EngineX/xc-llm-ascend
3
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

[Kernel] Add moe normal ops (#4810)

Browse Source 
### 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
2025-12-10 17:15:28 +08:00
committed by GitHub
parent c77dca54b2
commit bd8be2e759
39 changed files with 5365 additions and 4 deletions
Download Patch File Download Diff File Expand all files Collapse all files

49
csrc/moe_dispatch_normal/op_host/CMakeLists.txt Normal file
View File

@@ -0,0 +1,49 @@
# Copyright (c) 2025 Huawei Technologies Co., Ltd.
# This file is a part of the CANN Open Software.
# Licensed under CANN Open Software License Agreement Version 1.0 (the "License").
# Please refer to the License for details. You may not use this file except in compliance with the License.
# THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED,
# INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.
# See LICENSE in the root of the software repository for the full text of the License.
# ======================================================================================================================
add_ops_compile_options(
OP_NAME MoeDispatchNormal
OPTIONS --cce-auto-sync=off
-Wno-deprecated-declarations
-Werror
)
target_sources(op_host_aclnnInner PRIVATE
moe_dispatch_normal.cpp
)
target_sources(opapi PRIVATE
aclnn_moe_dispatch_normal.cpp
)
if (NOT BUILD_OPEN_PROJECT)
target_sources(aclnn_ops_train PRIVATE
aclnn_moe_dispatch_normal.cpp
)
target_sources(aclnn_ops_infer PRIVATE
aclnn_moe_dispatch_normal.cpp
)
endif ()
target_sources(optiling PRIVATE
moe_dispatch_normal_tiling.cpp
)
target_include_directories(optiling PRIVATE
${CMAKE_CURRENT_SOURCE_DIR}
)
target_sources(opsproto PRIVATE)
file(GLOB _GMM_Aclnn_header "${CMAKE_CURRENT_SOURCE_DIR}/aclnn_moe_dispatch_normal.h")
install(FILES ${_GMM_Aclnn_header}
DESTINATION ${ACLNN_INC_INSTALL_DIR} OPTIONAL
)

84
csrc/moe_dispatch_normal/op_host/aclnn_moe_dispatch_normal.cpp Normal file
View File

@@ -0,0 +1,84 @@
#include <string.h>
#include "graph/types.h"
#include "aclnn_moe_dispatch_normal.h"
enum NnopbaseHcclServerType {
NNOPBASE_HCCL_SERVER_TYPE_AICPU = 0,
NNOPBASE_HCCL_SERVER_TYPE_MTE,
NNOPBASE_HCCL_SERVER_TYPE_END
};
extern "C" void __attribute__((weak)) NnopbaseSetHcclServerType(void *executor, NnopbaseHcclServerType sType);
#ifdef __cplusplus
extern "C" {
#endif
extern aclnnStatus aclnnInnerMoeDispatchNormalGetWorkspaceSize(
const aclTensor *x,
const aclTensor *topkIdx,
const aclTensor *sendOffset,
const aclTensor *sendTokenIdx,
const aclTensor *recvOffset,
const aclTensor *recvCount,
char *groupEp,
int64_t epWorldSize,
int64_t epRankId,
char *groupTpOptional,
int64_t tpWorldSize,
int64_t tpRankId,
int64_t moeExpertNum,
int64_t quantMode,
int64_t globalBs,
const aclTensor *recvX,
const aclTensor *recvXScales,
const aclTensor *assistInfoForCombine,
uint64_t *workspaceSize,
aclOpExecutor **executor);
extern aclnnStatus aclnnInnerMoeDispatchNormal(
void *workspace,
uint64_t workspaceSize,
aclOpExecutor *executor,
aclrtStream stream);
aclnnStatus aclnnMoeDispatchNormalGetWorkspaceSize(const aclTensor *x, const aclTensor *topkIdx,
const aclTensor *sendOffset, const aclTensor *sendTokenIdx, const aclTensor *recvOffset, const aclTensor *recvCount,
char *groupEp, int64_t epWorldSize, int64_t epRankId, char *groupTpOptional, int64_t tpWorldSize, int64_t tpRankId,
int64_t moeExpertNum, int64_t quantMode, int64_t globalBs, const aclTensor *recvX,
const aclTensor *recvXScales, const aclTensor *assistInfoForCombine, uint64_t *workspaceSize,
aclOpExecutor **executor)
{
return aclnnInnerMoeDispatchNormalGetWorkspaceSize(x,
topkIdx,
sendOffset,
sendTokenIdx,
recvOffset,
recvCount,
groupEp,
epWorldSize,
epRankId,
groupTpOptional,
tpWorldSize,
tpRankId,
moeExpertNum,
quantMode,
globalBs,
recvX,
recvXScales,
