[Kernel] add custom moe ops for prefill (#4194)
### What this PR does / why we need it?
1.Add the implementation of normal Aclnn operators: MoeCombineNormal,
MoeDispatchNormal, NotifyDispatch，and DispatchLayout.
- MoeCombineNormal: Implements the combine logic within MoE operations.
- MoeDispatchNormal: Implements the dispatch logic within MoE
operations.
- NotifyDispatch: Exchanges topk_idx information among different ranks
to calculate the device memory required for the dispatch stage.
- DispatchLayout: Used to calculate information related to the device
memory layout for the dispatch stage.
2.Provide PyTorch interfaces for normal operators—get_dispatch_layout,
dispatch_prefill, and combine_prefill—to be used for MoE communication
during the prefill stage in vLLM.
- get_dispatch_layout: Calculates information related to the device
memory layout for the dispatch operator, and is called before
dispatch_prefill.
- dispatch_prefill: Initiates the dispatch operation.
- combine_prefill: Initiates the combine operation.
### Does this PR introduce _any_ user-facing change?
No
### How was this patch tested?
The functionality has already been validated using the local Qwen model.
Test cases will be added after support for multi-NPU use cases in the CI
pipeline is finalized.
- vLLM version: v0.12.0
- vLLM main:
ad32e3e19c
Signed-off-by: shiro-zzzz <zhangdianhao@huawei.com>
This commit is contained in:
shiro-zzzz
56
csrc/moe_dispatch_normal/op_kernel/moe_dispatch_normal.cpp
Normal file
56
csrc/moe_dispatch_normal/op_kernel/moe_dispatch_normal.cpp
Normal file
@@ -0,0 +1,56 @@
|
||||
#include "kernel_operator.h"
|
||||
#include "moe_dispatch_normal_tiling.h"
|
||||
#include "moe_dispatch_normal.h"
|
||||
|
||||
using namespace AscendC;
|
||||
using namespace MoeDispatchNormalImpl;
|
||||
|
||||
#define TILINGKEY_NO_QUANT 10000
|
||||
#define TILINGKEY_QUANT 10002
|
||||
|
||||
extern "C" __global__ __aicore__ void moe_dispatch_normal(GM_ADDR x, GM_ADDR expertIds, GM_ADDR send_offset,
|
||||
GM_ADDR send_token_idx, GM_ADDR recv_offset, GM_ADDR recv_count, GM_ADDR expandXOut, GM_ADDR dynamicScalesOut,
|
||||
GM_ADDR assist_info_for_combine, GM_ADDR workspaceGM, GM_ADDR tilingGM)
|
||||
{
|
||||
REGISTER_TILING_DEFAULT(MoeDispatchNormalTilingData);
|
||||
TPipe pipe;
|
||||
#if (ORIG_DTYPE_RECV_X == DT_BF16 || ORIG_DTYPE_RECV_X == DT_FLOAT16)
|
||||
if (TILING_KEY_IS(TILINGKEY_NO_QUANT)) {
|
||||
GET_TILING_DATA_WITH_STRUCT(MoeDispatchNormalTilingData, tilingData, tilingGM);
|
||||
MoeDispatchNormal<DTYPE_X, DTYPE_RECV_X, false, false, false> op;
|
||||
op.Init(x,
|
||||
expertIds,
|
||||
send_offset,
|
||||
send_token_idx,
|
||||
recv_offset,
|
||||
recv_count,
|
||||
expandXOut,
|
||||
dynamicScalesOut,
|
||||
assist_info_for_combine,
|
||||
workspaceGM,
|
||||
&pipe,
|
||||
&tilingData);
|
||||
op.Process();
|
||||
return;
|
||||
}
|
||||
#elif (ORIG_DTYPE_RECV_X == DT_INT8)
|
||||
if (TILING_KEY_IS(TILINGKEY_QUANT)) {
|
||||
GET_TILING_DATA_WITH_STRUCT(MoeDispatchNormalTilingData, tilingData, tilingGM);
|
||||
MoeDispatchNormal<DTYPE_X, DTYPE_RECV_X, true, false, false> op;
|
||||
op.Init(x,
|
||||
expertIds,
|
||||
send_offset,
|
||||
send_token_idx,
|
||||
recv_offset,
|
||||
recv_count,
|
||||
expandXOut,
|
||||
dynamicScalesOut,
|
||||
assist_info_for_combine,
|
||||
workspaceGM,
|
||||
&pipe,
|
||||
&tilingData);
|
||||
op.Process();
|
||||
return;
|
||||
}
|
||||
#endif
|
||||
}
|
||||
540
csrc/moe_dispatch_normal/op_kernel/moe_dispatch_normal.h
Normal file
540
csrc/moe_dispatch_normal/op_kernel/moe_dispatch_normal.h
Normal file
@@ -0,0 +1,540 @@
|
||||
#ifndef MOE_DISPATCH_NORMAL_H
|
||||
#define MOE_DISPATCH_NORMAL_H
