Skip to content

Conversation

@saurabhkale17
Copy link

@saurabhkale17 saurabhkale17 commented Mar 12, 2025

Issue

The QuantizedLinear node was being stripped off even when it was feeding into the graph output, which expected a quantized input. This occurred despite the presence of a check for QuantizedLinear feeding into src graph output. The root causes of this issue were:

  1. The order of execution in the Quantization rule set.
  2. Handling of the uint16 data type, which was not correctly accounted for.

Fix

This PR addresses both issues by:

  • Ensuring that the QuantizedLinear node is retained when it is required for graph output.
  • Correcting the execution order of the Quantization rule set.
  • Properly handling the uint16 data type to prevent unintended stripping.

This fix ensures that quantized models function correctly without unnecessary node removal, maintaining the integrity of the quantized computation.

https://jira.devtools.intel.com/browse/EISW-159866

@sfatimar
Copy link

@saurabhkale17 please update branch

@saurabhkale17 saurabhkale17 force-pushed the saurabh/stripping_issue branch from 5a0692d to e1053ff Compare March 18, 2025 12:29
@saurabhkale17
Copy link
Author

I have rebase the branch @sfatimar

@sfatimar sfatimar merged commit 23e17e2 into ovep-develop Mar 18, 2025
6 of 12 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

4 participants