[P2P] Remove collective backend from P2P, only use P2P native RDMA implementation

[zhongjiechen]

Description

Please include a summary of the changes and the related issue.

Fixes # (issue)

Type of Change

• Bug fix
• New feature
• Documentation update

How Has This Been Tested?

Include any tests here.

• Unit tests
• Integration tests
• Manual testing

Checklist

• My code follows the style guidelines, e.g. format.sh.
• I have run build_and_install.sh to verify compilation.
• I have removed redundant variables and comments.
• I have updated the documentation.
• I have added tests.

[YangZhou1997]

hi @zhongjie, we also have some lazy init functionality used by uccl_engine.h (from Nixl). Our IB and EFA would also want to support that.

[zhongjiechen]

hi @zhongjie, we also have some lazy init functionality used by uccl_engine.h (from Nixl). Our IB and EFA would also want to support that.

yeah! i have supported that with minor modification but need to run Nixl once to test

[praveingk]

@zhongjiechen Does the new NIC endpoint support initializing without gpu_idx for RDMA. This change was made to support NIXL, where the local_gpu_idx not passed in during initialization, but is discovered during memory registration, and then NIC is initialized accordingly.

[zhongjiechen]

Yeah! I deliver INVALID_GPU to its constructor here:

uccl/p2p/engine.cc

Lines 108 to 110 in f814432

	// Initialize the RDMA endpoint with lazy creation.
	ep_ = std::shared_ptr<NICEndpoint>(
	new NICEndpoint(INVALID_GPU, INVALID_RANK_ID, 0, false));

uccl/p2p/rdma/rdma_endpoint.h

Lines 24 to 35 in f814432

	if (gpu_index != INVALID_GPU) {
	std::vector<size_t> actual_device_ids;
	if (device_ids.size() == 0) {
	actual_device_ids =
	RdmaDeviceManager::instance().get_best_dev_idx(gpu_index);
	} else {
	actual_device_ids = device_ids;
	}
	initializeContexts(actual_device_ids);
	LOG(INFO) << "NICEndpoint initialized with " << contexts_.size()
	<< " context(s) for GPU " << gpu_index;
	}

And call initialize_engine here during memory registration:

uccl/p2p/engine.cc

Line 167 in f814432

unified::initialize_engine_by_dev(ep_, local_gpu_idx_, false);

With P2P native RDMA's initialize_engine_by_dev:

uccl/p2p/rdma/rdma_endpoint.h

Lines 383 to 396 in f814432

	bool initialize_engine_by_dev(int gpu_index, bool enable_p2p_listen) {
	(void)enable_p2p_listen;
	gpu_index_ = gpu_index;
	std::vector<size_t> device_ids =
	RdmaDeviceManager::instance().get_best_dev_idx(gpu_index_);
	initializeContexts(device_ids);
	LOG(INFO) << "NICEndpoint initialized with " << contexts_.size()
	<< " context(s) for GPU " << gpu_index_;
	return true;
	}

[praveingk]

Got it 👍🏽 Thanks for clarifying.
Also, should we simplify the directory structure? for e.g move rdma, tcp, tcp-x inside transport\?

[zhongjiechen]

Second you! It looks a bit messy right now... Let me reorganize them.

[YangZhou1997]

@zhongjiechen , quick thought: it would be great to get rid of the extra chunk in the engine.cc and engine.h, as I think they should be just done once in the underlying transport providers.