|
||||
|
||||
#include "kernel_operator.h"
|
||||
#include "kernel_tiling/kernel_tiling.h"
|
||||
#include "../common/moe_distribute_base.h"
|
||||
#include "moe_dispatch_normal_tiling.h"
|
||||
|
||||
namespace MoeDispatchNormalImpl {
|
||||
constexpr uint8_t BUFFER_NUM = 2;
|
||||
constexpr uint32_t STATE_OFFSET = 32U;
|
||||
constexpr uint32_t UB_ALIGN = 32U;
|
||||
constexpr uint8_t COMM_NUM = 2;
|
||||
constexpr uint8_t COMM_EP_IDX = 0;
|
||||
constexpr uint8_t COMM_TP_IDX = 1;
|
||||
|
||||
constexpr uint64_t WIN_STATE_OFFSET = 500UL * 1024UL;
|
||||
constexpr uint64_t STATE_WIN_OFFSET = 950UL * 1024UL;
|
||||
constexpr uint64_t WIN_ADDR_ALIGN = 512UL;
|
||||
constexpr uint32_t EXPAND_IDX_INFO = 3U;
|
||||
constexpr uint64_t COMBINE_STATE_WIN_OFFSET = 3UL * 1024UL * 1024UL;
|
||||
|
||||
template <AscendC::HardEvent event>
|
||||
__aicore__ inline void SyncFunc()
|
||||
{
|
||||
int32_t eventID = static_cast<int32_t>(GetTPipePtr()->FetchEventID(event));
|
||||
AscendC::SetFlag<event>(eventID);
|
||||
AscendC::WaitFlag<event>(eventID);
|
||||
}
|
||||
|
||||
#define CamTypeClass \
|
||||
typename XType, typename ExpandXOutType, bool DynamicQuant, bool IsSmoothScaleExist, bool IsShareExpertRank
|
||||
|
||||
#define CamTypeFunc XType, ExpandXOutType, DynamicQuant, IsSmoothScaleExist, IsShareExpertRank
|
||||
|
||||
using namespace AscendC;
|
||||
template <CamTypeClass>
|
||||
class MoeDispatchNormal {
|
||||
public:
|
||||
__aicore__ inline MoeDispatchNormal(){};
|
||||
__aicore__ inline void Init(GM_ADDR x, GM_ADDR expertIds, GM_ADDR send_offset, GM_ADDR send_tokenIdx,
|
||||
GM_ADDR recv_offset, GM_ADDR recv_count, GM_ADDR expandXOut, GM_ADDR dynamicScalesOut, GM_ADDR expandIdxOut,
|
||||
GM_ADDR workspaceGM, TPipe *pipe, const MoeDispatchNormalTilingData *tilingData);
|
||||
__aicore__ inline void Process();
|
||||
|
||||
private:
|
||||
__aicore__ inline void InputToShare();
|
||||
__aicore__ inline void SetStatus();
|
||||
__aicore__ inline void WaitStatus();
|
||||
__aicore__ inline void ShareToOutput();
|
||||
__aicore__ inline void UpdateOutput();
|
||||
__aicore__ inline void FillTriple(LocalTensor<ExpandXOutType> &xOutTensor, uint32_t tokenIndex, uint32_t k);
|
||||
__aicore__ inline void QuantInit();
|
||||
__aicore__ inline void ReduceMaxInplace(const LocalTensor<float> &srcLocal, uint32_t count);
|
||||
__aicore__ inline void QuantProcess();
|
||||
__aicore__ inline GM_ADDR GetWindAddrByRankId(uint8_t ctxIdx, const int32_t rankId)
|
||||
{
|
||||
uint32_t curRankId = ((ctxIdx == COMM_EP_IDX) ? epRankId : tpRankId);
|
||||
if (curRankId == rankId) {
|
||||
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;
|
||||
}
|
||||
|
||||
__aicore__ inline GM_ADDR GetWindStateAddrByRankId(uint8_t ctxIdx, const int32_t rankId)
|
||||
{
|
||||
uint32_t curRankId = ctxIdx == COMM_EP_IDX ? epRankId : tpRankId;
|
||||
if (curRankId == rankId) {
|
||||
return (GM_ADDR)(winContext_[ctxIdx]->localWindowsExp) + dataState * WIN_STATE_OFFSET;
|
||||
}
|
||||
return (GM_ADDR)(((HcclRankRelationResV2 *)(winContext_[ctxIdx]->remoteRes[rankId].nextDevicePtr))
|
||||
->windowsExp) +
|
||||
dataState * WIN_STATE_OFFSET;
|
||||
}
|
||||
|
||||
TPipe *tpipe_{nullptr};
|
||||
GlobalTensor<XType> xGT;
|
||||
GlobalTensor<int32_t> expertIdsGT;
|
||||
GlobalTensor<int32_t> sendOffsetGT;
|
||||
GlobalTensor<int32_t> sendTokenIdxGT;
|
||||
GlobalTensor<int32_t> recvOffsetGT;
|
||||
GlobalTensor<int32_t> recvCountGT;
|
||||
GlobalTensor<float> dynamicScalesOutGT;
|
||||
GlobalTensor<int32_t> expandIdxOutGT;
|
||||
GlobalTensor<ExpandXOutType> dstGT;
|
||||
GlobalTensor<int32_t> dstStatusGT;
|
||||
|
||||
