Revert "Fix the memory verification issue within intranode dispatch (#83)" (#88)

This reverts commit 50903a9.

Author: iforgetmyname

csrc/deepep/ops/op_host/cam_moe_combine_normal_tiling.cc

@@ -19,11 +19,9 @@
 #include "graph/utils/type_utils.h"
 #include "register/op_def_registry.h"
 #include "../op_kernel/cam_moe_combine_normal_tiling.h"
-#include "tiling_args.h"
 
 using namespace AscendC;
 using namespace ge;
-using namespace Moe;
 
 namespace {
 class Mc2TilingUtils
@@ -85,6 +83,7 @@ constexpr uint64_t MB_SIZE = 1024UL * 1024UL;
 constexpr uint64_t TRIPLE = 3;
 constexpr uint64_t WIN_ADDR_ALIGN = 512UL;
 constexpr uint64_t SCALE_RECV_IDX_BUFFER = 44UL;  // scale32B + 3*4 src info
+constexpr uint64_t COMBINE_STATE_WIN_OFFSET = 3U * 1024UL * 1024UL;
 constexpr uint64_t DOUBLE_DATA_BUFFER = 2UL;
 constexpr uint64_t MAX_OUT_DTYPE_SIZE = 2UL;
 constexpr uint64_t UB_ALIGN = 32UL;
@@ -515,17 +514,20 @@ static ge::graphStatus CamMoeCombineNormalA3TilingFuncImpl(gert::TilingContext *
     uint64_t maxBs = static_cast<uint64_t>(tilingData->camMoeCombineNormalInfo.globalBs) / epWorldSize;
     // combine数据区 token首地址对齐512
     uint64_t tokenNeedSizeCombine = ((h * MAX_OUT_DTYPE_SIZE + WIN_ADDR_ALIGN - 1UL) / WIN_ADDR_ALIGN) * WIN_ADDR_ALIGN;
+    // dispatch数据区 token首对齐512，有效token长度h_align_32b + scale(32b) + 三元组(3*4b)
+    uint64_t tokenActualLen = ((h * MAX_OUT_DTYPE_SIZE + UB_ALIGN - 1UL) / UB_ALIGN) * UB_ALIGN + SCALE_RECV_IDX_BUFFER;
+    uint64_t tokenNeedSizeDispatch = ((tokenActualLen + WIN_ADDR_ALIGN - 1UL) / WIN_ADDR_ALIGN) * WIN_ADDR_ALIGN;
     uint64_t actualSize =
-        (maxBs * k * tokenNeedSizeCombine + COMBINE_STATE_WIN_OFFSET + NOTIFY_DISPATCH_WIN_OFFSET) * DOUBLE_DATA_BUFFER;
+        (maxBs * k * (tokenNeedSizeCombine + tokenNeedSizeDispatch) + COMBINE_STATE_WIN_OFFSET) * DOUBLE_DATA_BUFFER;
     OP_TILING_CHECK(
         (actualSize > maxWindowSize),
         OP_LOGE(nodeName,
                 "HCCL_BUFFSIZE is too SMALL, maxBs = %lu, h = %lu, epWorldSize = %lu, localMoeExpertNum = %u,"
-                " tokenNeedSizeCombine = %lu, k = %lu, NEEDED_HCCL_BUFFSIZE("
-                "((maxBs * k * tokenNeedSizeCombine)) + 3MB + 204MB) * 2) = %luMB, "
+                " tokenNeedSizeDispatch = %lu, tokenNeedSizeCombine = %lu, k = %lu, NEEDED_HCCL_BUFFSIZE("
+                "((maxBs * tokenNeedSizeDispatch) + (maxBs * tokenNeedSizeCombine * k) + 3MB) * 2) = %luMB, "
                 "HCCL_BUFFSIZE=%luMB.",
-                maxBs, h, epWorldSize, localMoeExpertNum, tokenNeedSizeCombine, k, actualSize / MB_SIZE + 1UL,
-                maxWindowSize / MB_SIZE),
+                maxBs, h, epWorldSize, localMoeExpertNum, tokenNeedSizeDispatch, tokenNeedSizeCombine, k,
+                actualSize / MB_SIZE + 1UL, maxWindowSize / MB_SIZE),
         return ge::GRAPH_FAILED);
     tilingData->camMoeCombineNormalInfo.totalWinSize = maxWindowSize;
 

