Address DML crashes in WebNN QDQ subgraph tests #26822
Merged
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
fdwr
reviewed
Dec 18, 2025
onnxruntime/core/optimizer/qdq_transformer/selectors_actions/qdq_selector_action_transformer.cc
Outdated
Show resolved
Hide resolved
edgchen1
reviewed
Dec 19, 2025
onnxruntime/core/optimizer/qdq_transformer/selectors_actions/qdq_selector_action_transformer.cc
Outdated
Show resolved
Hide resolved
edgchen1
reviewed
Dec 19, 2025
onnxruntime/core/optimizer/qdq_transformer/selectors_actions/qdq_selector_action_transformer.cc
Outdated
Show resolved
Hide resolved
edgchen1
approved these changes
Dec 19, 2025
adrastogi
added a commit
that referenced
this pull request
Jan 9, 2026
### Description <!-- Describe your changes. --> This change proposes a fix to some DML crashes observed in WebNN QDQ subgraph tests. The root cause of the problem as analyzed in the issue linked below is that the DML EP does not have kernels implemented for the QLinear versions of LeakyRelu and Softmax. The QDQ transformer has logic to fuse a 3-node sequence in the graph to the associated QLinear custom op. There is an operator registry which contains the mapping between EPs and the ops that the transformer consults for computing the quantized variants. The registry currently has the CPU and DML EPs supporting the same sets of operators (thereby assuming that the fused node would also be supported by the EP), so what I did was split out a separate registration for the LeakyRelu and Softmax ops and associate them with the CPU EP, while removing those ops from the prior list. I was trying to also explore other options like seeing if I could somehow check to see if the EP has a supported kernel for the op before proceeding with the assignment- I could not figure out a way to do this, though. I validated this change by building a private version of ORT and running the affected tests in Edge Canary. The crash did not occur for those tests with the fix in place. ### Motivation and Context <!-- - Why is this change required? What problem does it solve? - If it fixes an open issue, please link to the issue here. --> This bug causes crashes in the web browser's GPU sandbox process, which is a blocker for WebNN origin trials. The crashes need to be resolved. Fixes #26531. --------- Co-authored-by: Aditya Rastogi <[email protected]>
adrastogi
added a commit
that referenced
this pull request
Jan 14, 2026
* **Address DML crashes in WebNN QDQ subgraph tests (#26822)** * **Update to next 15.2.6 (e7d04b949bd4550694a72af365ca97fa3ab4bd7d)** --------- Co-authored-by: Aditya Rastogi <[email protected]> Co-authored-by: Kevin Taha <[email protected]>
alex-spacemit
pushed a commit
to spacemit-com/onnxruntime
that referenced
this pull request
Jan 20, 2026
### Description <!-- Describe your changes. --> This change proposes a fix to some DML crashes observed in WebNN QDQ subgraph tests. The root cause of the problem as analyzed in the issue linked below is that the DML EP does not have kernels implemented for the QLinear versions of LeakyRelu and Softmax. The QDQ transformer has logic to fuse a 3-node sequence in the graph to the associated QLinear custom op. There is an operator registry which contains the mapping between EPs and the ops that the transformer consults for computing the quantized variants. The registry currently has the CPU and DML EPs supporting the same sets of operators (thereby assuming that the fused node would also be supported by the EP), so what I did was split out a separate registration for the LeakyRelu and Softmax ops and associate them with the CPU EP, while removing those ops from the prior list. I was trying to also explore other options like seeing if I could somehow check to see if the EP has a supported kernel for the op before proceeding with the assignment- I could not figure out a way to do this, though. I validated this change by building a private version of ORT and running the affected tests in Edge Canary. The crash did not occur for those tests with the fix in place. ### Motivation and Context <!-- - Why is this change required? What problem does it solve? - If it fixes an open issue, please link to the issue here. --> This bug causes crashes in the web browser's GPU sandbox process, which is a blocker for WebNN origin trials. The crashes need to be resolved. Fixes microsoft#26531. --------- Co-authored-by: Aditya Rastogi <[email protected]>
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Description
This change proposes a fix to some DML crashes observed in WebNN QDQ subgraph tests. The root cause of the problem as analyzed in the issue linked below is that the DML EP does not have kernels implemented for the QLinear versions of LeakyRelu and Softmax. The QDQ transformer has logic to fuse a 3-node sequence in the graph to the associated QLinear custom op. There is an operator registry which contains the mapping between EPs and the ops that the transformer consults for computing the quantized variants.
The registry currently has the CPU and DML EPs supporting the same sets of operators (thereby assuming that the fused node would also be supported by the EP), so what I did was split out a separate registration for the LeakyRelu and Softmax ops and associate them with the CPU EP, while removing those ops from the prior list.
I was trying to also explore other options like seeing if I could somehow check to see if the EP has a supported kernel for the op before proceeding with the assignment- I could not figure out a way to do this, though.
I validated this change by building a private version of ORT and running the affected tests in Edge Canary. The crash did not occur for those tests with the fix in place.
Motivation and Context
This bug causes crashes in the web browser's GPU sandbox process, which is a blocker for WebNN origin trials. The crashes need to be resolved.
Fixes #26531.