LocalTensor<XType> xInTensor;
|
||||
LocalTensor<ExpandXOutType> xOutTensor;
|
||||
LocalTensor<ExpandXOutType> xTmpTensor;
|
||||
LocalTensor<int32_t> expertIdsTensor;
|
||||
LocalTensor<int32_t> sendOffsetTensor;
|
||||
LocalTensor<int32_t> sendTokenIdxTensor;
|
||||
LocalTensor<int32_t> recvOffsetTensor;
|
||||
LocalTensor<int32_t> recvCountTensor;
|
||||
LocalTensor<int32_t> statusTensor;
|
||||
|
||||
TBuf<> expertIdsBuf;
|
||||
TBuf<> sendOffsetBuf;
|
||||
TBuf<> sendTokenIdxBuf;
|
||||
TBuf<> recvOffsetBuf;
|
||||
TBuf<> recvCountBuf;
|
||||
TBuf<> statusBuf;
|
||||
TBuf<> waitStatusBuf;
|
||||
TBuf<> gatherMaskOutBuf;
|
||||
TBuf<> scalarBuf;
|
||||
TBuf<> tokenCastFloatBuf;
|
||||
TBuf<> tokenAbsFloatBuf;
|
||||
|
||||
GM_ADDR expandXOutGM;
|
||||
GM_ADDR shareGM;
|
||||
|
||||
uint32_t batchSize{0};
|
||||
uint32_t globalBatchSize{0};
|
||||
uint32_t h{0};
|
||||
uint32_t topK{0};
|
||||
uint32_t blockNum{0};
|
||||
uint32_t blockIdx{0};
|
||||
uint32_t epRankSize{0};
|
||||
uint32_t epRankId{0};
|
||||
uint32_t tpRankSize{0};
|
||||
uint32_t tpRankId{0};
|
||||
uint32_t moeExpertNum{0};
|
||||
uint32_t moeExpertNumPerRank{0};
|
||||
|
||||
uint32_t hUBAlignSize{0};
|
||||
uint32_t hOutGMAlignSize{0};
|
||||
uint32_t hOutUBAlignSize{0};
|
||||
uint32_t hGMAlignCnt{0};
|
||||
uint32_t expandIdxStartIdx{0};
|
||||
uint32_t expertIdsCnt{0};
|
||||
uint32_t stateOffset{0};
|
||||
uint32_t dataState{0};
|
||||
uint32_t winDataSizeOffset{0};
|
||||
|
||||
uint32_t startStatusId;
|
||||
uint32_t endStatusId;
|
||||
uint32_t statusNumPerCore;
|
||||
uint32_t remainStatus;
|
||||
|
||||
TQueBind<QuePosition::VECIN, QuePosition::VECOUT, 1> xQueue;
|
||||
TQue<QuePosition::VECIN, 1> xInQueue;
|
||||
TQue<QuePosition::VECOUT, 1> xOutQueue;
|
||||
|
||||
__gm__ HcclOpResParam *winContext_[COMM_NUM]{nullptr, nullptr};
|
||||
|
||||
DataCopyExtParams hCommuCopyOutParams;
|
||||
};
|
||||
|
||||
template <CamTypeClass>
|
||||
__aicore__ inline void MoeDispatchNormal<CamTypeFunc>::Init(GM_ADDR x, GM_ADDR expertIds, GM_ADDR send_offset,
|
||||
GM_ADDR send_tokenIdx, GM_ADDR recv_offset, GM_ADDR recv_count, GM_ADDR expandXOut, GM_ADDR dynamicScalesOut,
|
||||
GM_ADDR expandIdxOut, GM_ADDR workspaceGM, TPipe *pipe, const MoeDispatchNormalTilingData *tilingData)
|
||||
{
|
||||
tpipe_ = pipe;
|
||||
blockIdx = GetBlockIdx();
|
||||
|
||||
winContext_[COMM_EP_IDX] = (__gm__ HcclOpResParam *)AscendC::GetHcclContext<HCCL_GROUP_ID_0>();
|
||||
winContext_[COMM_TP_IDX] = (__gm__ HcclOpResParam *)AscendC::GetHcclContext<1>();
|
||||
|
||||
GlobalTensor<int32_t> selfDataStatusTensor;
|
||||
GM_ADDR statusDataSpaceGm = (GM_ADDR)(winContext_[COMM_EP_IDX]->localWindowsExp);
|
||||
selfDataStatusTensor.SetGlobalBuffer(
|
||||
(__gm__ int32_t *)(statusDataSpaceGm + STATE_WIN_OFFSET + blockIdx * WIN_ADDR_ALIGN));
|
||||
|
||||
batchSize = tilingData->moeDispatchNormalInfo.bs;
|
||||
globalBatchSize = tilingData->moeDispatchNormalInfo.globalBs;
|
||||
h = tilingData->moeDispatchNormalInfo.h;
|
||||
topK = tilingData->moeDispatchNormalInfo.k;
|
||||
blockNum = tilingData->moeDispatchNormalInfo.aivNum;
|
||||
epRankSize = tilingData->moeDispatchNormalInfo.epWorldSize;
|
||||
epRankId = tilingData->moeDispatchNormalInfo.epRankId;
|
||||
moeExpertNum = tilingData->moeDispatchNormalInfo.moeExpertNum;
|
||||
moeExpertNumPerRank = moeExpertNum / epRankSize;
|
||||
|
||||
xGT.SetGlobalBuffer((__gm__ XType *)x);
|
||||
expertIdsGT.SetGlobalBuffer((__gm__ int32_t *)expertIds);
|
||||
sendOffsetGT.SetGlobalBuffer((__gm__ int32_t *)(send_offset));
|
||||
sendTokenIdxGT.SetGlobalBuffer((__gm__ int32_t *)(send_tokenIdx));
|
||||
