TensorRT backend not work when the device_id is greater than 0

[jax11235]

Describe the issue

Cuda backend runs fine on multiple gpus, but TensorRT fails when the device_id is specified as a value greater than 0.
One successful workaround I have tried so far is using multiple processes, and each using different environment variables CUDA_VISIBLE_DEVICES=device_id.

To reproduce

...

Urgency

Urgent, because the project is organized in multiple threads, and no workaround works.

Platform

Linux

OS Version

20.04

ONNX Runtime Installation

Released Package

ONNX Runtime Version or Commit ID

1.16.3 and 1.17.0

ONNX Runtime API

C++

Architecture

X64

Execution Provider

TensorRT

Execution Provider Library Version

TensorRT 8.6

[chilo-ms]

Some conversations here
#16274

[chilo-ms]

Can you share your C++ code of how you set device id?

We suggest to use session_options.AppendExecutionProvider_TensorRT_V2(tensorrt_options);
please see here for reference:
https://onnxruntime.ai/docs/execution-providers/TensorRT-ExecutionProvider.html#click-below-for-c-api-example

[jax11235]

This is part of my code to append tensorrt backend to sess_options:

    OrtApi const& ortApi = Ort::GetApi(); // Uses ORT_API_VERSION
    OrtTensorRTProviderOptionsV2* tensorrt_options = nullptr;
    Ort::ThrowOnError(ortApi.CreateTensorRTProviderOptions(&tensorrt_options));
    std::vector<const char*> keys{"device_id", "trt_max_workspace_size", "trt_fp16_enable", "trt_engine_cache_enable", "trt_engine_cache_path", "trt_timing_cache_enable"};
    std::vector<const char*> values{device_id.c_str(), memory_limit.c_str(), "1", "1", tensorrt_engine_cache_path.c_str(), "1"};
    Ort::ThrowOnError(ortApi.UpdateTensorRTProviderOptions(tensorrt_options, keys.data(), values.data(), keys.size()));
    Ort::ThrowOnError(ortApi.SessionOptionsAppendExecutionProvider_TensorRT_V2(so, tensorrt_options));
    ortApi.ReleaseTensorRTProviderOptions(tensorrt_options);

[chilo-ms]

Thanks, the code to append TRT EP to session options looks okay.

Could you also share the

• application code where it creates multiple sessions and specifies different device_id to intialize session with TRT EP? and where you set CUDA_VISIBLE_DEVICES
• the error message

Please note that users can only specify device_id at session initialization time and it's one gpu for one session.
Once the session is created, you can't change the device. It means when you call session.Run(), TRT EP always uses that device to perform inference, i think even though you set CUDA_VISIBLE_DEVICES=device_id before session.Run(), it won't have effect.

I'm still not sure what your application code is for now.
But we tested from our side, TRT EP can run different device within one session through provider options "device_id":
onnxruntime_perf_test -e tensorrt -r 1 -i "device_id|2" model.onnx

[chilo-ms]

Reply other discussion thread here

What I mean is that different threads uses different sessions, but if the first thread init the session with the environment variable CUDA_VISIBLE_DEVICES, subsequent threads can not change the device by modifying CUDA_VISIBLE_DEVICES, they use the same device with the first thread.

"subsequent threads can not change device by modifying CUDA_VISIBLE_DEVICES", other threads try to change the device at session creation? Or session.Run()?

[jax11235]

Thanks, the code to append TRT EP to session options looks okay.

Could you also share the

• application code where it creates multiple sessions and specifies different device_id to intialize session with TRT EP? and where you set CUDA_VISIBLE_DEVICES
• the error message

Please note that users can only specify device_id at session initialization time and it's one gpu for one session. Once the session is created, you can't change the device. It means when you call session.Run(), TRT EP always uses that device to perform inference, i think even though you set CUDA_VISIBLE_DEVICES=device_id before session.Run(), it won't have effect.

I'm still not sure what your application code is for now. But we tested from our side, TRT EP can run different device within one session through provider options "device_id": onnxruntime_perf_test -e tensorrt -r 1 -i "device_id|2" model.onnx

Without setting CUDA_VISIBLE_DEVICES, the way I create sessions is equivalent to creating a session array, each session with a different device_id appended to a new session_option.
The error message is TensorRT EP execution context enqueue failed, the same as #16274.

[jax11235]

Reply other discussion thread here

What I mean is that different threads uses different sessions, but if the first thread init the session with the environment variable CUDA_VISIBLE_DEVICES, subsequent threads can not change the device by modifying CUDA_VISIBLE_DEVICES, they use the same device with the first thread.

"subsequent threads can not change device by modifying CUDA_VISIBLE_DEVICES", other threads try to change the device at session creation? Or session.Run()?

My previous description was inaccurate, I create theses sessions sequentially in a single thread, like:

for (int i...) {
    setenv(CUDA_VISIBLE_DEVICES, device_ids[i], 1);
    sessions[i] = ...; // create a new session with a different session_options all with device_id=0
    sessions[i].run(...); // warmup
}

Above code can use a device with device_id > 0, but all sessions use the same device with device_id=device_ids[0].

[chilo-ms]

Without setting CUDA_VISIBLE_DEVICES, the way I create sessions is equivalent to creating a session array, each session with a different device_id appended to a new session_option. The error message is TensorRT EP execution context enqueue failed, the same as #16274.