assistInfoForCombine,
workspaceSize,
executor);
}
aclnnStatus aclnnMoeDispatchNormal(
void *workspace, uint64_t workspaceSize, aclOpExecutor *executor, aclrtStream stream)
{
if (NnopbaseSetHcclServerType) {
NnopbaseSetHcclServerType(executor, NNOPBASE_HCCL_SERVER_TYPE_MTE);
}
return aclnnInnerMoeDispatchNormal(workspace, workspaceSize, executor, stream);
}
#ifdef __cplusplus
}
#endif

24
csrc/moe_dispatch_normal/op_host/aclnn_moe_dispatch_normal.h Normal file
View File

@@ -0,0 +1,24 @@
#ifndef ACLNN_MOE_DISPATCH_NORMAL_H_
#define ACLNN_MOE_DISPATCH_NORMAL_H_
#include "aclnn/acl_meta.h"
#ifdef __cplusplus
extern "C" {
#endif
__attribute__((visibility("default"))) aclnnStatus aclnnMoeDispatchNormalGetWorkspaceSize(const aclTensor *x,
const aclTensor *topkIdx, const aclTensor *sendOffset, const aclTensor *sendTokenIdx, const aclTensor *recvOffset,
const aclTensor *recvCount, char *groupEp, int64_t epWorldSize, int64_t epRankId, char *groupTpOptional,
int64_t tpWorldSize, int64_t tpRankId, int64_t moeExpertNum, int64_t quantMode, int64_t globalBs,
const aclTensor *recvX, const aclTensor *recvXScales, const aclTensor *assistInfoForCombine,
uint64_t *workspaceSize, aclOpExecutor **executor);
__attribute__((visibility("default"))) aclnnStatus aclnnMoeDispatchNormal(
void *workspace, uint64_t workspaceSize, aclOpExecutor *executor, aclrtStream stream);
#ifdef __cplusplus
}
#endif
#endif

92
csrc/moe_dispatch_normal/op_host/moe_dispatch_normal.cpp Normal file
View File

@@ -0,0 +1,92 @@
#include "register/op_def_registry.h"
namespace ops {
class MoeDispatchNormal : public OpDef {
public:
explicit MoeDispatchNormal(const char *name) : OpDef(name)
{
this->Input("x")
.ParamType(REQUIRED)
.DataType({ge::DT_BF16, ge::DT_BF16, ge::DT_FLOAT16, ge::DT_FLOAT16})
.Format({ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND})
.UnknownShapeFormat({ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND})
.AutoContiguous();
this->Input("topk_idx")
.ParamType(REQUIRED)
.DataType({ge::DT_INT32, ge::DT_INT32, ge::DT_INT32, ge::DT_INT32})
.Format({ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND})
.UnknownShapeFormat({ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND})
.AutoContiguous();
this->Input("send_offset")
.ParamType(REQUIRED)
.DataType({ge::DT_INT32, ge::DT_INT32, ge::DT_INT32, ge::DT_INT32})
.Format({ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND})
.UnknownShapeFormat({ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND})
.AutoContiguous();
this->Input("send_tokenIdx")
.ParamType(REQUIRED)
.DataType({ge::DT_INT32, ge::DT_INT32, ge::DT_INT32, ge::DT_INT32})
.Format({ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND})
.UnknownShapeFormat({ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND})
.AutoContiguous();
this->Input("recv_offset")
.ParamType(REQUIRED)
.DataType({ge::DT_INT32, ge::DT_INT32, ge::DT_INT32, ge::DT_INT32})
.Format({ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND})
.UnknownShapeFormat({ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND})
.AutoContiguous();
this->Input("recv_count")
.ParamType(REQUIRED)
.DataType({ge::DT_INT32, ge::DT_INT32, ge::DT_INT32, ge::DT_INT32})
.Format({ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND})
.UnknownShapeFormat({ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND})
.AutoContiguous();
this->Output("recv_x")
.ParamType(REQUIRED)
.DataType({ge::DT_BF16, ge::DT_INT8, ge::DT_FLOAT16, ge::DT_INT8})
.Format({ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND})
.UnknownShapeFormat({ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND});
this->Output("x_scales")
.ParamType(REQUIRED)
.DataType({ge::DT_FLOAT, ge::DT_FLOAT, ge::DT_FLOAT, ge::DT_FLOAT})
.Format({ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND})