recvOffsetGT.SetGlobalBuffer((__gm__ int32_t *)(recv_offset));
|
||||
recvCountGT.SetGlobalBuffer((__gm__ int32_t *)(recv_count));
|
||||
dynamicScalesOutGT.SetGlobalBuffer((__gm__ float *)dynamicScalesOut);
|
||||
expandIdxOutGT.SetGlobalBuffer((__gm__ int32_t *)(expandIdxOut));
|
||||
|
||||
expandXOutGM = expandXOut;
|
||||
|
||||
hUBAlignSize = Ceil(h * sizeof(ExpandXOutType), UB_ALIGN) * UB_ALIGN;
|
||||
uint32_t hScaleSizeAlign = hUBAlignSize + UB_ALIGN;
|
||||
expandIdxStartIdx = hScaleSizeAlign / sizeof(int32_t);
|
||||
|
||||
uint32_t hScaleIdxSize = hScaleSizeAlign + EXPAND_IDX_INFO * sizeof(int32_t);
|
||||
hOutGMAlignSize = Ceil(hScaleIdxSize, WIN_ADDR_ALIGN) * WIN_ADDR_ALIGN;
|
||||
hGMAlignCnt = hOutGMAlignSize / sizeof(ExpandXOutType);
|
||||
|
||||
expertIdsCnt = batchSize * topK;
|
||||
statusNumPerCore = moeExpertNum / blockNum;
|
||||
remainStatus = moeExpertNum % blockNum;
|
||||
startStatusId = statusNumPerCore * blockIdx;
|
||||
if (blockIdx < remainStatus) {
|
||||
statusNumPerCore += 1;
|
||||
startStatusId += blockIdx;
|
||||
} else {
|
||||
startStatusId += remainStatus;
|
||||
}
|
||||
endStatusId = startStatusId + statusNumPerCore;
|
||||
stateOffset = STATE_OFFSET;
|
||||
DataCacheCleanAndInvalid<int32_t, CacheLine::SINGLE_CACHE_LINE, DcciDst::CACHELINE_OUT>(selfDataStatusTensor);
|
||||
dataState = selfDataStatusTensor(0);
|
||||
if (dataState == 0) {
|
||||
selfDataStatusTensor(0) = 1;
|
||||
} else {
|
||||
selfDataStatusTensor(0) = 0;
|
||||
}
|
||||
DataCacheCleanAndInvalid<int32_t, CacheLine::SINGLE_CACHE_LINE, DcciDst::CACHELINE_OUT>(selfDataStatusTensor);
|
||||
PipeBarrier<PIPE_ALL>();
|
||||
|
||||
uint64_t hSizeAlignCombine = Ceil(h * sizeof(XType), WIN_ADDR_ALIGN) * WIN_ADDR_ALIGN;
|
||||
winDataSizeOffset = dataState * (tilingData->moeDispatchNormalInfo.totalWinSize / 2) +
|
||||
globalBatchSize / epRankSize * topK * hSizeAlignCombine;
|
||||
shareGM = GetWindAddrByRankId(COMM_EP_IDX, epRankId);
|
||||
|
||||
hOutUBAlignSize = Ceil(hScaleIdxSize, UB_ALIGN) * UB_ALIGN;
|
||||
if constexpr (DynamicQuant) {
|
||||
QuantInit();
|
||||
} else {
|
||||
tpipe_->InitBuffer(xQueue, BUFFER_NUM, hOutUBAlignSize); // 2 * 14K = 28K
|
||||
}
|
||||
|
||||
tpipe_->InitBuffer(sendOffsetBuf, moeExpertNum * sizeof(int32_t)); // 4 * moeNum
|
||||
sendOffsetTensor = sendOffsetBuf.Get<int32_t>();
|
||||
|
||||
hCommuCopyOutParams = {1U, static_cast<uint32_t>(hScaleIdxSize), 0U, 0U, 0U};
|
||||
}
|
||||
|
||||
template <CamTypeClass>
|
||||
__aicore__ inline void MoeDispatchNormal<CamTypeFunc>::QuantInit()
|
||||
{
|
||||
uint32_t hAlignSize = Ceil(h * sizeof(XType), UB_ALIGN) * UB_ALIGN;
|
||||
tpipe_->InitBuffer(xInQueue, BUFFER_NUM, hAlignSize); // 14K * 2
|
||||
tpipe_->InitBuffer(xOutQueue, BUFFER_NUM, hOutUBAlignSize); // 7K * 2
|
||||
|
||||
tpipe_->InitBuffer(tokenCastFloatBuf, h * sizeof(float)); // 28K
|
||||
tpipe_->InitBuffer(tokenAbsFloatBuf, h * sizeof(float)); // 28K
|
||||
}
|
||||
|
||||
template <CamTypeClass>
|
||||
__aicore__ inline void MoeDispatchNormal<CamTypeFunc>::ReduceMaxInplace(
|
||||
const LocalTensor<float> &srcLocal, uint32_t count)
|
||||
{
|
||||
uint64_t repsFp32 = count >> 6; // 6 is count / elemPerRefFp32
|
||||
uint64_t offsetsFp32 = repsFp32 << 6; // 6 is repsFp32 * elemPerRefFp32
|
||||
uint64_t remsFp32 = count & 0x3f; // 0x3f 63, count % elemPerRefFp32
|
||||
const uint64_t elemPerRefFp32 = 64UL; // 256 bit / sizeof(float)
|
||||
if (likely(repsFp32 > 1)) {
|
||||
// 8 is rep stride
|
||||
Max(srcLocal, srcLocal[elemPerRefFp32], srcLocal, elemPerRefFp32, repsFp32 - 1, {1, 1, 1, 0, 8, 0});
|
||||
PipeBarrier<PIPE_V>();
|
||||
}
|
||||
if (unlikely(remsFp32 > 0) && unlikely(offsetsFp32 > 0)) {
