CVS-175447-[OVEP] Add a check for type mismatches in QDQ stripping

[mdvoretc-intel]

Description

When rewiring the graph after eliminating QDQ pairs, the runtime now checks whether the type matches before and after the eliminated nodes and inserts a Cast node if there is a mismatch.

Motivation and Context

At present, QDQ elimination assumes the floating point type is the same before the QuantizeLinear node and after the following DequantizeLinear, producing errors if the types mismatch.

If feature goes to new ABI?

Yes

Jira Ticket :

CVS-175447

[Kotomi-Du]

The casting node is for QuantizeLinear node correct? Could you add the before and after graph after this casting? Will it introduce any performance penalty?

[mdvoretc-intel]

The cast is introduced when rewiring the graph to remove the QuantizeLinear and DequantizeLinear nodes, so that instead of a direct edge from e.g. a f32 input to a f16 output there's a Cast node added between them. The before and after are:


The resulting Convert node in OV is then optimized away in graph transformation stage, adding no specific performance impact.

[mklimenk]

Please use the output type of the DQ node directly and fix the C++ Lint warnings (C-style casts).

[mdvoretc-intel]

The issue that causes the use of a C-style cast for the input arg is that it is provided as const by OutputDefs but cannot be accepted as such by AddNode.

[mklimenk]

Let's roll back this unordered_map which is mainly unused. As for the C-style cast, this warning could be potentially suppressed by using the const_cast<T&>().

[mklimenk]

Looks good to me, thanks!

[Kotomi-Du]

Could you open the PR to get review from reviewers with write access? Please also add the Jira ticket.

[Kotomi-Du]

@mklimenk This feature only impacts GPU, no impact on NPU right?

[mklimenk]

@mklimenk This feature only impacts GPU, no impact on NPU right?

Yes, this code isn't called by NPU path

[Copilot]

Pull Request Overview

This PR adds type mismatch detection and automatic Cast node insertion during QDQ (Quantize-Dequantize) pair elimination in the OpenVINO provider. The change addresses issues where floating point types differed before QuantizeLinear and after DequantizeLinear nodes.

Key Changes:

• Added type comparison logic before rewiring graph edges after QDQ elimination
• Inserted automatic Cast node creation when type mismatches are detected between float and float16

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[MayureshV1]

Thanks for making the modifications. Looks Good !