Share memory in single session without user specification

[okdshin]

Description

Removed is_shared_initializer from the decision to cache pre-packed weights. This allows memory to be shared within a single session even if the user does not specify it, thus reducing memory usage.

Motivation and Context

In the current implementation, memory is not shared even if the model shares parameters among multiple layers. There is a mechanism to share memory specified by the user, but the memory usage is large. By modifying the code slightly, we found that memory can be shared within a single session even if the user does not specify it, thus reducing memory usage.

In the current implementation, the decision to share or not to share memory is based on whether the memory is specified by the user or not.

onnxruntime/onnxruntime/core/framework/session_state.cc

Lines 426 to 430 in 5930e7e

	auto iter = initializers_to_share_map.find(input_name);
	bool is_shared_initializer = (iter != initializers_to_share_map.end());
	// Caching pre-packed weights is limited to shared initializers associated with the CPU EP for now
	if (is_shared_initializer && should_cache_prepacked_weights_for_shared_initializers &&

Even if this conditional branch is removed, memory that should not be shared will not be shared because there will be a conditional branch later in the Murmur hash to check whether the memory contents are the same or not.

onnxruntime/onnxruntime/core/framework/session_state.cc

Lines 460 to 465 in 5930e7e

	const std::string& prepacked_weights_container_key = GenerateKeyForPrepackedWeightsMap(op_type,
	weights_to_be_filled_in);
	bool container_contains_packed_weight = prepacked_weights_container_->HasWeight(prepacked_weights_container_key);
	if (container_contains_packed_weight) {

[take-cheeze]

I think we should work from don't break anything. So how about adding "share weights aggressively" option to SessionOptions. If it won't break anything in onnxruntime CI, maybe it could be true by default