|
||||
Max(srcLocal, srcLocal[offsetsFp32], srcLocal, remsFp32, 1, {1, 1, 1, 0, 8, 0});
|
||||
PipeBarrier<PIPE_V>();
|
||||
}
|
||||
uint32_t mask = (repsFp32 > 0) ? elemPerRefFp32 : count;
|
||||
// 8 is rep stride
|
||||
WholeReduceMax(srcLocal, srcLocal, mask, 1, 8, 1, 8);
|
||||
}
|
||||
|
||||
template <CamTypeClass>
|
||||
__aicore__ inline void MoeDispatchNormal<CamTypeFunc>::QuantProcess()
|
||||
{
|
||||
float dynamicScale = 0.0;
|
||||
LocalTensor<float> floatLocalTemp;
|
||||
floatLocalTemp = tokenCastFloatBuf.Get<float>();
|
||||
|
||||
Cast(floatLocalTemp, xInTensor, RoundMode::CAST_NONE, h);
|
||||
xInQueue.FreeTensor<XType>(xInTensor);
|
||||
PipeBarrier<PIPE_V>();
|
||||
|
||||
if constexpr (DynamicQuant) {
|
||||
LocalTensor<float> floatLocalAbsTemp = tokenAbsFloatBuf.Get<float>();
|
||||
|
||||
Abs(floatLocalAbsTemp, floatLocalTemp, h);
|
||||
PipeBarrier<PIPE_V>();
|
||||
ReduceMaxInplace(floatLocalAbsTemp, h);
|
||||
|
||||
SyncFunc<AscendC::HardEvent::V_S>();
|
||||
dynamicScale = float(127.0) / (floatLocalAbsTemp.GetValue(0) + 1e-12f);
|
||||
SyncFunc<AscendC::HardEvent::S_V>();
|
||||
Muls(floatLocalTemp, floatLocalTemp, dynamicScale, h);
|
||||
PipeBarrier<PIPE_V>();
|
||||
}
|
||||
LocalTensor<half> halfLocalTemp = floatLocalTemp.ReinterpretCast<half>();
|
||||
LocalTensor<int32_t> int32LocalTemp = floatLocalTemp.ReinterpretCast<int32_t>();
|
||||
Cast(int32LocalTemp, floatLocalTemp, RoundMode::CAST_RINT, h);
|
||||
PipeBarrier<PIPE_V>();
|
||||
SetDeqScale((half)1.000000e+00f);
|
||||
PipeBarrier<PIPE_V>();
|
||||
|
||||
Cast(halfLocalTemp, int32LocalTemp, RoundMode::CAST_ROUND, h);
|
||||
|
||||
PipeBarrier<PIPE_V>();
|
||||
Cast(xOutTensor, halfLocalTemp, RoundMode::CAST_TRUNC, h);
|
||||
|
||||
floatLocalTemp = xOutTensor.template ReinterpretCast<float>();
|
||||
floatLocalTemp.SetValue(hUBAlignSize / sizeof(float), float(1.0) / dynamicScale); // int8->float32
|
||||
}
|
||||
|
||||
template <CamTypeClass>
|
||||
__aicore__ inline void MoeDispatchNormal<CamTypeFunc>::FillTriple(
|
||||
LocalTensor<ExpandXOutType> &xOutTensor, uint32_t tokenIndex, uint32_t k)
|
||||
{
|
||||
SyncFunc<AscendC::HardEvent::MTE3_S>();
|
||||
LocalTensor<int32_t> xOutTint32 = xOutTensor.template ReinterpretCast<int32_t>();
|
||||
xOutTint32(expandIdxStartIdx) = epRankId;
|
||||
xOutTint32(expandIdxStartIdx + 1) = tokenIndex;
|
||||
xOutTint32(expandIdxStartIdx + 2) = k;
|
||||
SyncFunc<AscendC::HardEvent::S_MTE3>();
|
||||
}
|
||||
|
||||
template <CamTypeClass>
|
||||
__aicore__ inline void MoeDispatchNormal<CamTypeFunc>::InputToShare()
|
||||
{
|
||||
DataCopyExtParams sendOffsetParams = {1U, static_cast<uint32_t>(moeExpertNum * sizeof(uint32_t)), 0U, 0U, 0U};
|
||||
DataCopyPadExtParams<int32_t> sendOffsetCopyPadParams{false, 0U, 0U, 0U};
|
||||
DataCopyPad(sendOffsetTensor, sendOffsetGT, sendOffsetParams, sendOffsetCopyPadParams);
|
||||
SyncFunc<AscendC::HardEvent::MTE2_S>();
|
||||
|
||||
uint32_t startTokenId, endTokenId, sendTokenNum, remainTokenNum;
|
||||
sendTokenNum = expertIdsCnt / blockNum;
|
||||
remainTokenNum = expertIdsCnt % blockNum;
|
||||
startTokenId = sendTokenNum * blockIdx;
|
||||
if (blockIdx < remainTokenNum) {
|
||||
sendTokenNum += 1;
|
||||
startTokenId += blockIdx;
|
||||
} else {
|
||||
startTokenId += remainTokenNum;
|
||||
}
|
||||
endTokenId = startTokenId + sendTokenNum;
|
||||
|
||||
if (startTokenId >= expertIdsCnt) {
|
||||
return;
|
||||
}
|
||||
tpipe_->InitBuffer(expertIdsBuf, sendTokenNum * sizeof(int32_t)); // 4 * bs * k / 48
|
||||
tpipe_->InitBuffer(sendTokenIdxBuf, sendTokenNum * sizeof(int32_t)); // 4 * bs * k / 48
|
||||