.UnknownShapeFormat({ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND});
this->Output("assist_info_for_combine")
.ParamType(REQUIRED)
.DataType({ge::DT_INT32, ge::DT_INT32, ge::DT_INT32, ge::DT_INT32})
.Format({ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND})
.UnknownShapeFormat({ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND});
this->Attr("group_ep").AttrType(REQUIRED).String();
this->Attr("ep_world_size").AttrType(REQUIRED).Int();
this->Attr("ep_rank_id").AttrType(REQUIRED).Int();
this->Attr("group_tp").AttrType(OPTIONAL).String("");
this->Attr("tp_world_size").AttrType(OPTIONAL).Int(0);
this->Attr("tp_rank_id").AttrType(OPTIONAL).Int(0);
this->Attr("moe_expert_num").AttrType(REQUIRED).Int();
this->Attr("quant_mode").AttrType(OPTIONAL).Int(0);
this->Attr("global_bs").AttrType(OPTIONAL).Int(0);
OpAICoreConfig aicore_config;
aicore_config.DynamicCompileStaticFlag(true)
.DynamicFormatFlag(true)
.DynamicRankSupportFlag(true)
.DynamicShapeSupportFlag(true)
.NeedCheckSupportFlag(false)
.PrecisionReduceFlag(true)
.ExtendCfgInfo("aclnnSupport.value", "support_aclnn")
.ExtendCfgInfo("jitCompile.flag", "static_true")
.ExtendCfgInfo("multiKernelSupportDynamicGraph.value", "multi_kernel");
this->AICore().AddConfig("ascend910_93", aicore_config);
this->MC2().HcclGroup({"group_ep", "group_tp"});
}
};
OP_ADD(MoeDispatchNormal);
} // namespace ops

635
csrc/moe_dispatch_normal/op_host/moe_dispatch_normal_tiling.cpp Normal file
View File

@@ -0,0 +1,635 @@
#include <queue>
#include <vector>
#include <dlfcn.h>
#include <fcntl.h>
#include <cstdio>
#include <cstdlib>
#include <cstring>
#include <sys/stat.h>
#include <sys/types.h>
#include <unistd.h>
#include <cmath>
#include <cstdint>
#include <string>
#include "register/tilingdata_base.h"
#include "tiling/tiling_api.h"
#include "log/ops_log.h"
#include "graph/utils/type_utils.h"
#include "register/op_def_registry.h"
#include "../op_kernel/moe_dispatch_normal_tiling.h"
using namespace AscendC;
using namespace ge;
namespace {
class Mc2TilingUtils {
public:
#define HCCL_BUFFSIZE "HCCL_BUFFSIZE"
static uint64_t GetMaxWindowSize()
{
uint16_t defaultWindowSize = 200;
if (getenv(HCCL_BUFFSIZE) == nullptr) {
OPS_LOG_D("", "Env HCCL_BUFFSIZE don't set");
} else {
try {
std::string envStr(getenv(HCCL_BUFFSIZE));
defaultWindowSize = std::stoi(envStr);
} catch (const std::invalid_argument &ia) {
OPS_LOG_E("", "Invalid argument when parsing HCCL_BUFFSIZE: %s", ia.what());
} catch (const std::out_of_range &oor) {
OPS_LOG_E("", "Out of range when parsing HCCL_BUFFSIZE: %s", oor.what());
}
}
const uint64_t maxWindowSize = static_cast<uint64_t>(defaultWindowSize) * 1024UL * 1024UL;
OPS_LOG_I("", "Get maxWindowSize is %lu", maxWindowSize);
return maxWindowSize;
}
};
constexpr uint32_t X_INDEX = 0U;
constexpr uint32_t EXPERT_IDS_INDEX = 1U;
constexpr uint32_t SEND_OFFSET_INDEX = 2U;
constexpr uint32_t SEND_TOKENIDX_INDEX = 3U;
constexpr uint32_t RECV_OFFSET_INDEX = 4U;
constexpr uint32_t RECV_COUNT_INDEX = 5U;
constexpr uint32_t OUTPUT_EXPAND_X_INDEX = 0U;
constexpr uint32_t OUTPUT_DYNAMIC_SCALES_INDEX = 1U;
constexpr uint32_t OUTPUT_ASSIST_INFO_INDEX = 2U;
constexpr uint32_t ATTR_GROUP_EP_INDEX = 0;
constexpr uint32_t ATTR_EP_WORLD_SIZE_INDEX = 1;
constexpr uint32_t ATTR_EP_RANK_ID_INDEX = 2;
constexpr uint32_t ATTR_GROUP_TP_INDEX = 3;
constexpr uint32_t ATTR_TP_WORLD_SIZE_INDEX = 4;
constexpr uint32_t ATTR_TP_RANK_ID_INDEX = 5;
constexpr uint32_t ATTR_MOE_EXPERT_NUM_INDEX = 6;
constexpr uint32_t ATTR_QUANT_MODE_INDEX = 7;
constexpr uint32_t ATTR_GLOBAL_BS_INDEX = 8;
constexpr uint32_t TWO_DIMS = 2;
constexpr uint32_t ONE_DIM = 1;
constexpr uint32_t DYNAMIC_SCALE_DIM_NUM = 1;
constexpr uint64_t INIT_TILINGKEY = 10000;
constexpr uint32_t OP_TYPE_ALL_TO_ALL = 8;
constexpr uint32_t NO_SCALES = 0;
