IAttention failure of TensorRT 10.14 when running serilization on GPU RTX4080

Description

Cannot use IAttention. Originally i tried build network SDK (i.e. network->addAttention), it failed, so i did some tests.

I build a dummy attention layer with ONNX opset v23. Error when exporting.

command:

trtexec --verbose --stronglyTyped --onnx=models/attn_export/1attn_attention_opset23.onnx

1 layer:

[02/26/2026-09:23:18] [I] Start parsing network model.
[02/26/2026-09:23:18] [I] [TRT] ----------------------------------------------------------------
[02/26/2026-09:23:18] [I] [TRT] Input filename:   models/attn_export/1attn_attention_opset23.onnx
[02/26/2026-09:23:18] [I] [TRT] ONNX IR version:  0.0.11
[02/26/2026-09:23:18] [I] [TRT] Opset version:    23
[02/26/2026-09:23:18] [I] [TRT] Producer name:    
[02/26/2026-09:23:18] [I] [TRT] Producer version: 
[02/26/2026-09:23:18] [I] [TRT] Domain:           
[02/26/2026-09:23:18] [I] [TRT] Model version:    0
[02/26/2026-09:23:18] [I] [TRT] Doc string:       
[02/26/2026-09:23:18] [I] [TRT] ----------------------------------------------------------------
[02/26/2026-09:23:18] [V] [TRT] Adding network input: x with dtype: float16, dimensions: (1, 8, 128, 512)
[02/26/2026-09:23:18] [V] [TRT] Registering tensor: x for ONNX tensor: x
[02/26/2026-09:23:18] [V] [TRT] Static check for parsing node: attn_1 [Attention]
[02/26/2026-09:23:18] [V] [TRT] Parsing node: attn_1 [Attention]
[02/26/2026-09:23:18] [V] [TRT] Searching for input: x
[02/26/2026-09:23:18] [V] [TRT] Searching for input: x
[02/26/2026-09:23:18] [V] [TRT] Searching for input: x
[02/26/2026-09:23:18] [V] [TRT] attn_1 [Attention] inputs: [x -> (1, 8, 128, 512)[HALF]], [x -> (1, 8, 128, 512)[HALF]], [x -> (1, 8, 128, 512)[HALF]], 
[02/26/2026-09:23:18] [V] [TRT] Registering layer: ONNXTRT_castHelper required by ONNX-TRT
[02/26/2026-09:23:18] [V] [TRT] Registering layer: ONNXTRT_castHelper_1 required by ONNX-TRT
[02/26/2026-09:23:18] [V] [TRT] Registering tensor: y1_2 for ONNX tensor: y1
[02/26/2026-09:23:18] [V] [TRT] attn_1 [Attention] outputs: [y1 -> (1, 8, 128, 512)[HALF]], 
[02/26/2026-09:23:18] [V] [TRT] Marking y1_2 as output: y1
[02/26/2026-09:23:18] [I] Finished parsing network model. Parse time: 0.000553871
[02/26/2026-09:23:18] [V] [TRT] could not open /sys/fs/cgroup/memory/memory.limit_in_bytes or /sys/fs/cgroup/memory.max
[02/26/2026-09:23:18] [V] [TRT] Trying to load shared library libnvinfer_builder_resource_sm89.so.10.14.1
[02/26/2026-09:23:18] [V] [TRT] Loaded shared library libnvinfer_builder_resource_sm89.so.10.14.1
[02/26/2026-09:23:18] [I] [TRT] [MemUsageChange] Init builder kernel library: CPU +204, GPU +4, now: CPU 434, GPU 1154 (MiB)
[02/26/2026-09:23:18] [V] [TRT] Original: 8 layers
[02/26/2026-09:23:18] [V] [TRT] After dead-layer removal: 8 layers
[02/26/2026-09:23:18] [V] [TRT] SYMBOLIC CHECKS
[02/26/2026-09:23:18] [V] [TRT] GRAPH NODES
[02/26/2026-09:23:18] [V] [TRT] CONSTANT (Unnamed Layer* 0) [Constant]
[02/26/2026-09:23:18] [V] [TRT]     Output 0
[02/26/2026-09:23:18] [V] [TRT]         1       1
[02/26/2026-09:23:18] [V] [TRT]         1       1
[02/26/2026-09:23:18] [V] [TRT]         1       1
[02/26/2026-09:23:18] [V] [TRT]         1       1
[02/26/2026-09:23:18] [V] [TRT] CAST ONNXTRT_castHelper
[02/26/2026-09:23:18] [V] [TRT]     Input 0
[02/26/2026-09:23:18] [V] [TRT]         1       1
[02/26/2026-09:23:18] [V] [TRT]         1       1