expertIdsTensor = expertIdsBuf.Get<int32_t>();
|
||||
sendTokenIdxTensor = sendTokenIdxBuf.Get<int32_t>();
|
||||
DataCopyExtParams expertIdsCntParams = {1U, static_cast<uint32_t>(sendTokenNum * sizeof(uint32_t)), 0U, 0U, 0U};
|
||||
DataCopyExtParams sendTokenIdxParams = {1U, static_cast<uint32_t>(sendTokenNum * sizeof(uint32_t)), 0U, 0U, 0U};
|
||||
DataCopyPadExtParams<int32_t> copyPadExtParams{false, 0U, 0U, 0U};
|
||||
DataCopyPadExtParams<XType> tokenCopyPadExtParams{false, 0U, 0U, 0U};
|
||||
DataCopyPad(expertIdsTensor, expertIdsGT[startTokenId], expertIdsCntParams, copyPadExtParams);
|
||||
DataCopyPad(sendTokenIdxTensor, sendTokenIdxGT[startTokenId], sendTokenIdxParams, copyPadExtParams);
|
||||
SyncFunc<AscendC::HardEvent::MTE2_S>();
|
||||
|
||||
DataCopyExtParams xCopyParams = {1U, static_cast<uint32_t>(h * sizeof(XType)), 0U, 0U, 0U};
|
||||
for (int32_t tokenIndex = startTokenId; tokenIndex < endTokenId; ++tokenIndex) {
|
||||
uint32_t dstExpertId = expertIdsTensor(tokenIndex - startTokenId);
|
||||
int32_t curExpertCnt = sendTokenIdxTensor(tokenIndex - startTokenId);
|
||||
int32_t dstExpertOffset = sendOffsetTensor(dstExpertId);
|
||||
GM_ADDR rankGM =
|
||||
(__gm__ uint8_t *)(shareGM + hOutGMAlignSize * (dstExpertOffset + curExpertCnt));
|
||||
dstGT.SetGlobalBuffer((__gm__ ExpandXOutType *)rankGM);
|
||||
|
||||
if constexpr (DynamicQuant) {
|
||||
xInTensor = xInQueue.AllocTensor<XType>();
|
||||
DataCopyPad(xInTensor, xGT[tokenIndex / topK * h], xCopyParams, tokenCopyPadExtParams);
|
||||
xInQueue.EnQue(xInTensor);
|
||||
xInTensor = xInQueue.DeQue<XType>();
|
||||
xOutTensor = xOutQueue.AllocTensor<ExpandXOutType>();
|
||||
QuantProcess();
|
||||
xOutQueue.EnQue(xOutTensor);
|
||||
xOutTensor = xOutQueue.DeQue<ExpandXOutType>();
|
||||
FillTriple(xOutTensor, tokenIndex / topK, tokenIndex % topK);
|
||||
DataCopyPad(dstGT, xOutTensor, hCommuCopyOutParams);
|
||||
xOutQueue.FreeTensor(xOutTensor);
|
||||
} else {
|
||||
xTmpTensor = xQueue.AllocTensor<ExpandXOutType>();
|
||||
DataCopyPad(xTmpTensor, xGT[tokenIndex / topK * h], xCopyParams, tokenCopyPadExtParams);
|
||||
xQueue.EnQue(xTmpTensor);
|
||||
xTmpTensor = xQueue.DeQue<ExpandXOutType>();
|
||||
FillTriple(xTmpTensor, tokenIndex / topK, tokenIndex % topK);
|
||||
DataCopyPad(dstGT, xTmpTensor, hCommuCopyOutParams);
|
||||
xQueue.FreeTensor<ExpandXOutType>(xTmpTensor);
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
template <CamTypeClass>
|
||||
__aicore__ inline void MoeDispatchNormal<CamTypeFunc>::SetStatus()
|
||||
{
|
||||
uint32_t startExpId, endExpId, expNumPerCore;
|
||||
expNumPerCore = statusNumPerCore;
|
||||
startExpId = startStatusId;
|
||||
endExpId = endStatusId;
|
||||
if (startExpId > moeExpertNum) {
|
||||
SyncAll<true>();
|
||||
return;
|
||||
}
|
||||
uint32_t statusCntAlign = Ceil(expNumPerCore, 8) * 8;
|
||||
tpipe_->InitBuffer(statusBuf, statusCntAlign * UB_ALIGN); // moeNum / 48 * 32
|
||||
statusTensor = statusBuf.Get<int32_t>();
|
||||
Duplicate<int32_t>(statusTensor, 0, expNumPerCore * 8);
|
||||
uint64_t mask[2] = {0x101010101010101, 0};
|
||||
PipeBarrier<PIPE_V>();
|
||||
Duplicate<int32_t>(statusTensor, 0x3F800000, mask, statusCntAlign / 8, 1, 8);
|
||||
PipeBarrier<PIPE_ALL>();
|
||||
SyncAll<true>();
|
||||
for (uint32_t i = startExpId; i < endExpId; ++i) {
|
||||
uint32_t targetRankId = i / moeExpertNumPerRank;
|
||||
uint32_t offset = stateOffset * (epRankId + i % moeExpertNumPerRank * epRankSize);
|
||||
GM_ADDR rankGM = (__gm__ uint8_t *)(GetWindStateAddrByRankId(COMM_EP_IDX, targetRankId) + offset);
|
||||
dstStatusGT.SetGlobalBuffer((__gm__ int32_t *)rankGM);
|
||||
