[Build] ‘struct onnxruntime::ProviderHostCPU’ has no member named ‘UpsampleBase__AdjustOutputSizeAsPolicy’ when CONTRIB ops are disabled. about onnxruntime HOT 15 CLOSED

The "com.microsoft:EPContext" op is a contrib op, for implementing TensorRT EP.

from onnxruntime.

Thanks for reporting this issue, it seems method AdjustOutputSizeAsPolicy is sometimes excluded if DISABLE_CONTRIB_OPS is on, sometimes not like here https://github.com/microsoft/onnxruntime/blob/main/onnxruntime/core/providers/shared_library/provider_bridge_provider.cc#L586

from onnxruntime.

It indeed builds without onnxruntime_DISABLE_CONTRIB_OPS=ON. Any tips how to fix it for 1.18.1?

from onnxruntime.

I submitted a PR for this.

from onnxruntime.

I tried this fix already, but then I could not create CUDAExecutionProvider:

[E:onnxruntime:Default, provider_bridge_ort.cc:1745 TryGetProviderInfo_CUDA] <redacted>/onnxruntime-1.18.1/onnxruntime/core/session/provider_bridge_ort.cc:1426 onnxruntime::Provider& onnxruntime::ProviderLibrary::Get() [ONNXRuntimeError] : 1 : FAIL : Failed to load library libonnxruntime_providers_cuda.so with error: <redacted>/python3.11/site-packages/onnxruntime/capi/libonnxruntime_providers_cuda.so: undefined symbol: _ZN11onnxruntime4cuda21BuildKernelCreateInfoINS0_57kCudaExecutionProvider_GridSample_kOnnxDomain_ver16_floatEEENS_16KernelCreateInfoEv

which is after demangling

onnxruntime::KernelCreateInfo onnxruntime::cuda::BuildKernelCreateInfo<onnxruntime::cuda::kCudaExecutionProvider_GridSample_kOnnxDomain_ver16_float>()

from onnxruntime.

A side question: may I know why do you need to set onnxruntime_DISABLE_CONTRIB_OPS to ON? The build option is for reducing binary size. However, CUDA itself is already very large. Disabling contrib ops does not help much.

from onnxruntime.

I will update the PR.

from onnxruntime.

I like the smaller size, but I also appreciate the much faster compilation. I think it's more than 2-3 times faster on my PC, and I'm currently compiling for just one CUDA architecture. This difference includes the fact that I can use more threads, as compiling contrib kernels takes up too much RAM.

from onnxruntime.

Would you mind trying it again? Now it should be good.

from onnxruntime.

I applied the patch to version 1.18.1 with TRT 8.6 as described here and CUDAExecutionProvider works.

Probably not related to the PR:
However, when I run my model for the first time with the TensorRT Execution Provider (EP) and contrib ops disabled, it creates an .engine file and throws the following error:

RUNTIME_EXCEPTION : Exception during initialization: <path>/onnxruntime/core/providers/tensorrt/onnx_ctx_model_helper.cc:127 onnx::ModelProto* onnxruntime::CreateCtxModel(const onnxruntime::GraphViewer&, std::string, char*, size_t, int64_t, std::string, std::string, const onnxruntime::logging::Logger*) graph_build.Resolve().IsOK() was false. 

Subsequent runs are successful if the .engine file is present in the cache. However, if I delete the .engine file, it fails again. This issue does not occur when contrib ops are enabled. I will investigate further and maybe open an another issue.

from onnxruntime.

Open onnxruntime/core/providers/tensorrt/onnx_ctx_model_helper.cc, find line 127, replace ORT_ENFORCE with ORT_THROW_IF_ERROR.

The new code should be:

ORT_THROW_IF_ERROR(graph_build.Resolve());

Then please let us know what error message you see.

from onnxruntime.

[E:onnxruntime:, inference_session.cc:2044 operator()] Exception during initialization: <path>/onnxruntime/core/providers/tensorrt/onnx_ctx_model_helper.cc:127 onnx::ModelProto* onnxruntime::CreateCtxModel(const onnxruntime::GraphViewer&, std::string, char*, size_t, int64_t, std::string, std::string, const onnxruntime::logging::Logger*) [ONNXRuntimeError] : 1 : FAIL : Fatal error: com.microsoft:EPContext(-1) is not a registered function/op

from onnxruntime.

Ok .. As it says , it is not supported. You may open a new issue for this error.

from onnxruntime.

The interesting thing is that it will create an .engine file and it works with subsequent runs, until I delete the .engine file, with no errors and correct output.

from onnxruntime.

The issue has been fixed in #20695

from onnxruntime.