constexpr uint32_t DYNAMIC_SCALES = 2;
constexpr uint32_t OP_TYPE_ALL_GATHER = 6;
constexpr size_t MAX_GROUP_NAME_LENGTH = 128UL;
constexpr int64_t MAX_EP_WORLD_SIZE = 384;
constexpr int64_t MIN_EP_WORLD_SIZE = 2;
constexpr int64_t MAX_TP_WORLD_SIZE = 2;
constexpr int64_t BS_UPPER_BOUND = 8000; // Maximum bs
constexpr uint32_t TILINGKEY_TP_WORLD_SIZE = 100;
constexpr uint32_t TP_WORLD_SIZE_TWO = 2;
constexpr int64_t MOE_EXPERT_MAX_NUM = 512;
constexpr int64_t K_MAX = 16;
constexpr uint32_t SYSTEM_NEED_WORKSPACE = 16 * 1024 * 1024;
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
constexpr uint64_t DOUBLE_DATA_BUFFER = 2UL;
constexpr uint64_t MAX_OUT_DTYPE_SIZE = 2UL;
constexpr uint64_t UB_ALIGN = 32UL;
constexpr int64_t DISPATCH_STATUS_MAX_SUPPORT_NUM = 1280UL;
} // namespace
namespace optiling {
static void PrintTilingDataInfo(const char *nodeName, MoeDispatchNormalTilingData &tilingData)
{
OPS_LOG_D(nodeName, "epWorldSize is %u.", tilingData.moeDispatchNormalInfo.epWorldSize);
OPS_LOG_D(nodeName, "tpWorldSize is %u.", tilingData.moeDispatchNormalInfo.tpWorldSize);
OPS_LOG_D(nodeName, "epRankId is %u.", tilingData.moeDispatchNormalInfo.epRankId);
OPS_LOG_D(nodeName, "tpRankId is %u.", tilingData.moeDispatchNormalInfo.tpRankId);
OPS_LOG_D(nodeName, "moeExpertNum is %u.", tilingData.moeDispatchNormalInfo.moeExpertNum);
OPS_LOG_D(nodeName, "quantMode is %u.", tilingData.moeDispatchNormalInfo.quantMode);
OPS_LOG_D(nodeName, "globalBs is %u.", tilingData.moeDispatchNormalInfo.globalBs);
OPS_LOG_D(nodeName, "bs is %u.", tilingData.moeDispatchNormalInfo.bs);
OPS_LOG_D(nodeName, "k is %u.", tilingData.moeDispatchNormalInfo.k);
OPS_LOG_D(nodeName, "h is %u.", tilingData.moeDispatchNormalInfo.h);
OPS_LOG_D(nodeName, "aivNum is %u.", tilingData.moeDispatchNormalInfo.aivNum);
OPS_LOG_D(nodeName, "totalUbSize is %lu.", tilingData.moeDispatchNormalInfo.totalUbSize);
OPS_LOG_D(nodeName, "totalWinSize is %lu.", tilingData.moeDispatchNormalInfo.totalWinSize);
}
static bool CheckTensorDim(gert::TilingContext *context, const char *nodeName, const uint32_t quantMode)
{
const gert::StorageShape *xStorageShape = context->GetInputShape(X_INDEX);
OPS_CHECK(xStorageShape == nullptr, OPS_LOG_E(nodeName, "xShape is null."), return false);
OPS_CHECK(xStorageShape->GetStorageShape().GetDimNum() != TWO_DIMS,
OPS_LOG_E(nodeName,
"xShape dims must be 2, but current dim num is %lu.",
xStorageShape->GetStorageShape().GetDimNum()),
return false);
int64_t xDim0 = xStorageShape->GetStorageShape().GetDim(0);
int64_t xDim1 = xStorageShape->GetStorageShape().GetDim(1);
OPS_LOG_D(nodeName, "x dim0 = %ld", xDim0);
OPS_LOG_D(nodeName, "x dim1 = %ld", xDim1);
const gert::StorageShape *expertIdStorageShape = context->GetInputShape(EXPERT_IDS_INDEX);
OPS_CHECK(expertIdStorageShape == nullptr, OPS_LOG_E(nodeName, "expertIdShape is null."), return false);
OPS_CHECK(expertIdStorageShape->GetStorageShape().GetDimNum() != TWO_DIMS,
OPS_LOG_E(nodeName,
"expertIdShape dims must be 2, but current dim num is %lu.",
expertIdStorageShape->GetStorageShape().GetDimNum()),
return false);
OPS_LOG_D(nodeName, "expertId dim0 = %ld", expertIdStorageShape->GetStorageShape().GetDim(0));
OPS_LOG_D(nodeName, "expertId dim1 = %ld", expertIdStorageShape->GetStorageShape().GetDim(1));
const gert::StorageShape *expandXStorageShape = context->GetOutputShape(OUTPUT_EXPAND_X_INDEX);
OPS_CHECK(expandXStorageShape == nullptr, OPS_LOG_E(nodeName, "expandXShape is null."), return false);
OPS_CHECK(expandXStorageShape->GetStorageShape().GetDimNum() != TWO_DIMS,
OPS_LOG_E(nodeName,
"expandXShape dims must be 2, but current dim num is %lu.",
expandXStorageShape->GetStorageShape().GetDimNum()),
return false);
OPS_LOG_D(nodeName, "expandX dim0 = %ld", expandXStorageShape->GetStorageShape().GetDim(0));
