Revert "Separate the buffers used by D/C and notify_dispatch to avoid conflic…" (#89)

This reverts commit 29aec86.

Author: iforgetmyname

csrc/deepep/ops/op_kernel/cam_moe_combine_normal.h

@@ -5,7 +5,6 @@
 #include "kernel_tiling/kernel_tiling.h"
 #include "moe_distribute_base.h"
 #include "cam_moe_combine_normal_tiling.h"
-#include "comm_args.h"
 
 namespace CamMoeCombineNormalImpl {
 constexpr uint32_t RANK_ID_OFFSET_IN_SRC_INFO = 0U;
@@ -64,7 +63,7 @@ class CamMoeCombineNormal
         } else {
             bufferAddr = (GM_ADDR)((HcclRankRelationResV2 *)epWinContext_->remoteRes[rankId].nextDevicePtr)->windowsIn;
         }
-        return (GM_ADDR)(bufferAddr + winDataSizeOffset_ + Moe::NOTIFY_DISPATCH_BUFF_OFFSET);
+        return (GM_ADDR)(bufferAddr + winDataSizeOffset_);
     }
 
     __aicore__ GM_ADDR GetBufferAddrByRankId(const int32_t rankId)

csrc/deepep/ops/op_kernel/cam_moe_dispatch_normal.h

@@ -5,7 +5,6 @@
 #include "kernel_tiling/kernel_tiling.h"
 #include "moe_distribute_base.h"
 #include "cam_moe_dispatch_normal_tiling.h"
-#include "comm_args.h"
 
 namespace CamMoeDispatchNormalImpl {
 constexpr uint8_t BUFFER_NUM = 2;
@@ -60,11 +59,10 @@ class CamMoeDispatchNormal
     {
         uint32_t curRankId = ((ctxIdx == COMM_EP_IDX) ? epRankId : tpRankId);
         if (curRankId == rankId) {
-            return (GM_ADDR)(winContext_[ctxIdx]->localWindowsIn) + winDataSizeOffset + COMBINE_STATE_WIN_OFFSET +
-                   Moe::NOTIFY_DISPATCH_BUFF_OFFSET;
+            return (GM_ADDR)(winContext_[ctxIdx]->localWindowsIn) + winDataSizeOffset + COMBINE_STATE_WIN_OFFSET;
         }
         return (GM_ADDR)(((HcclRankRelationResV2 *)(winContext_[ctxIdx]->remoteRes[rankId].nextDevicePtr))->windowsIn) +
-               winDataSizeOffset + COMBINE_STATE_WIN_OFFSET + Moe::NOTIFY_DISPATCH_BUFF_OFFSET;
+               winDataSizeOffset + COMBINE_STATE_WIN_OFFSET;
     }
 
     __aicore__ inline GM_ADDR GetWindStateAddrByRankId(uint8_t ctxIdx, const int32_t rankId)

csrc/deepep/ops/op_kernel/comm_args.h

@@ -8,7 +8,6 @@
 namespace Moe {
 constexpr int CAM_MAX_RANK_SIZE = 384;  // Maximum number of NPU cards supported by the communication library
 
-constexpr int64_t NOTIFY_DISPATCH_BUFF_OFFSET = 204UL * 1024UL * 1024UL;
 constexpr int64_t IPC_BUFF_MAX_SIZE = 100 * 1024 * 1024;
 constexpr int64_t IPC_DATA_OFFSET = 2 * 1024 * 1024;  // First 2MB as flag, then 100MB as data storage
 constexpr int64_t PING_PONG_SIZE = 2;