Skip to content

feat(topology): add GPU entries in topology discovery #673

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
wants to merge 2 commits into
base: main
Choose a base branch
from
Open

feat(topology): add GPU entries in topology discovery #673

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
3e052c0
feat(topology): add GPU entries alongside CUDA entries in topo discovery
staryxchen Jul 25, 2025
650ad76
feat(transfer_engine): enhance topology logging
staryxchen Jul 27, 2025
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
5 changes: 4 additions & 1 deletion mooncake-transfer-engine/src/topology.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -203,8 +203,11 @@ static std::vector<TopologyEntry> discoverCudaTopology(
avail_hca.push_back(hca.name);
}
}
topology.push_back(TopologyEntry{.name = "cuda:" + std::to_string(i),
.preferred_hca = preferred_hca,
.avail_hca = avail_hca});
topology.push_back(
TopologyEntry{.name = "cuda:" + std::to_string(i),
TopologyEntry{.name = "gpu:" + std::to_string(i),
Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Just wonder what the difference is between cuda:x and gpu:x, it seems the same thing in torch.

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Yes, there is no difference for the Torch API. However, now, in Mooncake, we select the device to use based on the segment buffer name, which is a string comparison. Therefore, if the user registers the memory and declares it as “gpu:”, Mooncake will fall back to using the device corresponding to kWildcardLocation because it cannot find a matching string.

Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

nope, I do think people should use cuda: instead of gpu:, see document here: https://github.com/kvcache-ai/Mooncake/blob/main/doc/zh/transfer-engine.md#transferengineregisterlocalmemory

Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Changing the name may lead compatibility problem after merging it.

Copy link
Collaborator

@alogfans alogfans Jul 31, 2025 •
edited
Loading

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

In other words, you can check whether there is any hard-coded "cuda:" in elsewhere.

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

In other words, you can check whether there is any hard-coded "cuda:" in elsewhere.

mooncake internal (location and topology) uniformly uses cuda:xx.

.preferred_hca = std::move(preferred_hca),
.avail_hca = std::move(avail_hca)});
}
Expand Down
9 changes: 8 additions & 1 deletion mooncake-transfer-engine/src/transfer_engine.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -20,6 +20,7 @@
#include <fstream>
#include <string>

#include "config.h"
#include "transfer_metadata_plugin.h"
#include "transport/transport.h"

Expand Down Expand Up @@ -150,8 +151,11 @@ int TransferEngine::init(const std::string &metadata_conn_string,
} else {
local_topology_->discover(filter_);
}
LOG(INFO) << "Topology discovery complete. Found "
LOG(INFO) << "Auto-discovering topology complete: Found "
<< local_topology_->getHcaList().size() << " HCAs.";
if (globalConfig().trace) {
LOG(INFO) << "Topology:\n" << local_topology_->toString();
}

#ifdef USE_MNNVL
if (local_topology_->getHcaList().size() > 0 &&
Expand Down Expand Up @@ -215,6 +219,9 @@ Transport *TransferEngine::installTransport(const std::string &proto,
LOG(ERROR) << "Failed to parse NIC priority matrix";
return nullptr;
}
if (globalConfig().trace) {
LOG(INFO) << "Install transport " << proto << " with custom topology:\n" << local_topology_->toString();
}
}

transport = multi_transports_->installTransport(proto, local_topology_);
Expand Down
Loading