OPS_LOG_D(nodeName, "expandX dim1 = %ld", expandXStorageShape->GetStorageShape().GetDim(1));
if (quantMode == DYNAMIC_SCALES) {
const gert::StorageShape *dynamicScalesStorageShape = context->GetOutputShape(OUTPUT_DYNAMIC_SCALES_INDEX);
OPS_CHECK(
dynamicScalesStorageShape == nullptr, OPS_LOG_E(nodeName, "dynamicScalesShape is null."), return false);
OPS_CHECK(dynamicScalesStorageShape->GetStorageShape().GetDimNum() != DYNAMIC_SCALE_DIM_NUM,
OPS_LOG_E(nodeName,
"dynamicScalesShape dims must be %u, but current dim num is %lu.",
DYNAMIC_SCALE_DIM_NUM,
dynamicScalesStorageShape->GetStorageShape().GetDimNum()),
return false);
OPS_LOG_D(nodeName, "dynamicScales dim0 = %ld", dynamicScalesStorageShape->GetStorageShape().GetDim(0));
}
const gert::StorageShape *assistInfoStorageShape = context->GetOutputShape(OUTPUT_ASSIST_INFO_INDEX);
OPS_CHECK(assistInfoStorageShape == nullptr, OPS_LOG_E(nodeName, "assistInfoShape is null."), return false);
OPS_CHECK(assistInfoStorageShape->GetStorageShape().GetDimNum() != ONE_DIM,
OPS_LOG_E(nodeName,
"assistInfoShape dims must be 1, but current dim num is %lu.",
assistInfoStorageShape->GetStorageShape().GetDimNum()),
return false);
OPS_LOG_D(nodeName, "assistInfoForCombine dim0 = %ld", assistInfoStorageShape->GetStorageShape().GetDim(0));
return true;
}
static bool CheckTensorDataType(gert::TilingContext *context, const char *nodeName, const uint32_t quantMode)
{
auto xDesc = context->GetInputDesc(X_INDEX);
OPS_CHECK(xDesc == nullptr, OPS_LOG_E(nodeName, "xDesc is null."), return false);
OPS_CHECK((xDesc->GetDataType() != ge::DT_BF16) && (xDesc->GetDataType() != ge::DT_FLOAT16),
OPS_LOG_E(nodeName, "x dataType is invalid, dataType should be bf16 or float16, but is ."),
return false);
auto expertIdDesc = context->GetInputDesc(EXPERT_IDS_INDEX);
OPS_CHECK(expertIdDesc == nullptr, OPS_LOG_E(nodeName, "expertIdDesc is null."), return false);
OPS_CHECK(expertIdDesc->GetDataType() != ge::DT_INT32,
OPS_LOG_E(nodeName, "expertId dataType is invalid, dataType should be int32, but is ."),
return false);
auto expandXDesc = context->GetOutputDesc(OUTPUT_EXPAND_X_INDEX);
OPS_CHECK(expandXDesc == nullptr, OPS_LOG_E(nodeName, "expandXDesc is null."), return false);
if (quantMode != NO_SCALES) {
OPS_CHECK(expandXDesc->GetDataType() != ge::DT_INT8,
OPS_LOG_E(nodeName, "expandX dataType is invalid, dataType should be int8, but is."),
return false);
} else {
OPS_CHECK(expandXDesc->GetDataType() != xDesc->GetDataType(),
OPS_LOG_E(nodeName, "expandX dataType is invalid, dataType should be equal to x dataType , but is."),
return false);
}
if (quantMode == DYNAMIC_SCALES) {
auto dynamicScalesDesc = context->GetOutputDesc(OUTPUT_DYNAMIC_SCALES_INDEX);
OPS_CHECK(dynamicScalesDesc == nullptr, OPS_LOG_E(nodeName, "dynamicScalesDesc is null."), return false);
OPS_CHECK(dynamicScalesDesc->GetDataType() != ge::DT_FLOAT,
OPS_LOG_E(nodeName, "dynamicScales dataType is invalid, dataType should be float, but is ."),
return false);
}
auto assistInfoDesc = context->GetOutputDesc(OUTPUT_ASSIST_INFO_INDEX);
OPS_CHECK(assistInfoDesc == nullptr, OPS_LOG_E(nodeName, "assistInfoDesc is null."), return false);
OPS_CHECK(assistInfoDesc->GetDataType() != ge::DT_INT32,
OPS_LOG_E(nodeName, "assistInfoForCombine dataType is invalid, dataType should be int32, but is ."),
return false);
return true;
}
static bool CheckTensorFormat(gert::TilingContext *context, const char *nodeName, const uint32_t quantMode)
{
auto xDesc = context->GetInputDesc(X_INDEX);
OPS_CHECK(xDesc == nullptr, OPS_LOG_E(nodeName, "xDesc is null."), return false);
OPS_CHECK(static_cast<ge::Format>(ge::GetPrimaryFormat(xDesc->GetStorageFormat())) == ge::FORMAT_FRACTAL_NZ,
OPS_LOG_E(nodeName, "x format is invalid."),
return false);
auto expertIdDesc = context->GetInputDesc(EXPERT_IDS_INDEX);
