Native WebGPU EP fails to run model with in-memory external data

[huningxin]

Description

For models that are built by ModelEditorAPI, they may have initializers that are TensorProto with external data in existing memory buffer. For those TensorProto, the 'location' field of the external data info is set to a special marker */_ORT_MEM_ADDR_/*, and the 'offset' field contains the address of the memory buffer.

When running those models with native WebGPU EP, it reports error, for example the error log of running a "Conv" node with initializer using in-memory external data:

"ORT status error code: 6 error message: Non-zero status code returned while running Conv node. Name:'_0_token_1' Status Message: C:\workspace\onnxruntime\onnxruntime\core\providers\webgpu\webgpu_context.cc:210 onnxruntime::webgpu::WebGpuContext::Run std::all_of(inputs.begin(), inputs.end(), [](const ProgramInput& input) { const auto* tensor = input.tensor; return tensor != nullptr && tensor->Location().mem_type == OrtMemType::OrtMemTypeDefault && tensor->Location().device.Type() == OrtDevice::GPU && !strcmp(tensor->Location().name, WEBGPU_BUFFER); }) was false. All inputs must be tensors on WebGPU buffers."

It seems WebGPU EP doesn't upload the in-memory external weights to GPU buffer before running it.

/cc @fdwr @skottmckay @guschmue

[fs-eire]

could you please share the repro steps ? It would be great if I can start debugging with a small model.

[huningxin]

I met this issue by using Chromium WebNN ONNX Runtime backend prototype: https://github.com/shiyi9801/chromium/tree/ort_backend, with WebGPU EP.

To reproduce it, it needs to build the Chromium and test one of WebNN samples, for example image classification, with GPU device type: https://webmachinelearning.github.io/webnn-samples/image_classification/

I know it is not easy to build a Chromium. I'll try to find a simpler reproducer.

[huningxin]

This issue can be reproduced by enabling WebGPU EP in Model Editor API test ModelEditorAPITest.Basic_CApi.

Steps:
Build ORT with WebGPU EP support

$ ./build.sh --config Release --build_dir build/WGPU --use_webgpu --build_shared_lib --parallel

Enable WebGPU EP in Model Editor API test:
main...huningxin:onnxruntime:test_model_builder_api_webgpu

Run ModelEditorAPITest.Basic_CApi with following commandline:

$ ./build/WGPU/Release/onnxruntime_shared_lib_test --gtest_filter=ModelEditorAPITest.Basic_CApi

The test fails with the following message
"
[==========] Running 1 test from 1 test suite.
[----------] Global test environment set-up.
[----------] 1 test from ModelEditorAPITest
[ RUN ] ModelEditorAPITest.Basic_CApi
2025-06-03 22:59:42.896178887 [E:onnxruntime:, sequential_executor.cc:572 ExecuteKernel] Non-zero status code returned while running Gemm node. Name:'Gemm1' Status Message: /home/nhu/code/onnxruntime_native/onnxruntime/core/providers/webgpu/webgpu_context.cc:203 onnxruntime::common::Status onnxruntime::webgpu::WebGpuContext::Run(onnxruntime::webgpu::ComputeContext&, const onnxruntime::webgpu::ProgramBase&) std::all_of(inputs.begin(), inputs.end(), [](const ProgramInput& input) { const auto* tensor = input.tensor; return tensor != nullptr && tensor->Location().mem_type == OrtMemType::OrtMemTypeDefault && tensor->Location().device.Type() == OrtDevice::GPU && !strcmp(tensor->Location().name, WEBGPU_BUFFER); }) was false. All inputs must be tensors on WebGPU buffers.

unknown file: Failure
C++ exception with description "Non-zero status code returned while running Gemm node. Name:'Gemm1' Status Message: /home/nhu/code/onnxruntime_native/onnxruntime/core/providers/webgpu/webgpu_context.cc:203 onnxruntime::common::Status onnxruntime::webgpu::WebGpuContext::Run(onnxruntime::webgpu::ComputeContext&, const onnxruntime::webgpu::ProgramBase&) std::all_of(inputs.begin(), inputs.end(), [](const ProgramInput& input) { const auto* tensor = input.tensor; return tensor != nullptr && tensor->Location().mem_type == OrtMemType::OrtMemTypeDefault && tensor->Location().device.Type() == OrtDevice::GPU && !strcmp(tensor->Location().name, WEBGPU_BUFFER); }) was false. All inputs must be tensors on WebGPU buffers.
" thrown in the test body.

[ FAILED ] ModelEditorAPITest.Basic_CApi (237 ms)
[----------] 1 test from ModelEditorAPITest (237 ms total)

[----------] Global test environment tear-down
[==========] 1 test from 1 test suite ran. (237 ms total)
[ PASSED ] 0 tests.
[ FAILED ] 1 test, listed below:
[ FAILED ] ModelEditorAPITest.Basic_CApi

1 FAILED TEST
"

[fs-eire]

I will debug the test and figure out how to fix it.

[fs-eire]

@huningxin could you please help to verify if #23979 fixes the problem?