DNNL Execution Provider produces incorrect output for LayerNormalization with axis = -1

[eKevinHoang]

Describe the issue

Description:

When using the DNNL Execution Provider (EP) in ONNX Runtime, LayerNormalization nodes with the attribute axis=-1 produce incorrect results. Currently, DNNL does not resolve negative axes, which leads to wrong outputs.

Environment:

• ONNX Runtime version: 1.23.1
• Platform:(no CUDA, Intel + AMD Graphics)
• Models with LayerNormalization using axis=-1 (e.g., SAM2)

Expected behavior:

The negative axis should be correctly resolved before DNNL execution, or DNNL EP should handle negative axes to produce correct output.

Current workaround:

Exclude LayerNormalization nodes with axis=-1 when using DNNL EP.

Additional notes:

DNNL provides excellent performance, but this issue causes incorrect results if negative axes are not handled.

To reproduce

1. Load a model containing LayerNormalization nodes with axis=-1.
2. Run inference using the DNNL Execution Provider.
3. Observe that the output differs from other (e.g., CPU)

Urgency

No response

Platform

Mac

OS Version

15.7.1

ONNX Runtime Installation

Built from Source

ONNX Runtime Version or Commit ID

1.23.1

ONNX Runtime API

C++

Architecture

X64

Execution Provider

oneDNN

Execution Provider Library Version

oneDNN-3.9.2

[eKevinHoang]

After looking more closely, the problem is not just about axis = -1. The real cause is that src_mem was being overwritten in-place during the LayerNormalization forward pass. This caused the input to downstream layers to be incorrect, leading to wrong outputs. In my patch attched, I created a separate dst_mem for the LayerNormalization output, leaving the original src_mem intact.

onnxruntime-1.23.1-layernorm.patch

Note: This fixes the issue I observed on my device. It worked for me locally, but I haven’t tested it in other environments.