Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Fix hybrid quantization reg issue #27687

Conversation

hyunback
Copy link
Contributor

Details:

Tickets:

  • ticket-id

vladimir-paramuzov and others added 8 commits November 5, 2024 19:47
### Details:
- set LPT callbacks to handle compression and avoid constant folding for
it (taken from openvinotoolkit#20973)
 - Allow u8/i8 output data type for compressed onednn FC
- Disable Dequantize propagation through Transpose if it's a dependency
of SDPA to keep Transpose+SDPA fusion
Many daily int8 models have Perf regression.(incorrect conv data type and bias)
Fix kernel selection issue.

Signed-off-by: hyunback <[email protected]>
Convolution is expected int8 data type in int8 model, but with mixed weight compressed occurs, it run fp16.

Signed-off-by: hyunback <[email protected]>
Currently dynamic quantized int8 onednn convolution has the problem. Working with ref convolution.
So replace to run fp16 mode instead.

Signed-off-by: hyunback <[email protected]>
@hyunback hyunback added the category: GPU OpenVINO GPU plugin label Nov 22, 2024
@hyunback hyunback requested review from a team as code owners November 22, 2024 00:05
@isanghao isanghao added this pull request to the merge queue Nov 29, 2024
Merged via the queue into openvinotoolkit:releases/2024/5 with commit 868a568 Nov 29, 2024
285 checks passed
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