OPS_CHECK(expertIdDesc == nullptr, OPS_LOG_E(nodeName, "expertIdDesc is null."), return false);
OPS_CHECK(
static_cast<ge::Format>(ge::GetPrimaryFormat(expertIdDesc->GetStorageFormat())) == ge::FORMAT_FRACTAL_NZ,
OPS_LOG_E(nodeName, "expertId format is invalid."),
return false);
auto expandXDesc = context->GetOutputDesc(OUTPUT_EXPAND_X_INDEX);
OPS_CHECK(expandXDesc == nullptr, OPS_LOG_E(nodeName, "expandXDesc is null."), return false);
OPS_CHECK(
static_cast<ge::Format>(ge::GetPrimaryFormat(expandXDesc->GetStorageFormat())) == ge::FORMAT_FRACTAL_NZ,
OPS_LOG_E(nodeName, "expandX format is invalid."),
return false);
if (quantMode == DYNAMIC_SCALES) {
auto dynamicScalesDesc = context->GetOutputDesc(OUTPUT_DYNAMIC_SCALES_INDEX);
OPS_CHECK(dynamicScalesDesc == nullptr, OPS_LOG_E(nodeName, "dynamicScalesDesc is null."), return false);
OPS_CHECK(static_cast<ge::Format>(ge::GetPrimaryFormat(dynamicScalesDesc->GetStorageFormat())) ==
ge::FORMAT_FRACTAL_NZ,
OPS_LOG_E(nodeName, "dynamicScales format is invalid."),
return false);
}
auto assistInfoDesc = context->GetOutputDesc(OUTPUT_ASSIST_INFO_INDEX);
OPS_CHECK(assistInfoDesc == nullptr, OPS_LOG_E(nodeName, "assistInfoDesc is null."), return false);
OPS_CHECK(
static_cast<ge::Format>(ge::GetPrimaryFormat(assistInfoDesc->GetStorageFormat())) == ge::FORMAT_FRACTAL_NZ,
OPS_LOG_E(nodeName, "assistInfoForCombine format is invalid."),
return false);
return true;
}
static ge::graphStatus GetAttrAndSetTilingData(gert::TilingContext *context, const char *nodeName,
MoeDispatchNormalTilingData &tilingData, std::string &groupEp, std::string &groupTp)
{
auto attrs = context->GetAttrs();
OPS_CHECK(attrs == nullptr, OPS_LOG_E(nodeName, "attrs is nullptr."), return ge::GRAPH_FAILED);
auto groupEpPtr = attrs->GetAttrPointer<char>(static_cast<int>(ATTR_GROUP_EP_INDEX));
auto groupTpPtr = attrs->GetAttrPointer<char>(static_cast<int>(ATTR_GROUP_TP_INDEX));
auto epWorldSizePtr = attrs->GetAttrPointer<int64_t>(ATTR_EP_WORLD_SIZE_INDEX);
auto tpWorldSizePtr = attrs->GetAttrPointer<int64_t>(ATTR_TP_WORLD_SIZE_INDEX);
auto epRankIdPtr = attrs->GetAttrPointer<int64_t>(ATTR_EP_RANK_ID_INDEX);
auto tpRankIdPtr = attrs->GetAttrPointer<int64_t>(ATTR_TP_RANK_ID_INDEX);
auto moeExpertNumPtr = attrs->GetAttrPointer<int64_t>(ATTR_MOE_EXPERT_NUM_INDEX);
auto quantModePtr = attrs->GetAttrPointer<int64_t>(ATTR_QUANT_MODE_INDEX);
// Check for null
OPS_CHECK((groupEpPtr == nullptr) || (strnlen(groupEpPtr, MAX_GROUP_NAME_LENGTH) == 0) ||
(strnlen(groupEpPtr, MAX_GROUP_NAME_LENGTH) == MAX_GROUP_NAME_LENGTH),
OPS_LOG_E(nodeName, "groupEpPtr is null."),
return ge::GRAPH_FAILED);
OPS_CHECK(epWorldSizePtr == nullptr, OPS_LOG_E(nodeName, "epWorldSizePtr is null."), return ge::GRAPH_FAILED);
OPS_CHECK(tpWorldSizePtr == nullptr, OPS_LOG_E(nodeName, "tpWorldSizePtr is null."), return ge::GRAPH_FAILED);
OPS_CHECK(epRankIdPtr == nullptr, OPS_LOG_E(nodeName, "epRankIdPtr is null."), return ge::GRAPH_FAILED);
OPS_CHECK(tpRankIdPtr == nullptr, OPS_LOG_E(nodeName, "tpRankIdPtr is null."), return ge::GRAPH_FAILED);
OPS_CHECK(moeExpertNumPtr == nullptr, OPS_LOG_E(nodeName, "moeExpertNumPtr is null."), return ge::GRAPH_FAILED);
OPS_CHECK(quantModePtr == nullptr, OPS_LOG_E(nodeName, "quantModePtr is null."), return ge::GRAPH_FAILED);
// Check if it meets uint32_t and other constraints
int64_t moeExpertNum = *moeExpertNumPtr;
int64_t epWorldSize = *epWorldSizePtr;
OPS_CHECK((epWorldSize < MIN_EP_WORLD_SIZE) || (epWorldSize > MAX_EP_WORLD_SIZE),
OPS_LOG_E(nodeName,
"epWorldSize is invalid, only support [%ld, %ld], but got epWorldSize=%ld.",
MIN_EP_WORLD_SIZE,
MAX_EP_WORLD_SIZE,
epWorldSize),
return ge::GRAPH_FAILED);
OPS_CHECK((*tpWorldSizePtr < 0) || (*tpWorldSizePtr > MAX_TP_WORLD_SIZE),
OPS_LOG_E(nodeName,
"tpWorldSize is invalid, only support [0, %ld], but got tpWorldSize=%ld.",