DataCopy<int32_t>(dstStatusGT, statusTensor[(i - startExpId) * 8], 8UL);
|
||||
}
|
||||
SyncFunc<AscendC::HardEvent::MTE3_S>();
|
||||
}
|
||||
|
||||
template <CamTypeClass>
|
||||
__aicore__ inline void MoeDispatchNormal<CamTypeFunc>::WaitStatus()
|
||||
{
|
||||
tpipe_->Reset();
|
||||
uint32_t waitStatusBufSize = (((statusNumPerCore * UB_ALIGN) > 256) ? (statusNumPerCore * UB_ALIGN) : 256);
|
||||
tpipe_->InitBuffer(waitStatusBuf, waitStatusBufSize); // moeNum /48 * 32B = 43 * 32B
|
||||
tpipe_->InitBuffer(gatherMaskOutBuf, moeExpertNum * sizeof(float)); // moeNum * 4B
|
||||
tpipe_->InitBuffer(scalarBuf, UB_ALIGN * 3); // 96B
|
||||
tpipe_->InitBuffer(xQueue, BUFFER_NUM, hOutUBAlignSize); // 28K
|
||||
tpipe_->InitBuffer(recvOffsetBuf, moeExpertNum * sizeof(int32_t)); // moeNum * 4B
|
||||
tpipe_->InitBuffer(recvCountBuf, moeExpertNum * sizeof(int32_t)); // moeNum * 4B
|
||||
|
||||
recvOffsetTensor = recvOffsetBuf.Get<int32_t>();
|
||||
recvCountTensor = recvCountBuf.Get<int32_t>();
|
||||
DataCopyExtParams recvOffsetParams = {1U, static_cast<uint32_t>(moeExpertNum * sizeof(uint32_t)), 0U, 0U, 0U};
|
||||
DataCopyExtParams recvCountParams = {1U, static_cast<uint32_t>(moeExpertNum * sizeof(uint32_t)), 0U, 0U, 0U};
|
||||
DataCopyPadExtParams<int32_t> copyPadExtParams{false, 0U, 0U, 0U};
|
||||
DataCopyPad(recvOffsetTensor, recvOffsetGT, recvOffsetParams, copyPadExtParams);
|
||||
DataCopyPad(recvCountTensor, recvCountGT, recvCountParams, copyPadExtParams);
|
||||
|
||||
if (startStatusId >= moeExpertNum) {
|
||||
SyncAll<true>();
|
||||
return;
|
||||
}
|
||||
|
||||
LocalTensor<float> gatherMaskOutTensor = gatherMaskOutBuf.Get<float>();
|
||||
LocalTensor<float> statusSumOutTensor = scalarBuf.GetWithOffset<float>(UB_ALIGN / sizeof(float), UB_ALIGN);
|
||||
LocalTensor<float> statusFp32Tensor = waitStatusBuf.Get<float>();
|
||||
GlobalTensor<float> windowInstatusFp32Tensor;
|
||||
windowInstatusFp32Tensor.SetGlobalBuffer((__gm__ float *)(GetWindStateAddrByRankId(COMM_EP_IDX, epRankId)));
|
||||
uint32_t mask = 1;
|
||||
float compareTarget = static_cast<float>(1.0) * statusNumPerCore;
|
||||
float sumOfFlag = static_cast<float>(-1.0);
|
||||
DataCopyParams intriParams{static_cast<uint16_t>(statusNumPerCore), 1, 0, 0};
|
||||
SyncFunc<AscendC::HardEvent::S_V>();
|
||||
while (sumOfFlag != compareTarget) {
|
||||
DataCopy(statusFp32Tensor, windowInstatusFp32Tensor[startStatusId * stateOffset / sizeof(float)], intriParams);
|
||||
SyncFunc<AscendC::HardEvent::MTE2_V>();
|
||||
ReduceSum(statusSumOutTensor, statusFp32Tensor, gatherMaskOutTensor, mask, statusNumPerCore, 1);
|
||||
SyncFunc<AscendC::HardEvent::V_S>();
|
||||
sumOfFlag = statusSumOutTensor.GetValue(0);
|
||||
}
|
||||
|
||||
// Clear state
|
||||
SyncFunc<AscendC::HardEvent::MTE3_S>();
|
||||
DataCopyParams intriOutParams{static_cast<uint16_t>(statusNumPerCore), 1, 0, 0};
|
||||
uint64_t duplicateMask[2] = {0x101010101010101, 0};
|
||||
LocalTensor<int32_t> cleanStateTensor = waitStatusBuf.Get<int32_t>();
|
||||
SyncFunc<AscendC::HardEvent::S_V>();
|
||||
Duplicate<int32_t>(cleanStateTensor, 0, duplicateMask, Ceil(statusNumPerCore, 8), 1, 8);
|
||||
SyncFunc<AscendC::HardEvent::V_MTE3>();
|
||||
DataCopy(windowInstatusFp32Tensor[startStatusId * stateOffset / sizeof(float)],
|
||||
cleanStateTensor.ReinterpretCast<float>(),
|
||||
intriOutParams);
|
||||
SyncFunc<AscendC::HardEvent::MTE3_S>();
|
||||
SyncAll<true>();
|
||||
}
|
||||
|
||||
template <CamTypeClass>
|
||||
__aicore__ inline void MoeDispatchNormal<CamTypeFunc>::ShareToOutput()
|
||||
{
|
||||
if (startStatusId >= moeExpertNum) {
|
||||
return;
|
||||
}
|
||||
uint32_t fromRank, count, preCount, recvOffset, targetOffset;