That's a bit strange, i can't repro from my side with devide_id specified by provider option:

for (int i...) {
    std::vector<const char*> keys{"device_id"};
    std::vector<const char*> values{device_id[i].c_str()}
    Ort::ThrowOnError(ortApi.UpdateTensorRTProviderOptions(tensorrt_options, keys.data(), values.data(), keys.size()));
    sessions[i] = ...; // create a new session with a different session_options with different device id
    sessions[i].run(...); // warmup
}

The error message is TensorRT EP execution context enqueue failed, the same as #16274.

Is that all the error message?
The issue has another error message: Error Code 1: Cuda Runtime (invalid resource handle)

Could you help enable verbose log by adding a line of code as below to see the TRT EP's full log?

Ort::Env env(ORT_LOGGING_LEVEL_VERBOSE, "test");

you will see the device_id configured by TRT EP provider options: (you should see different devide_id for different session)

...
2024-02-13 18:45:22.629147053 [V:onnxruntime:test, tensorrt_execution_provider.cc:1700 TensorrtExecutionProvider] [TensorRT EP] TensorRT provider options: device_id: 2, trt_max_partition_iterations: 1000, trt_min_subgraph_size: 1, trt_max_workspace_size: 1073741824, trt_fp16_enable: 0, trt_int8_enable: 0, trt_int8_calibration_cache_name: , int8_calibration_cache_available: 0, trt_int8_use_native_tensorrt_calibration_table: 0, trt_dla_enable: 0, trt_dla_core: 0, trt_dump_subgraphs: 0, trt_engine_cache_enable: 0, trt_cache_path: , trt_global_cache_path: , trt_engine_decryption_enable: 0, trt_engine_decryption_lib_path: , trt_force_sequential_engine_build: 0, trt_context_memory_sharing_enable: 0, trt_layer_norm_fp32_fallback: 0, trt_build_heuristics_enable: 0, trt_sparsity_enable: 0, trt_builder_optimization_level: 3, trt_auxiliary_streams: -1, trt_tactic_sources: , trt_profile_min_shapes: , trt_profile_max_shapes: , trt_profile_opt_shapes: , trt_cuda_graph_enable: 0, trt_dump_ep_context_model: 0, trt_ep_context_file_path: , trt_ep_context_embed_mode: 0, trt_cache_prefix:
...

[chilo-ms]

My previous description was inaccurate, I create theses sessions sequentially in a single thread, like:

for (int i...) {
    setenv(CUDA_VISIBLE_DEVICES, device_ids[i], 1);
    sessions[i] = ...; // create a new session with a different session_options all with device_id=0
    sessions[i].run(...); // warmup
}

Above code can use a device with device_id > 0, but all sessions use the same device with device_id=device_ids[0].

I can't repro from my side using CUDA_VISIBLE_DEVICES either

[chilo-ms]

Follow-up questions here:

1. How do you check that session using which devicde_id?
2. Can you create one session with device_id > 0 via TRT EP provider options and run the session successfully?
   Or the session can only be run successfully with device_id = 0

[jax11235]

Without setting CUDA_VISIBLE_DEVICES, the way I create sessions is equivalent to creating a session array, each session with a different device_id appended to a new session_option. The error message is TensorRT EP execution context enqueue failed, the same as #16274.

That's a bit strange, i can't repro from my side with devide_id specified by provider option:

for (int i...) {
    std::vector<const char*> keys{"device_id"};
    std::vector<const char*> values{device_id[i].c_str()}
    Ort::ThrowOnError(ortApi.UpdateTensorRTProviderOptions(tensorrt_options, keys.data(), values.data(), keys.size()));
    sessions[i] = ...; // create a new session with a different session_options with different device id
    sessions[i].run(...); // warmup
}

The error message is TensorRT EP execution context enqueue failed, the same as #16274.

Is that all the error message? The issue has another error message: Error Code 1: Cuda Runtime (invalid resource handle)

Could you help enable verbose log by adding a line of code as below to see the TRT EP's full log?

Ort::Env env(ORT_LOGGING_LEVEL_VERBOSE, "test");

you will see the device_id configured by TRT EP provider options: (you should see different devide_id for different session)

...
2024-02-13 18:45:22.629147053 [V:onnxruntime:test, tensorrt_execution_provider.cc:1700 TensorrtExecutionProvider] [TensorRT EP] TensorRT provider options: device_id: 2, trt_max_partition_iterations: 1000, trt_min_subgraph_size: 1, trt_max_workspace_size: 1073741824, trt_fp16_enable: 0, trt_int8_enable: 0, trt_int8_calibration_cache_name: , int8_calibration_cache_available: 0, trt_int8_use_native_tensorrt_calibration_table: 0, trt_dla_enable: 0, trt_dla_core: 0, trt_dump_subgraphs: 0, trt_engine_cache_enable: 0, trt_cache_path: , trt_global_cache_path: , trt_engine_decryption_enable: 0, trt_engine_decryption_lib_path: , trt_force_sequential_engine_build: 0, trt_context_memory_sharing_enable: 0, trt_layer_norm_fp32_fallback: 0, trt_build_heuristics_enable: 0, trt_sparsity_enable: 0, trt_builder_optimization_level: 3, trt_auxiliary_streams: -1, trt_tactic_sources: , trt_profile_min_shapes: , trt_profile_max_shapes: , trt_profile_opt_shapes: , trt_cuda_graph_enable: 0, trt_dump_ep_context_model: 0, trt_ep_context_file_path: , trt_ep_context_embed_mode: 0, trt_cache_prefix:
...

Thanks for your reply, I have switched to multiple processes version, it runs fine on multiple cuda devices.
Because you can't reproduce it, currently I think it may be other non-code issues, I will do more tests if have time.