MAX_TP_WORLD_SIZE,
*tpWorldSizePtr),
return ge::GRAPH_FAILED);
OPS_CHECK((*epRankIdPtr < 0) || (*epRankIdPtr >= epWorldSize),
OPS_LOG_E(
nodeName, "epRankId is invalid, only support [0, %ld), but got epRankId=%ld.", epWorldSize, *epRankIdPtr),
return ge::GRAPH_FAILED);
if (*tpWorldSizePtr > 1) {
OPS_CHECK((*tpRankIdPtr < 0) || (*tpRankIdPtr >= *tpWorldSizePtr),
OPS_LOG_E(nodeName,
"tpRankId is invalid, only support [0, %ld), but got tpRankId=%ld.",
*tpWorldSizePtr,
*tpRankIdPtr),
return ge::GRAPH_FAILED);
OPS_CHECK((groupTpPtr == nullptr) || (strnlen(groupTpPtr, MAX_GROUP_NAME_LENGTH) == 0) ||
(strnlen(groupTpPtr, MAX_GROUP_NAME_LENGTH) == MAX_GROUP_NAME_LENGTH),
OPS_LOG_E(nodeName, "groupTpPtr is null."),
return ge::GRAPH_FAILED);
groupTp = std::string(groupTpPtr);
} else {
OPS_CHECK(*tpRankIdPtr != 0,
OPS_LOG_E(nodeName, "tpRankId is invalid, NoTp mode only support 0, but got tpRankId=%ld.", *tpRankIdPtr),
return ge::GRAPH_FAILED);
}
OPS_CHECK((moeExpertNum <= 0) || (moeExpertNum > MOE_EXPERT_MAX_NUM),
OPS_LOG_E(nodeName,
"moeExpertNum is invalid, only support (0, %ld], but got moeExpertNum=%ld.",
MOE_EXPERT_MAX_NUM,
moeExpertNum),
return ge::GRAPH_FAILED);
OPS_CHECK(
(*quantModePtr < static_cast<int64_t>(NO_SCALES)) || (*quantModePtr > static_cast<int64_t>(DYNAMIC_SCALES)),
OPS_LOG_E(nodeName,
"quantMode is invalid, only support [0, %u], but got quantMode=%ld.",
DYNAMIC_SCALES,
*quantModePtr),
return ge::GRAPH_FAILED);
int64_t moePerRankNum = moeExpertNum / epWorldSize;
int64_t curDispatchStatusNum = moePerRankNum * epWorldSize;
OPS_CHECK((curDispatchStatusNum > DISPATCH_STATUS_MAX_SUPPORT_NUM),
OPS_LOG_E(nodeName,
"The moe experts num must meet the conditions,"
" (moeExpertNum / epWorldSize * epWorldSize <= 1280, but cur is %ld.",
curDispatchStatusNum),
return ge::GRAPH_FAILED);
groupEp = std::string(groupEpPtr);
tilingData.moeDispatchNormalInfo.epWorldSize = static_cast<uint32_t>(epWorldSize);
tilingData.moeDispatchNormalInfo.tpWorldSize = static_cast<uint32_t>(*tpWorldSizePtr);
tilingData.moeDispatchNormalInfo.epRankId = static_cast<uint32_t>(*epRankIdPtr);
tilingData.moeDispatchNormalInfo.tpRankId = static_cast<uint32_t>(*tpRankIdPtr);
tilingData.moeDispatchNormalInfo.moeExpertNum = static_cast<uint32_t>(moeExpertNum);
tilingData.moeDispatchNormalInfo.quantMode = static_cast<uint32_t>(*quantModePtr);
return ge::GRAPH_SUCCESS;
}
static ge::graphStatus CheckAttrs(
gert::TilingContext *context, const char *nodeName, MoeDispatchNormalTilingData &tilingData, uint32_t &localMoeExpertNum)
{
uint32_t epWorldSize = tilingData.moeDispatchNormalInfo.epWorldSize;
uint32_t tpWorldSize = tilingData.moeDispatchNormalInfo.tpWorldSize;
uint32_t moeExpertNum = tilingData.moeDispatchNormalInfo.moeExpertNum;
// Validate if moe expert number can be evenly distributed across multiple machines
localMoeExpertNum = moeExpertNum / epWorldSize;
OPS_CHECK(moeExpertNum % epWorldSize != 0,
OPS_LOG_E(nodeName,
"moeExpertNum should be divisible by epWorldSize, "
"but moeExpertNum=%u, epWorldSize=%u.",
moeExpertNum,
epWorldSize),
return ge::GRAPH_FAILED);
OPS_CHECK(localMoeExpertNum <= 0,
OPS_LOG_E(nodeName, "localMoeExpertNum is invalid, localMoeExpertNum = %d", localMoeExpertNum),
return ge::GRAPH_FAILED);
// Validate input x dimension 0 and set bs
const gert::StorageShape *xStorageShape = context->GetInputShape(X_INDEX);
const int64_t xDim0 = xStorageShape->GetStorageShape().GetDim(0);
OPS_CHECK((xDim0 > BS_UPPER_BOUND) || (xDim0 <= 0),
OPS_LOG_E(
nodeName, "xDim0(BS) is invalid. Should be between [1, %ld], but got xDim0=%ld.", BS_UPPER_BOUND, xDim0),
return ge::GRAPH_FAILED);
tilingData.moeDispatchNormalInfo.bs = static_cast<uint32_t>(xDim0);
// Validate globalBS
auto attrs = context->GetAttrs();
OPS_CHECK(attrs == nullptr, OPS_LOG_E(nodeName, "attrs is nullptr."), return ge::GRAPH_FAILED);
