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

Improving build time by removing the gfx11xx and host code from rccl_float8.h (#1789)

Browse Source 
* removing extra build time by removing the gfx11xx arch from using hip_fp8

---------

Co-authored-by: Marzieh Berenjkoub <mberenjk@amd.com>
This commit is contained in:
mberenjk
2025-07-09 14:03:47 -05:00
committed by GitHub
parent 9c89573580
commit 697bee4ee8
3 changed files with 6 additions and 39 deletions
Download Patch File Download Diff File Expand all files Collapse all files
test/common/PtrUnion.cpp
+2 -2
View File
@@ -234,11 +234,11 @@ namespace RcclUnitTesting
case ncclUint32: U4[idx] *= scalarsPerRank.U4[rank]; break;
case ncclInt64: I8[idx] *= scalarsPerRank.I8[rank]; break;
case ncclUint64: U8[idx] *= scalarsPerRank.U8[rank]; break;
case ncclFloat8e4m3: F1[idx] = rccl_float8(F1[idx] * scalarsPerRank.F1[rank]); break;
case ncclFloat8e4m3: F1[idx] = rccl_float8((float)F1[idx] * (float)scalarsPerRank.F1[rank]); break;
case ncclFloat16: F2[idx] = __float2half(__half2float(F2[idx]) * __half2float(scalarsPerRank.F2[rank])); break;
case ncclFloat32: F4[idx] *= scalarsPerRank.F4[rank]; break;
case ncclFloat64: F8[idx] *= scalarsPerRank.F8[rank]; break;
case ncclFloat8e5m2: B1[idx] = rccl_bfloat8(B1[idx] * scalarsPerRank.B1[rank]); break;
case ncclFloat8e5m2: B1[idx] = rccl_bfloat8((float)B1[idx] * (float)scalarsPerRank.B1[rank]); break;
case ncclBfloat16: B2[idx] *= scalarsPerRank.B2[rank]; break;
default:
ERROR("Unsupported datatype\n");