[GPU] Enable qdq_stripping path for GPU #694
Conversation
MayureshV1
requested review from
ankitm3k,
Copilot,
javier-intel,
preetha-intel and
saurabhkale17
There was a problem hiding this comment.
Pull Request Overview
This PR enables the QDQ stripping path for GPU devices by updating condition checks in both capability.cc and backend_manager.cc.
- Extends support for GPU devices in QDQ optimizer logic.
- Updates condition checks to include "GPU" in device type validations.
Reviewed Changes
Copilot reviewed 2 out of 2 changed files in this pull request and generated no comments.
| File | Description |
|---|---|
| onnxruntime/core/providers/openvino/ov_versions/capability.cc | Updated condition check to support GPUs |
| onnxruntime/core/providers/openvino/backend_manager.cc | Updated condition check and comment to support GPUs |
Comments suppressed due to low confidence (1)
onnxruntime/core/providers/openvino/backend_manager.cc:383
- Update this comment to reflect that QDQ stripping is now enabled for both NPU and GPU devices.
// QDQ stripping enabled only for the NPU
|
@intbf , This is a straight forward PR which I can approve for experimentation but I have a fundamental question. Today we have logic to check for QDQ stripping support in NPU compiler. If it does not pass then we use OVEP stripping logic which was designed for NPU. Would the exact same ruleset work for GPU? |
Signed-off-by: bfilipek <[email protected]>
|
@MayureshV1 according to the jira, the plan is to test this implementation and then adapt rules if needed. But I run some tests on a few models and they runs on GPU with the flag enabled. |
MayureshV1
left a comment
MayureshV1
left a comment
There was a problem hiding this comment.
Changes look good to run OVEP QDQ stripping on GPU. Ruleset needs to be verified !
Experimental feature for GPU.
* update the statement so that we run CreateModelWithStrippedQDQNodes on GPU * ensure the capability checks are also updated * update the comment Signed-off-by: bfilipek <[email protected]> --------- Signed-off-by: bfilipek <[email protected]>
2 participants
Try enabling enable_qdq_optimizer for the GPU device
Description
Update the conditions so that we can run CreateModelWithStrippedQDQNodes for the GPU device
Motivation and Context
Implementation for https://jira.devtools.intel.com/browse/CVS-167484
It uses (initially) the same rules as implemented in microsoft#20622