Fix gfx950 gating conditions to match ROCm 7.0.2 (#2003)

此提交包含在:
Nilesh M Negi
2025-10-29 23:27:04 -05:00
提交者 GitHub
父節點 12f51ba8bf
當前提交 8444b3c6e9
共有 2 個檔案被更改,包括 5 行新增5 行删除
+1 -1
查看文件
@@ -1410,7 +1410,7 @@ static ncclResult_t initTransportsRank(struct ncclComm* comm, struct ncclComm* p
comm -> gfx9CheapFenceOff = 0;
}
else if(IsArchMatch(comm->topo->nodes[GPU].nodes[idx].gpu.gcn, "gfx950")){
comm -> gfx9CheapFenceOff = ROCM_VERSION < 70200 && nNodes > 1; // Enable for single node only prior to ROCm 7.0.2
comm -> gfx9CheapFenceOff = ROCM_VERSION < 70002 && nNodes > 1; // Enable for single node only prior to ROCm 7.0.2
}
}
INFO(NCCL_INIT, "GFX9 cheap fence is %s", comm -> gfx9CheapFenceOff ? "OFF" : "ON");