[02/26/2026-09:23:18] [V] [TRT]         1       1
[02/26/2026-09:23:18] [V] [TRT]         1       1
[02/26/2026-09:23:18] [V] [TRT]     Output 0
[02/26/2026-09:23:18] [V] [TRT]         1       1
[02/26/2026-09:23:18] [V] [TRT]         1       1
[02/26/2026-09:23:18] [V] [TRT]         1       1
[02/26/2026-09:23:18] [V] [TRT]         1       1
[02/26/2026-09:23:18] [V] [TRT] ELEMENTWISE (Unnamed Layer* 2) [ElementWise]
[02/26/2026-09:23:18] [V] [TRT]     Input 0
[02/26/2026-09:23:18] [V] [TRT]         1       1
[02/26/2026-09:23:18] [V] [TRT]         8       8
[02/26/2026-09:23:18] [V] [TRT]         128     128
[02/26/2026-09:23:18] [V] [TRT]         512     512
[02/26/2026-09:23:18] [V] [TRT]     Input 1
[02/26/2026-09:23:18] [V] [TRT]         1       1
[02/26/2026-09:23:18] [V] [TRT]         1       1
[02/26/2026-09:23:18] [V] [TRT]         1       1
[02/26/2026-09:23:18] [V] [TRT]         1       1
[02/26/2026-09:23:18] [V] [TRT]     Output 0
[02/26/2026-09:23:18] [V] [TRT]         1       1
[02/26/2026-09:23:18] [V] [TRT]         8       8
[02/26/2026-09:23:18] [V] [TRT]         128     128
[02/26/2026-09:23:18] [V] [TRT]         512     512
[02/26/2026-09:23:18] [V] [TRT] CONSTANT (Unnamed Layer* 3) [Constant]
[02/26/2026-09:23:18] [V] [TRT]     Output 0
[02/26/2026-09:23:18] [V] [TRT]         1       1
[02/26/2026-09:23:18] [V] [TRT]         1       1
[02/26/2026-09:23:18] [V] [TRT]         1       1
[02/26/2026-09:23:18] [V] [TRT]         1       1
[02/26/2026-09:23:18] [V] [TRT] CAST ONNXTRT_castHelper_1
[02/26/2026-09:23:18] [V] [TRT]     Input 0
[02/26/2026-09:23:18] [V] [TRT]         1       1
[02/26/2026-09:23:18] [V] [TRT]         1       1
[02/26/2026-09:23:18] [V] [TRT]         1       1
[02/26/2026-09:23:18] [V] [TRT]         1       1
[02/26/2026-09:23:18] [V] [TRT]     Output 0
[02/26/2026-09:23:18] [V] [TRT]         1       1
[02/26/2026-09:23:18] [V] [TRT]         1       1
[02/26/2026-09:23:18] [V] [TRT]         1       1
[02/26/2026-09:23:18] [V] [TRT]         1       1
[02/26/2026-09:23:18] [V] [TRT] ELEMENTWISE (Unnamed Layer* 5) [ElementWise]
[02/26/2026-09:23:18] [V] [TRT]     Input 0
[02/26/2026-09:23:18] [V] [TRT]         1       1
[02/26/2026-09:23:18] [V] [TRT]         8       8
[02/26/2026-09:23:18] [V] [TRT]         128     128
[02/26/2026-09:23:18] [V] [TRT]         512     512
[02/26/2026-09:23:18] [V] [TRT]     Input 1
[02/26/2026-09:23:18] [V] [TRT]         1       1
[02/26/2026-09:23:18] [V] [TRT]         1       1
[02/26/2026-09:23:18] [V] [TRT]         1       1
[02/26/2026-09:23:18] [V] [TRT]         1       1
[02/26/2026-09:23:18] [V] [TRT]     Output 0
[02/26/2026-09:23:18] [V] [TRT]         1       1
[02/26/2026-09:23:18] [V] [TRT]         8       8
[02/26/2026-09:23:18] [V] [TRT]         128     128
[02/26/2026-09:23:18] [V] [TRT]         512     512
[02/26/2026-09:23:18] [V] [TRT] ATTENTION_INPUT (Unnamed Layer* 7) [AttentionInput]
[02/26/2026-09:23:18] [V] [TRT]     Input 0
[02/26/2026-09:23:18] [V] [TRT]         1       1
[02/26/2026-09:23:18] [V] [TRT]         8       8
[02/26/2026-09:23:18] [V] [TRT]         128     128
[02/26/2026-09:23:18] [V] [TRT]         512     512
[02/26/2026-09:23:18] [V] [TRT]     Input 1
[02/26/2026-09:23:18] [V] [TRT]         1       1
[02/26/2026-09:23:18] [V] [TRT]         8       8
[02/26/2026-09:23:18] [V] [TRT]         128     128
[02/26/2026-09:23:18] [V] [TRT]         512     512
