Skip to content

[None][feat] add flag for EPLB to force using GDRCopy #8650

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

Merged
merged 1 commit into from
Oct 29, 2025
Merged

[None][feat] add flag for EPLB to force using GDRCopy #8650

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
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: 5 additions & 0 deletions cpp/tensorrt_llm/common/envUtils.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -495,4 +495,9 @@ int getEnvMoeA2ACombineBlockSize()
return kBlock;
}

bool getEnvEplbForceGdrcopy()
{
return getBoolEnv("TRTLLM_EPLB_FORCE_GDRCOPY");
}

} // namespace tensorrt_llm::common
2 changes: 2 additions & 0 deletions cpp/tensorrt_llm/common/envUtils.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -147,4 +147,6 @@ int getEnvMoeA2ACombineBlockSize();

bool getEnvKVCacheTransferAllBlocksForWindow();

bool getEnvEplbForceGdrcopy();

} // namespace tensorrt_llm::common
15 changes: 13 additions & 2 deletions cpp/tensorrt_llm/runtime/moeLoadBalancer/hostAccessibleDeviceAllocator.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -27,6 +27,7 @@
#include "topologyDetector.h"

#include "tensorrt_llm/common/cudaUtils.h"
#include "tensorrt_llm/common/envUtils.h"
#include "tensorrt_llm/common/logger.h"

namespace tensorrt_llm::runtime
Expand Down Expand Up @@ -169,7 +170,8 @@ bool HostAccessibleDeviceAllocator::mAllowManagedFallback = false;

bool HostAccessibleDeviceAllocator::isSupported()
{
if (TopologyDetector::getInstance().getCurrentGpuMemoryNumaId() >= 0)
if (!tensorrt_llm::common::getEnvEplbForceGdrcopy()
&& TopologyDetector::getInstance().getCurrentGpuMemoryNumaId() >= 0)
{
// we are on systems that GPU memory is also a NUMA node.
return true;
Expand All @@ -195,7 +197,16 @@ void HostAccessibleDeviceAllocator::init()
}

TLLM_CUDA_CHECK(cudaGetDevice(&mDevId));
mGpuMemNumaId = TopologyDetector::getInstance().getCurrentGpuMemoryNumaId();
if (tensorrt_llm::common::getEnvEplbForceGdrcopy())
{
mGpuMemNumaId = -1;
TLLM_LOG_INFO("Force using GDRCopy for EPLB, ignore NUMA node for GPU memory.");
}
else
{
mGpuMemNumaId = TopologyDetector::getInstance().getCurrentGpuMemoryNumaId();
}

if (mGpuMemNumaId < 0)
{
// We only use GDRCopy when there is no NUMA node for GPU memory.
Expand Down