[call for help] Porting PHi-2 to HAILO8/8L

General Discussion
, ,
1

Hi, all:

I try to convert a phi-2 model to onnx according to Phi-2 tutorial | onnxruntime. The onnx file was successfully generated and tested.

When I tried to convert the model.onnx to HAR, the hailo parser failed to support those functions defined in onnexruntime as shown below. Was it a limitation or I did it in a wrong way? Thanks!

[Note] the virtual environment is the 2025/4 version.

[info] Current Time: 15:14:28, 04/29/25
[info] CPU: Architecture: x86_64, Model: Intel(R) Core(TM) i7-8700 CPU @ 3.20GHz, Number Of Cores: 12, Utilization: 1.1%
[info] Memory: Total: 31GB, Available: 25GB
[info] System info: OS: Linux, Kernel: 6.8.0-57-generic
[info] Hailo DFC Version: 3.31.0
[info] HailoRT Version: 4.21.0
[info] PCIe: 0000:03:00.0: Number Of Lanes: 4, Speed: 8.0 GT/s PCIe
[info] PCIe: 0000:04:00.0: Number Of Lanes: 4, Speed: 8.0 GT/s PCIe
[info] Running `hailo parser onnx --har-path ./har --hw-arch hailo8l ./example-models/phi2-int4-cpu/model.onnx`
[info] Translation started on ONNX model model
[warning] Large model detected. The graph may contain either a large number of operators, or weight variables with a very large capacity.
[warning] Translation time may be a bit long, and some features may be disabled (e.g. model augmentation, retry simplified model, onnx runtime hailo model extraction, etc.).
Traceback (most recent call last):
  File "/local/workspace/hailo_virtualenv/bin/hailo", line 8, in <module>
    sys.exit(main())
  File "/local/workspace/hailo_virtualenv/lib/python3.10/site-packages/hailo_sdk_client/tools/cmd_utils/main.py", line 111, in main
    ret_val = client_command_runner.run()
  File "/local/workspace/hailo_virtualenv/lib/python3.10/site-packages/hailo_platform/tools/hailocli/main.py", line 64, in run
    return self._run(argv)
  File "/local/workspace/hailo_virtualenv/lib/python3.10/site-packages/hailo_platform/tools/hailocli/main.py", line 104, in _run
    return args.func(args)
  File "/local/workspace/hailo_virtualenv/lib/python3.10/site-packages/hailo_sdk_client/tools/parser_cli.py", line 213, in run
    self._parse(net_name, args, tensor_shapes)
  File "/local/workspace/hailo_virtualenv/lib/python3.10/site-packages/hailo_sdk_client/tools/parser_cli.py", line 297, in _parse
    self.runner.translate_onnx_model(
  File "/local/workspace/hailo_virtualenv/lib/python3.10/site-packages/hailo_sdk_common/states/states.py", line 16, in wrapped_func
    return func(self, *args, **kwargs)
  File "/local/workspace/hailo_virtualenv/lib/python3.10/site-packages/hailo_sdk_client/runner/client_runner.py", line 1187, in translate_onnx_model
    parser.translate_onnx_model(
  File "/local/workspace/hailo_virtualenv/lib/python3.10/site-packages/hailo_sdk_client/sdk_backend/parser/parser.py", line 194, in translate_onnx_model
    onnx.checker.check_model(model)
  File "/local/workspace/hailo_virtualenv/lib/python3.10/site-packages/onnx/checker.py", line 163, in check_model
    C.check_model_path(
onnx.onnx_cpp2py_export.checker.ValidationError: No Op registered for LayerNormalization with domain_version of 14

==> Context: Bad node spec for node. Name: /model/layers.0/input_layernorm/LayerNorm OpType: LayerNormalization
2

Hey @ufocash,

Welcome to the Hailo Community!

You’re running into a known compatibility limitation with Hailo’s ONNX parser — specifically around the LayerNormalization operator.

What Hailo DFC Supports

DFC does support LayerNormalization, but there’s a catch:

In the Hailo Dataflow Compiler (DFC) v3.30.0 and above, LayerNormalization isn’t a native op. Instead, it’s broken down into a subgraph of primitive operations (mean, variance, normalization, scale, shift). This decomposition only works under specific conditions and ONNX opset versions.

The error you’re seeing:

onnx.onnx_cpp2py_export.checker.ValidationError: No Op registered for LayerNormalization with domain_version of 14

means that the ONNX parser can’t handle this op — most likely due to either:

  • Using an unsupported opset version (e.g. opset 14+), or

  • Having LayerNormalization under a custom domain (e.g. com.microsoft), which the parser ignores.

  • Most likely, your model uses LayerNormalization from opset 14 or newer.

  • Some models (like Phi-2 or ONNX Runtime exports) wrap this op with extra metadata or domain tags that Hailo doesn’t recognize.

How to Fix It

1. Simplify the ONNX model

Run:

python3 -m onnxsim model.onnx model-simplified.onnx

This may break down LayerNormalization into supported primitive ops — giving the Hailo parser a better chance of handling it.

2. Replace LayerNormalization manually (pre- or post-export)

If simplification doesn’t help, rebuild the op manually as a subgraph:

mean → sub → pow → mean → add ε → sqrt → div → scale + bias

These are all ops that DFC supports natively.
You can do this manually in PyTorch before export, or with a post-processing ONNX graph tool.