AI/rocm-systems
2
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Update direct AG and single node LL threshold (#1944)

Browse Source 
* update AG direct and single node LL threshold

* update thresholds based on MI350 expeirmental results

* disable using LL for direct AG

* enable direct AG for lower GPU counts

* direct AG single node tuning

* fix in-place buffer allocation for AG unit test

* whitespace fix

* gate direct AG for gfx950 and gfx942

---------

Co-authored-by: Nusrat Islam <nusislam@nova-login-gtu2.prov.gtu.zts.cpe.ice.amd.com>
This commit is contained in:
Nusrat Islam
2025-10-09 10:48:50 -05:00
committed by GitHub
parent 00a42c80f3
commit d22a39e954
3 changed files with 29 additions and 12 deletions
Download Patch File Download Diff File Expand all files Collapse all files
test/common/CollectiveArgs.cpp
+1 -1
View File
@@ -82,7 +82,7 @@ namespace RcclUnitTesting
CHECK_CALL(this->inputGpu.AllocateGpuMem(this->numInputBytesAllocated, useManagedMem, userRegistered));
this->outputGpu.Attach(this->inputGpu.U1 + (this->globalRank * this->numOutputBytesAllocated));
}
else if (this->funcType == ncclCollGather)
else if (this->funcType == ncclCollGather || this->funcType == ncclCollAllGather)
{
CHECK_CALL(this->outputGpu.AllocateGpuMem(this->numOutputBytesAllocated, useManagedMem, userRegistered));
this->inputGpu.Attach(this->outputGpu.U1 + (this->globalRank * this->numInputBytesAllocated));