Fixing the issue where the A2 notify_dispatch operator gets stuck on cann8.3 (#245)

Author: oagniqgnat

csrc/deepep/ops2/op_host/notify_dispatch_a2.cpp

@@ -92,25 +92,20 @@ class NotifyDispatchA2 : public OpDef
         this->Attr("local_rank_size").Int();
         this->Attr("local_rank_id").Int();
 
-        OpAICoreConfig aicore_config_base;
-        aicore_config_base.DynamicCompileStaticFlag(true)
+        OpAICoreConfig aicore_config_A2;
+        aicore_config_A2.DynamicCompileStaticFlag(true)
             .DynamicFormatFlag(true)
             .DynamicRankSupportFlag(true)
             .DynamicShapeSupportFlag(true)
             .NeedCheckSupportFlag(false)
             .PrecisionReduceFlag(true)
             .ExtendCfgInfo("aclnnSupport.value", "support_aclnn")
+            .ExtendCfgInfo("prebuildPattern.value", "Opaque")
+            .ExtendCfgInfo("jitCompile.flag", "static_false")
             .ExtendCfgInfo("multiKernelSupportDynamicGraph.value", "multi_kernel");
 
-        OpAICoreConfig aicore_config_A2 = aicore_config_base;
-        aicore_config_A2.ExtendCfgInfo("jitCompile.flag", "static_false");
-
-        OpAICoreConfig aicore_config = aicore_config_base;
-        aicore_config.ExtendCfgInfo("jitCompile.flag", "static_true");
-
-        this->AICore().AddConfig("ascend910_93", aicore_config);
         this->AICore().AddConfig("ascend910b", aicore_config_A2);
-        this->MC2().HcclGroup("comm_group");
+        this->MC2().HcclGroup({"comm_group"});
     }
 };
 

csrc/deepep/ops2/op_host/notify_dispatch_tiling_a2.cc

@@ -186,8 +186,8 @@ static void SetHcommCfg(const gert::TilingContext *context, NotifyDispatchA2Tili
 {
     const char *nodeName = context->GetNodeName();
     OP_LOGD(nodeName, "NotifyDispatchA2 commGroup = %s", commGroup.c_str());
-    uint32_t opType1 = OP_TYPE_ALL_TO_ALL;
-    std::string algConfigAllToAllStr = "AlltoAll=level0:fullmesh;level1:pairwise";
+    uint32_t opType1 = 18;  // batch write=18,
+    std::string algConfigAllToAllStr = "BatchWrite=level1:hierarchy";
 
     AscendC::Mc2CcTilingConfig mc2CcTilingConfig(commGroup, opType1, algConfigAllToAllStr);
     mc2CcTilingConfig.GetTiling(tiling->mc2InitTiling);
@@ -442,6 +442,7 @@ static ge::graphStatus NotifyDispatchA2TilingFuncImpl(gert::TilingContext *conte
     if (socVersion == "Ascend910B") {
         tilingKey = tilingKey + TILING_KEY_A2_TYPE;
     }
+    OP_LOGD(nodeName, "tilingKey is %lu", tilingKey);
     context->SetTilingKey(tilingKey);
 
     auto ascendcPlatform = platform_ascendc::PlatformAscendC(context->GetPlatformInfo());