[02/26/2026-09:23:18] [V] [TRT]     Input 2
[02/26/2026-09:23:18] [V] [TRT]         1       1
[02/26/2026-09:23:18] [V] [TRT]         8       8
[02/26/2026-09:23:18] [V] [TRT]         128     128
[02/26/2026-09:23:18] [V] [TRT]         512     512
[02/26/2026-09:23:18] [V] [TRT]     Output 0
[02/26/2026-09:23:18] [V] [TRT]         1       1
[02/26/2026-09:23:18] [V] [TRT]         8       8
[02/26/2026-09:23:18] [V] [TRT]         128     128
[02/26/2026-09:23:18] [V] [TRT]         512     512
[02/26/2026-09:23:18] [V] [TRT] ATTENTION_OUTPUT (Unnamed Layer* 8) [AttentionOutput]
[02/26/2026-09:23:18] [V] [TRT]     Input 0
[02/26/2026-09:23:18] [V] [TRT]         1       1
[02/26/2026-09:23:18] [V] [TRT]         8       8
[02/26/2026-09:23:18] [V] [TRT]         128     128
[02/26/2026-09:23:18] [V] [TRT]         512     512
[02/26/2026-09:23:18] [V] [TRT]     Output 0
[02/26/2026-09:23:18] [V] [TRT]         1       1
[02/26/2026-09:23:18] [V] [TRT]         8       8
[02/26/2026-09:23:18] [V] [TRT]         128     128
[02/26/2026-09:23:18] [V] [TRT]         512     512
[02/26/2026-09:23:18] [V] [TRT] Graph construction completed in 0.000633728 seconds.
[02/26/2026-09:23:18] [V] [TRT] After adding DebugOutput nodes: 8 layers
[02/26/2026-09:23:18] [V] [TRT] After Myelin optimization: 1 layers
[02/26/2026-09:23:18] [V] [TRT] Applying ScaleNodes fusions.
[02/26/2026-09:23:18] [V] [TRT] After scale fusion: 1 layers
[02/26/2026-09:23:18] [V] [TRT] After dupe layer removal: 1 layers
[02/26/2026-09:23:18] [V] [TRT] After final dead-layer removal: 1 layers
[02/26/2026-09:23:18] [V] [TRT] After tensor merging: 1 layers
[02/26/2026-09:23:18] [V] [TRT] After vertical fusions: 1 layers
[02/26/2026-09:23:18] [V] [TRT] After dupe layer removal: 1 layers
[02/26/2026-09:23:18] [V] [TRT] After final dead-layer removal: 1 layers
[02/26/2026-09:23:18] [V] [TRT] After tensor merging: 1 layers
[02/26/2026-09:23:18] [V] [TRT] After slice removal: 1 layers
[02/26/2026-09:23:18] [V] [TRT] After concat removal: 1 layers
[02/26/2026-09:23:18] [V] [TRT] Trying to split Reshape and strided tensor
[02/26/2026-09:23:18] [V] [TRT] Graph optimization time: 0.000291817 seconds.
[02/26/2026-09:23:18] [V] [TRT] Building graph using backend strategy 2
[02/26/2026-09:23:18] [I] [TRT] Local timing cache in use. Profiling results in this builder pass will not be stored.
[02/26/2026-09:23:18] [V] [TRT] Constructing optimization profile number 0 [1/1].
[02/26/2026-09:23:18] [V] [TRT] Applying generic optimizations to the graph for inference.
[02/26/2026-09:23:18] [V] [TRT] Reserving memory for host IO tensors. Host: 0 bytes
[02/26/2026-09:23:18] [V] [TRT] =============== Computing costs for {ForeignNode[(Unnamed Layer* 0) [Constant]...(Unnamed Layer* 8) [AttentionOutput]]}
[02/26/2026-09:23:18] [V] [TRT] ForeignNode {ForeignNode[(Unnamed Layer* 0) [Constant]...(Unnamed Layer* 8) [AttentionOutput]]} metadata: 
[02/26/2026-09:23:18] [V] [TRT] *************** Autotuning format combination: Half(524288,65536,512,1) -> Half(524288,65536,512,1) ***************
[02/26/2026-09:23:18] [V] [TRT] --------------- Timing Runner: {ForeignNode[(Unnamed Layer* 0) [Constant]...(Unnamed Layer* 8) [AttentionOutput]]} (Myelin[0x80000023])
[02/26/2026-09:23:18] [I] [TRT] Compiler backend is used during engine build.
[1]    13731 segmentation fault (core dumped)  trtexec --verbose --stronglyTyped