|
||||
DataCopyPadExtParams<ExpandXOutType> copyPadExtParams{false, 0U, 0U, 0U};
|
||||
DataCopyExtParams dataCopyExandIdxParams{1U, sizeof(int32_t) * EXPAND_IDX_INFO, 0U, 0U, 0U};
|
||||
DataCopyExtParams dataCopyOutParams{1U, static_cast<uint32_t>(statusNumPerCore * sizeof(int32_t)), 0U, 0U, 0U};
|
||||
DataCopyExtParams expandXCopyParams = {1U, static_cast<uint32_t>(h * sizeof(ExpandXOutType)), 0U, 0U, 0U};
|
||||
LocalTensor<int32_t> xTmpTensorInt;
|
||||
AscendC::TQueSync<PIPE_MTE2, PIPE_S> recvCountLocalSync;
|
||||
recvCountLocalSync.SetFlag(0);
|
||||
recvCountLocalSync.WaitFlag(0);
|
||||
for (uint32_t i = startStatusId; i < endStatusId; ++i) {
|
||||
preCount = 0;
|
||||
if (likely(i != 0)) {
|
||||
preCount = recvCountTensor(i - 1);
|
||||
}
|
||||
fromRank = i % epRankSize;
|
||||
count = recvCountTensor(i) - preCount;
|
||||
recvOffset = recvOffsetTensor(i);
|
||||
targetOffset = preCount;
|
||||
GM_ADDR recvStart =
|
||||
(__gm__ uint8_t *)(GetWindAddrByRankId(COMM_EP_IDX, fromRank)) + recvOffset * hOutGMAlignSize;
|
||||
GlobalTensor<ExpandXOutType> srcTokenGT, dstTokenGT;
|
||||
for (uint32_t j = 0; j < count; ++j) {
|
||||
srcTokenGT.SetGlobalBuffer((__gm__ ExpandXOutType *)(recvStart + j * hOutGMAlignSize));
|
||||
xTmpTensor = xQueue.AllocTensor<ExpandXOutType>();
|
||||
DataCopyPad(xTmpTensor, srcTokenGT, hCommuCopyOutParams, copyPadExtParams);
|
||||
xQueue.EnQue(xTmpTensor);
|
||||
xTmpTensor = xQueue.DeQue<ExpandXOutType>();
|
||||
xTmpTensorInt = xTmpTensor.template ReinterpretCast<int32_t>();
|
||||
DataCopyPad(expandIdxOutGT[(targetOffset + j) * EXPAND_IDX_INFO],
|
||||
xTmpTensorInt[expandIdxStartIdx],
|
||||
dataCopyExandIdxParams);
|
||||
if constexpr (DynamicQuant) {
|
||||
DataCopyExtParams floatDataCopyParams = {1U, sizeof(float), 0U, 0U, 0U};
|
||||
LocalTensor<float> xOutFp32Tensor = xTmpTensor.template ReinterpretCast<float>();
|
||||
DataCopyPad(dynamicScalesOutGT[targetOffset + j],
|
||||
xOutFp32Tensor[hUBAlignSize / sizeof(float)],
|
||||
floatDataCopyParams);
|
||||
}
|
||||
dstTokenGT.SetGlobalBuffer((__gm__ ExpandXOutType *)(expandXOutGM) + (targetOffset + j) * h, h);
|
||||
DataCopyPad(dstTokenGT, xTmpTensor, expandXCopyParams);
|
||||
xQueue.FreeTensor(xTmpTensor);
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
template <CamTypeClass>
|
||||
__aicore__ inline void MoeDispatchNormal<CamTypeFunc>::Process()
|
||||
{
|
||||
if ASCEND_IS_AIV {
|
||||
InputToShare();
|
||||
SetStatus();
|
||||
WaitStatus();
|
||||
ShareToOutput();
|
||||
}
|
||||
}
|
||||
|
||||
} // namespace MoeDispatchNormalImpl
|
||||
#endif
|
||||
@@ -0,0 +1,30 @@
|
||||
#ifndef MOE_DISPATCH_NORMAL_TILING_H
|
||||
#define MOE_DISPATCH_NORMAL_TILING_H
|
||||
|
||||
struct MoeDispatchNormalInfo {
|
||||
uint32_t epWorldSize; // epWorldSize
|
||||
uint32_t tpWorldSize; // tpWorldSize
|
||||
uint32_t epRankId; // epRankId
|
||||
uint32_t tpRankId; // tpRankId
|
||||
uint32_t moeExpertNum; // moe expert number
|
||||
uint32_t quantMode; // quant mode
|
||||
uint32_t globalBs; // globalBs = BS * worldSize
|
||||
uint32_t bs; // bs
|
||||
uint32_t k; // k
|
||||
uint32_t h; // h
|
||||
uint32_t aivNum; // aivNum
|
||||
bool isQuant; // whether quant or not
|
||||
bool reserved2; // reserved
|
||||
bool reserved3; // reserved
|
||||
uint64_t totalUbSize; // epWorldSize
|
||||
uint64_t totalWinSize;
|
||||
};
|
||||
|
||||
struct MoeDispatchNormalTilingData {
|
||||
Mc2InitTiling mc2InitTiling;
|
||||
Mc2CcTiling mc2CcTiling1;
|
||||
Mc2CcTiling mc2CcTiling2;
|
||||
MoeDispatchNormalInfo moeDispatchNormalInfo;
|
||||
};
|
||||
|
||||
#endif
|
||||
Reference in New Issue
Block a user