Skipping AllReduce test on more than 8 ranks for FP8 type on Hyabusa (#1598)

* Skipping AllReduce FP8 test on 9 to 16 ranks (gfx90a) as it's using Tree algorithm not RING

---------

Co-authored-by: Marzieh Berenjkoub <mberenjk@amd.com>

[ROCm/rccl commit: 5f691aaf65]

projects/rccl/test/common/EnvVars.cpp

@@ -202,6 +202,8 @@ namespace RcclUnitTesting
     getArchInfo(&isGfx94, "gfx94");
     isGfx12 = false;
     getArchInfo(&isGfx12, "gfx12");
+    isGfx90 = false;
+    getArchInfo(&isGfx90, "gfx90");
 
     showNames      = GetEnvVar("UT_SHOW_NAMES"  , 1);
     minGpus        = GetEnvVar("UT_MIN_GPUS"    , 1);

projects/rccl/test/common/EnvVars.hpp

@@ -32,6 +32,7 @@ namespace RcclUnitTesting
     bool useMultithreading; // Multi-thread single-process ranks   [UT_MULTITHREAD]
     bool isGfx94;        // Detects if architecture is gfx94
     bool isGfx12;        // Detects if architecture is gfx12
+    bool isGfx90;        // Detects if architecture is gfx90
 
     // Constructor that parses and collects environment variables
     EnvVars();

projects/rccl/test/common/TestBed.cpp

@@ -681,6 +681,14 @@ namespace RcclUnitTesting
 
       for (int ftIdx = 0; ftIdx < funcTypes.size()      && isCorrect; ++ftIdx)
       for (int dtIdx = 0; dtIdx < dataTypes.size()      && isCorrect; ++dtIdx)
+      {
+      //Skipping AllReduce FP8 test on 9 to 16 ranks (gfx90a).
+      if(ev.isGfx90 && numRanks > 8 && funcTypes[ftIdx] == ncclCollAllReduce
+                    && (dataTypes[dtIdx] == ncclFp8E4M3
+                    || dataTypes[dtIdx] == ncclFp8E5M2))
+      {
+            continue;
+      }
       for (int rdIdx = 0; rdIdx < redOps.size()         && isCorrect; ++rdIdx)
       for (int rtIdx = 0; rtIdx < roots.size()          && isCorrect; ++rtIdx)
       for (int ipIdx = 0; ipIdx < inPlaceList.size()    && isCorrect; ++ipIdx)
@@ -764,6 +772,7 @@ namespace RcclUnitTesting
         }
         this->DeallocateMem();
       }
+    }
       this->DestroyComms();
     }
   }