csrc/deepep/ops/op_host/cam_moe_dispatch_normal_tiling.cc

@@ -18,12 +18,9 @@
 #include "graph/utils/type_utils.h"
 #include "register/op_def_registry.h"
 #include "../op_kernel/cam_moe_dispatch_normal_tiling.h"
-#include "tiling_args.h"
 
 using namespace AscendC;
 using namespace ge;
-using namespace Moe;
-
 namespace {
 class Mc2TilingUtils
 {
@@ -94,7 +91,6 @@ constexpr uint32_t WORKSPACE_ELEMENT_OFFSET = 512;
 constexpr int64_t H_MIN = 1024;
 constexpr int64_t H_MAX = 7168;
 constexpr uint64_t MB_SIZE = 1024UL * 1024UL;
-
 constexpr uint64_t TRIPLE = 3;
 constexpr uint64_t WIN_ADDR_ALIGN = 512UL;
 constexpr uint64_t SCALE_EXPAND_IDX_BUFFER = 44UL;  // scale32B + 3*4expandIdx
@@ -543,18 +539,15 @@ static ge::graphStatus CamMoeDispatchNormalA3TilingFuncImpl(gert::TilingContext
     uint64_t tokenActualLen =
         ((h * MAX_OUT_DTYPE_SIZE + UB_ALIGN - 1UL) / UB_ALIGN) * UB_ALIGN + SCALE_EXPAND_IDX_BUFFER;
     uint64_t tokenNeedSizeDispatch = ((tokenActualLen + WIN_ADDR_ALIGN - 1UL) / WIN_ADDR_ALIGN) * WIN_ADDR_ALIGN;
-    uint64_t tokenNeedSizeCombine = ((h * MAX_OUT_DTYPE_SIZE + WIN_ADDR_ALIGN - 1UL) / WIN_ADDR_ALIGN) * WIN_ADDR_ALIGN;
     // 未考虑双流时大小
-    uint64_t actualSize = (maxBs * k * (tokenNeedSizeCombine + tokenNeedSizeDispatch) + COMBINE_STATE_WIN_OFFSET +
-                           NOTIFY_DISPATCH_WIN_OFFSET) *
-                          DOUBLE_DATA_BUFFER;
+    uint64_t actualSize = maxBs * k * tokenNeedSizeDispatch * DOUBLE_DATA_BUFFER;
     OP_TILING_CHECK((actualSize > maxWindowSize),
                     OP_LOGE(nodeName,
                             "HCCL_BUFFSIZE is too SMALL, maxBs = %lu, h = %lu, epWorldSize = %lu,"
-                            " localMoeExpertNum = %u, tokenNeedSizeDispatch = %lu, tokenNeedSizeCombine = %lu,"
-                            " k = %lu, NEEDED_HCCL_BUFFSIZE((maxBs * k * (tokenNeedSizeDispatch"
-                            " + tokenNeedSizeCombine) + 3MB + 204MB) * 2) = %luMB, HCCL_BUFFSIZE=%luMB.",
-                            maxBs, h, epWorldSize, localMoeExpertNum, tokenNeedSizeDispatch, tokenNeedSizeCombine, k,
+                            " localMoeExpertNum = %u, tokenNeedSizeDispatch = %lu,"
+                            " k = %lu, NEEDED_HCCL_BUFFSIZE(maxBs * k * tokenNeedSizeDispatch) = %luMB,"
+                            " HCCL_BUFFSIZE=%luMB.",
+                            maxBs, h, epWorldSize, localMoeExpertNum, tokenNeedSizeDispatch, k,
                             actualSize / MB_SIZE + 1UL, maxWindowSize / MB_SIZE),
                     return ge::GRAPH_FAILED);
     tilingData->camMoeDispatchNormalInfo.totalWinSize = maxWindowSize;