2 layers:

[02/26/2026-09:24:23] [I] Start parsing network model.
[02/26/2026-09:24:23] [I] [TRT] ----------------------------------------------------------------
[02/26/2026-09:24:23] [I] [TRT] Input filename:   models/attn_export/2attn_attention_opset23.onnx
[02/26/2026-09:24:23] [I] [TRT] ONNX IR version:  0.0.11
[02/26/2026-09:24:23] [I] [TRT] Opset version:    23
[02/26/2026-09:24:23] [I] [TRT] Producer name:    
[02/26/2026-09:24:23] [I] [TRT] Producer version: 
[02/26/2026-09:24:23] [I] [TRT] Domain:           
[02/26/2026-09:24:23] [I] [TRT] Model version:    0
[02/26/2026-09:24:23] [I] [TRT] Doc string:       
[02/26/2026-09:24:23] [I] [TRT] ----------------------------------------------------------------
[02/26/2026-09:24:23] [V] [TRT] Adding network input: x with dtype: float16, dimensions: (1, 32, 256, 512)
[02/26/2026-09:24:23] [V] [TRT] Registering tensor: x for ONNX tensor: x
[02/26/2026-09:24:23] [V] [TRT] Static check for parsing node: attn_1 [Attention]
[02/26/2026-09:24:23] [V] [TRT] Parsing node: attn_1 [Attention]
[02/26/2026-09:24:23] [V] [TRT] Searching for input: x
[02/26/2026-09:24:23] [V] [TRT] Searching for input: x
[02/26/2026-09:24:23] [V] [TRT] Searching for input: x
[02/26/2026-09:24:23] [V] [TRT] attn_1 [Attention] inputs: [x -> (1, 32, 256, 512)[HALF]], [x -> (1, 32, 256, 512)[HALF]], [x -> (1, 32, 256, 512)[HALF]], 
[02/26/2026-09:24:23] [V] [TRT] Registering layer: ONNXTRT_castHelper required by ONNX-TRT
[02/26/2026-09:24:23] [V] [TRT] Registering layer: ONNXTRT_castHelper_1 required by ONNX-TRT
[02/26/2026-09:24:23] [V] [TRT] Registering tensor: y1 for ONNX tensor: y1
[02/26/2026-09:24:23] [V] [TRT] attn_1 [Attention] outputs: [y1 -> (1, 32, 256, 512)[HALF]], 
[02/26/2026-09:24:23] [V] [TRT] Static check for parsing node: attn_2 [Attention]
[02/26/2026-09:24:23] [V] [TRT] Parsing node: attn_2 [Attention]
[02/26/2026-09:24:23] [V] [TRT] Searching for input: y1
[02/26/2026-09:24:23] [V] [TRT] Searching for input: y1
[02/26/2026-09:24:23] [V] [TRT] Searching for input: y1
[02/26/2026-09:24:23] [V] [TRT] attn_2 [Attention] inputs: [y1 -> (1, 32, 256, 512)[HALF]], [y1 -> (1, 32, 256, 512)[HALF]], [y1 -> (1, 32, 256, 512)[HALF]], 
[02/26/2026-09:24:23] [V] [TRT] Registering layer: ONNXTRT_castHelper_3 required by ONNX-TRT
[02/26/2026-09:24:23] [V] [TRT] Registering layer: ONNXTRT_castHelper_5 required by ONNX-TRT
[02/26/2026-09:24:23] [E] Error[3]: INetworkDefinition::addAttention: Error Code 3: API Usage Error (Parameter check failed, condition: mScopedOps.size() == mGraph.scopedOps.size().  In addAttention at /_src/optimizer/api/network.cpp:2125)
[02/26/2026-09:24:23] [E] [TRT] ModelImporter.cpp:135: While parsing node number 1 [Attention -> "y2"]:
[02/26/2026-09:24:23] [E] [TRT] ModelImporter.cpp:138: --- Begin node ---
input: "y1"
input: "y1"
input: "y1"
output: "y2"
name: "attn_2"
op_type: "Attention"
attribute {
  name: "scale"
  f: 0.0441941731
  type: FLOAT
}

