microsoft / microsoft/winml-cli

SA: Coverage report investigation — QNN NPU (2026-04-01)

Open
#308 0 comments 0 reactions 0 assignees View on GitHub

Nobody has claimed this yet.

P2 static-analyzer triaged
Dominant language
Python
Stars
40
Forks
11
Avg merge
1d 8h
Merged PRs (30d)
50

Description

Overview

Investigate regressions and unknown ops surfaced in the QNN NPU SA coverage report (2026-04-01 run). Report: https://icy-moss-029643d00.6.azurestaticapps.net/sa_e2e_coverage/QNNExecutionProvider_NPU/0401/sa_eval_report.html

Summary stats: Avg SUPPORTED (Pre): 95.0% → (Post): 94.9% | Avg delta: -0.1% | Avg UNKNOWN: 27 | All-SUPPORTED (Post): 98.3%

Regressions (Post < Pre)

Model Task Pre Post Delta
microsoft/swin-large-patch4-window7-224 image-classification 68.8% 66.7% -2.1%
PekingU/rtdetr_v2_r18vd object-detection 76.7% 74.2% -2.5%
facebook/bart-large-mnli text-classification 77.3% 76.2% -1.1%

Partial op patterns involved: Reshape, Transpose, Pad, Slice, Add, Mul, Sub, TopK

Unknown Ops

Model Unknown Count Op Types
facebook/bart-large-mnli 1 GatherND
PekingU/rtdetr_r50vd_coco_o365 2 GridSample, Tile
PekingU/rtdetr_r101vd_coco_o365 2 GridSample, Tile
PekingU/rtdetr_v2_r18vd 2 GridSample, Tile

Acceptance Criteria

  • Root cause identified for swin-large regression (-2.1%)
  • Root cause identified for rtdetr_v2_r18vd regression (-2.5%)
  • Root cause identified for bart-large-mnli regression (-1.1%)
  • GatherND, GridSample, Tile added to QNN coverage DB or flagged as unsupported
  • Next report shows 0 regressions and 0 unknown ops for the above models

Contributor guide

Open the contributing guide

First steps

  1. Read the whole issue, then the project's contributing guide.
  2. Comment on the issue to say you are picking it up — it saves two people doing the same work.
  3. Fork the repository and make your change on a branch.
  4. Open a pull request that references the issue number.

Research direction

Start with the linked QNN NPU SA coverage report for the 2026-04-01 run and compare the listed regressions with their partial op patterns. Investigate the three named models and the unknown GatherND, GridSample, and Tile operations. Done means root causes are identified, the operations are covered or explicitly unsupported, and a subsequent report has no listed regressions or unknown ops.

Written by the indexing model from the issue text.

Assessment

Tech stack
python
Domain
machine-learning, testing-qa
Issue type
Bug
Difficulty
5/5
Estimated time
Over a week
Activity status
Quiet
Clarity
Mostly clear
Newbie friendliness
35/100

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.