[Build] Issue with "include_ops_by_config" and DynamicQuantizeMatMul in WASM CPU execution provider

[sevagh]

Describe the issue

Here's my required_operators_and_types.config file:

ai.onnx;6;InstanceNormalization
ai.onnx;10;MatMulInteger
ai.onnx;11;Conv{"inputs": {"0": ["float"]}},ConvTranspose,DynamicQuantizeLinear{"outputs": {"0": ["uint8_t"]}}
ai.onnx;13;Cast{"inputs": {"0": ["int32_t"]}, "outputs": {"0": ["float"]}},Erf,Gather{"inputs": {"0": ["float"], "1": ["int64_t"]}},Gemm{"inputs": {"0": ["float"]}},MatMul{"inputs": {"0": ["float"]}},ReduceMean{"inputs": {"0": ["float"]}},Sigmoid{"inputs": {"0": ["float"]}},Slice{"inputs": {"0": ["float"], "1": ["int64_t"]}},Softmax{"inputs": {"0": ["float"]}},Split{"inputs": {"0": ["float"]}},Sqrt{"inputs": {"0": ["float"]}},Squeeze,Transpose{"inputs": {"0": ["float"]}},Unsqueeze
ai.onnx;14;Add{"inputs": {"0": ["float"]}},Div{"inputs": {"0": ["float"]}},Mul{"inputs": {"0": ["float"]}},Reshape,Sub{"inputs": {"0": ["float"]}}
ai.onnx;17;LayerNormalization
com.microsoft;1;BiasGelu,DynamicQuantizeMatMul,FusedMatMul,Gelu

When I build onnxruntime for wasm without the config file, ORT inference works:

python ./vendor/onnxruntime/tools/ci_build/build.py \
    --build_dir="./build/build-ort-wasm-simd" \
    --config=MinSizeRel \
    --build_wasm_static_lib \
    --parallel \
    --minimal_build \
    --disable_ml_ops \
    --disable_rtti \
    --use_preinstalled_eigen \
    --eigen_path=$(realpath "./vendor/eigen") \
    --skip_tests \
    --enable_wasm_simd
    #--include_ops_by_config="./onnx-models/required_operators_and_types.config" \
    #--enable_reduced_operator_type_support \
;;

When I build onnxruntime for wasm with the config file, ORT inference gives the following error:

2024-11-07 09:14:13.284000 [I:onnxruntime:, inference_session.cc:1699 Initialize] Initializing session. demucs_onnx_simd.js:1751:12
2024-11-07 09:14:13.284000 [I:onnxruntime:, inference_session.cc:1736 Initialize] Adding default CPU execution provider. demucs_onnx_simd.js:1751:12
Ort::Exception: Could not find an implementation for DynamicQuantizeMatMul(1) node with name '/crosstransformer/layers.0/self_attn/MatMul_quant'

In the cmake output, we see the file that contains this operator being replaced:

File '/home/sevagh/repos/demucs.onnx-pro/vendor/onnxruntime/onnxruntime/contrib_ops/cpu/cpu_contrib_kernels.cc' substituted with reduced op version '/home/sevagh/repos/demucs.onnx-pro/build/build-ort-wasm-simd/MinSizeRel/op_reduction.generated/onnxruntime/contrib_ops/cpu/cpu_contrib_kernels.cc'.

In this replacement file, there are some lines related to DynamicQuantizeMatMul:

$ rg 'DynamicQuantizeMatMul' build/build-ort-wasm-simd/build/build-ort-wasm-simd/MinSizeRel/op_reduction.generated/onnxruntime/contrib_ops/cpu/cpu_contrib_kernels.cc
95:class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCpuExecutionProvider, kMSDomain, 1, float, DynamicQuantizeMatMul);
260:      BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCpuExecutionProvider, kMSDomain, 1, float, DynamicQuantizeMatMul)>,

Urgency

No response

Target platform

WebAssembly

Build script

python ./vendor/onnxruntime/tools/ci_build/build.py
--build_dir="./build/build-ort-wasm-simd"
--config=MinSizeRel
--build_wasm_static_lib
--parallel
--minimal_build
--disable_ml_ops
--disable_rtti
--use_preinstalled_eigen
--eigen_path=$(realpath "./vendor/eigen")
--skip_tests
--enable_wasm_simd
#--include_ops_by_config="./onnx-models/required_operators_and_types.config"
#--enable_reduced_operator_type_support \

Error / output

Ort::Exception: Could not find an implementation for DynamicQuantizeMatMul(1) node with name '/crosstransformer/layers.0/self_attn/MatMul_quant'

Visual Studio Version

No response

GCC / Compiler Version

No response