[02/26/2026-09:24:23] [E] [TRT] ModelImporter.cpp:139: --- End node ---
[02/26/2026-09:24:23] [E] [TRT] ModelImporter.cpp:141: ERROR: ModelImporter.cpp:362 In function parseNode:
[6] Invalid Node - attn_2
INetworkDefinition::addAttention: Error Code 3: API Usage Error (Parameter check failed, condition: mScopedOps.size() == mGraph.scopedOps.size().  In addAttention at /_src/optimizer/api/network.cpp:2125)
[02/26/2026-09:24:23] [E] Failed to parse onnx file
[02/26/2026-09:24:23] [I] Finished parsing network model. Parse time: 0.116632
[02/26/2026-09:24:23] [E] Parsing model failed
[02/26/2026-09:24:23] [E] Failed to create engine from model or file.
[02/26/2026-09:24:23] [E] Engine set up failed
&&&& FAILED TensorRT.trtexec [TensorRT v101401] [b48] # trtexec --verbose --stronglyTyped --onnx=models/attn_export/2attn_attention_opset23.onnx

Environment

TensorRT Version: 10.14.1

NVIDIA GPU: RTX 4080

NVIDIA Driver Version: 590.48.01

CUDA Version: 13.1

CUDNN Version: /

Operating System: Ubuntu 22.04

Python Version (if applicable): 3.12

Baremetal or Container (if so, version): Container, based on torch NGC 25.12

Relevant Files

models:

attn_export.zip

Steps To Reproduce

use

export_2attn_onnx.py

Commands or scripts:

python export_2attn_onnx.py --layers 1 --seq 128 --heads 8

or

python export_2attn_onnx.py --layers 2 --seq 128 --heads 8

Have you tried the latest release?: tried TesnsorRT 10.15 on Windows, same error.

Can this model run on other frameworks? For example run ONNX model with ONNXRuntime (polygraphy run <model.onnx> --onnxrt): works on cpu, but cannot find a suitable onnxruntime-gpu version

Reproduce the crash with trtexec --verbose --stronglyTyped using models/attn_export/1attn_attention_opset23.onnx and 2attn_attention_opset23.onnx. Start from the Attention parsing and compiler-backend build stages shown in the logs; done means the TensorRT engine build no longer ends in a segmentation fault for the supplied models.