apr 7, 2026
Merged

26.1.1 hotfix#35170
peterchen-intel merged 2 commits intoopenvinotoolkit:releases/2026/1from
peterchen-intel:26.1.1_hotfix

Conversation

@peterchen-intel
Copy link
Copy Markdown
Contributor

Details:

  • Cherry pick a bug fixing from master branch
  • commit ID: 2f0404d

Tickets:

AI Assistance:

  • AI assistance used: yes
  • If yes, summarize how AI was used and what human validation was performed (build/tests/manual checks). Use the github agent to cherry pick commit from master branch

andrew-k-park and others added 2 commits April 5, 2026 00:04
… instead of runtime shape (openvinotoolkit#34878)

- **Symptom**: SwinTransformer accuracy regression on GPU (max abs diff:
7.53) after openvinotoolkit#34680.
- **Root-cause**: The condition `reshape_ps[0].get_length() == 1` used
runtime shape to reject spatial-flatten reshapes in the in-place crop
path, but it also incorrectly rejects batch-squeeze reshapes when the
runtime batch dim happens to be 1 (SwinTransformer layers.3,
window_count=1)
- **Resolution**: Use compile-time `output_pattern[0]` instead of
runtime shape. `output_pattern[0] == -1` means batch-squeeze (allow),
`0` or `1` means batch-preserved/spatial-flatten (reject)

directly)
- intel_gpu/src/graph/include/`reshape_inst.h` —
`is_runtime_propagatable_padding()`, batch axis=0 block
- intel_gpu/src/graph/graph_optimizer/`prepare_buffer_fusing.cpp` —
`update_in_place_crop_padding_simple_data_format()`, build-time dyn_pad
branch and runtime padding branch

customer model)
- $ pytest test_ovc_mo.py --modules pipelines/production/pytorch/heavy
-k Pytorch_SwinTransformer --dynamism_type=None --log-cli-level INFO

- SwinTransformer layers.3: input [3,1,24,49,32] → split axis=0 into 3
crops [1,1,24,49,32] → reshape [-1,24,49,32] (batch-squeeze,
output_pattern[0]==-1)
window_count=1 causes runtime reshape output shape[0]=1, which falsely
matches the spatial-flatten rejection condition
<img width="1937" height="626" alt="image"
src="https://github.com/user-attachments/assets/a91b977a-64f6-40d9-9cb5-69c855456ed7"
/>

 - [x] Is it a proper fix? (not a workaround)
 - [x] Did you include test case for this fix, if necessary?
- [x] Did you review existing test that can be extended to cover this
scenario? Which test did you review?
- `in_place_crop_dynamic_batch_axis_split_with_reshape` (batch_squeeze,
spatial_flatten variants)

 - [CVS-183289](https://jira.devtools.intel.com/browse/CVS-183289)

Signed-off-by: Andrew Park <andrew.park@intel.com>

Co-authored-by: peterchen-intel <19401820+peterchen-intel@users.noreply.github.com>
…-2f0404d

Cherry-pick: [GPU] Fix in-place crop batch-squeeze detection to use output_pattern instead of runtime shape (openvinotoolkit#34878)
Copy link
Copy Markdown

@geunhwan geunhwan left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Backport of #34878 which was already merged in the master.

@peterchen-intel peterchen-intel added this pull request to the merge queue Apr 7, 2026
Merged via the queue into openvinotoolkit:releases/2026/1 with commit 3687ccc Apr 7, 2026
188 checks passed
@peterchen-intel peterchen-intel deleted the 26.1.1_hotfix branch April 7, 2026 10:11
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

category: GPU OpenVINO GPU plugin

Projects

None yet

Development

Successfully merging this pull request may close these issues.

4 participants