[TRT RTX EP] Implement GetEPContextNodes()

[thevishalagarwal]

Implements GetEPContextNodes()

[thevishalagarwal]

cc @gedoensmax @ankan-ban @ishwar-raut1

@chilo-ms @jywu-msft to review

[HectorSVC]

/azp run Linux QNN CI Pipeline,Win_TRT_Minimal_CUDA_Test_CI,Windows ARM64 QNN CI Pipeline,Windows GPU Doc Gen CI Pipeline,Windows x64 QNN CI Pipeline

[azure-pipelines]

Azure Pipelines successfully started running 5 pipeline(s).

[HectorSVC]

ep_context_models_

The model instance here is just help to hold the EPContext nodes. Keep one instance is enough, why need a list of models?

[HectorSVC]

auto model_build = graph_viewer.CreateModel(*logger);

Seems wired to me. Why create a model from a existing graph_viewer? how the lifecycle is controlled with the existing graph_viewer? ep_context_model instance need to be valid as long as the EP instance is alive.
All EP need to do with GetEpContextNodes() is, the EP creates the model instance to hold all EPContext nodes for the graph partitioner to query. EP only need to add all EPContext nodes into that model instance.

onnxruntime/onnxruntime/core/providers/qnn/qnn_execution_provider.cc

Lines 1179 to 1191 in 9705b17

	qnn_ep_context_model_ = Factory<Model>::Create(std::string{"qnn_ep_context_model"}, false, logger);
	ORT_RETURN_IF_ERROR(qnn::CreateEPContextNodes(qnn_ep_context_model_.get(),
	context_buffer.get(),
	buffer_size,
	qnn_backend_manager_->GetSdkVersion(),
	fused_nodes_and_graphs,
	qnn_models_,
	context_model_path,
	qnn_context_embed_mode_,
	max_spill_fill_buffer_size,
	logger,
	share_ep_contexts_,
	stop_share_ep_contexts_));

[chilo-ms]

The name of the CreateCtxModel() might be a bit confusing.
It has some historic reasons

• At the beginning of TRT EP implementing EP Context feature, it didn't follow the rule to implement GetEpContextNodes()
• it only supported a model only containing a single EP Context node, meaning the whole graph can be run by TRT. There is no partitioning.
• It didn't implement EP API's GetEpContextNodes(), instead it directly creates the ONNX model, so that's why this function has this name to really create a onnx model and dump to file.

[chilo-ms]

Right now, since RTX EP and TRT EP are going to follow EP Context implementation rule to implement GetEpContextNodes(), we could rename it to avoid confusion.

[gedoensmax]

Sharing some source between TRT and TRT RTX seems like a good idea but will be kind of awkward. Any opinion on this ? Otherwise let's duplicate changes and source for now.

[HectorSVC]

There was a fix for the Web CI pipeline, please merge the code from latest main branch.

[chilo-ms]

When running with TRT RTX EP, will it also handle the case where model contains contrib ops that will be fallback to run on CUDA EP or CPU? I assume it will.

If that's the case, following change need to be added to this PR as well.

• In GetCapability, it should be able the correctly parse the EP Context model that contains some EP Context nodes and some onnx nodes.

[gedoensmax]

If we will be filling the source section on an EP Context node will the partitioner automatically force a node onto an EP ? I think if there are EP Context nodes there is no longer a real partitioning needed 🤔