auto globalBsPtr = attrs->GetAttrPointer<int64_t>(ATTR_GLOBAL_BS_INDEX);
OPS_CHECK(globalBsPtr == nullptr, OPS_LOG_E(nodeName, "globalBsPtr is nullptr."), return ge::GRAPH_FAILED);
OPS_LOG_D(nodeName, "MoeDispatchNormal *globalBsPtr = %ld, bs = %ld, epWorldSize = %u\n", *globalBsPtr, xDim0, epWorldSize);
OPS_CHECK(*globalBsPtr <= 0,
OPS_LOG_E(nodeName,
"globalBS is invalid, should be positive, but got globalBS=%ld.",
*globalBsPtr),
return ge::GRAPH_FAILED);
tilingData.moeDispatchNormalInfo.globalBs = static_cast<uint32_t>(*globalBsPtr);
return ge::GRAPH_SUCCESS;
}
static ge::graphStatus CheckTensorShape(gert::TilingContext *context, const char *nodeName,
MoeDispatchNormalTilingData &tilingData, const uint32_t quantMode, const int64_t localMoeExpertNum)
{
uint32_t A = 0U;
uint32_t globalBs = tilingData.moeDispatchNormalInfo.globalBs;
// Validate input x dimension 1 and set h, bs already validated
const gert::StorageShape *xStorageShape = context->GetInputShape(X_INDEX);
const int64_t xDim0 = xStorageShape->GetStorageShape().GetDim(0);
const int64_t xDim1 = xStorageShape->GetStorageShape().GetDim(1);
OPS_CHECK((xDim1 < H_MIN) || (xDim1 > H_MAX),
OPS_LOG_E(nodeName, "xShape dims1(H) should be in [%ld, %ld], but got %ld.", H_MIN, H_MAX, xDim1),
return ge::GRAPH_FAILED); // 32-byte aligned
tilingData.moeDispatchNormalInfo.h = static_cast<uint32_t>(xDim1);
// Validate expert_id dimensions and set k
int64_t moeExpertNum = static_cast<int64_t>(tilingData.moeDispatchNormalInfo.moeExpertNum);
const gert::StorageShape *expertIdStorageShape = context->GetInputShape(EXPERT_IDS_INDEX);
const int64_t expertIdsDim0 = expertIdStorageShape->GetStorageShape().GetDim(0);
const int64_t expertIdsDim1 = expertIdStorageShape->GetStorageShape().GetDim(1);
OPS_CHECK(xDim0 != expertIdsDim0,
OPS_LOG_E(nodeName,
"xShape's dim0 not equal to expertIdShape's dim0, "
"xShape's dim0 is %ld, expertIdShape's dim0 is %ld.",
xDim0,
expertIdsDim0),
return ge::GRAPH_FAILED);
OPS_CHECK((expertIdsDim1 <= 0) || (expertIdsDim1 > K_MAX) || (expertIdsDim1 > moeExpertNum),
OPS_LOG_E(nodeName,
"expertIdShape's dim1(k) should be in (0, min(%ld, moeExpertNum=%ld)], "
"but got expertIdShape's dim1=%ld.",
K_MAX,
moeExpertNum,
expertIdsDim1),
return ge::GRAPH_FAILED);
tilingData.moeDispatchNormalInfo.k = static_cast<uint32_t>(expertIdsDim1);
A = globalBs;
// Validate expandX dimensions
const gert::StorageShape *expandXStorageShape = context->GetOutputShape(OUTPUT_EXPAND_X_INDEX);
const int64_t expandXDim0 = expandXStorageShape->GetStorageShape().GetDim(0);
const int64_t expandXDim1 = expandXStorageShape->GetStorageShape().GetDim(1);
OPS_CHECK(xDim1 != expandXDim1,
OPS_LOG_E(nodeName,
"expandX's dim1 not equal to xShape's dim1, "
"xShape's dim1 is %ld, expandX's dim1 is %ld.",
xDim1,
expandXDim1),
return ge::GRAPH_FAILED);
// Validate dynamicScales dimensions
if (quantMode != NO_SCALES) {
const gert::StorageShape *dynamicScalesStorageShape = context->GetOutputShape(OUTPUT_DYNAMIC_SCALES_INDEX);
const int64_t dynamicScalesDim0 = dynamicScalesStorageShape->GetStorageShape().GetDim(0);
}
// Validate assistInfo dimensions
const gert::StorageShape *assistInfoStorageShape = context->GetOutputShape(OUTPUT_ASSIST_INFO_INDEX);
const int64_t assistInfoDim0 = assistInfoStorageShape->GetStorageShape().GetDim(0);
return ge::GRAPH_SUCCESS;
}
static ge::graphStatus TilingCheckMoeDispatchNormal(
gert::TilingContext *context, const char *nodeName, const uint32_t quantMode)
{
OPS_CHECK(!CheckTensorDim(context, nodeName, quantMode),
OPS_LOG_E(nodeName, "params shape is invalid."),
return ge::GRAPH_FAILED);
OPS_CHECK(!CheckTensorDataType(context, nodeName, quantMode),
OPS_LOG_E(nodeName, "params dataType is invalid."),
return ge::GRAPH_FAILED);
OPS_CHECK(!CheckTensorFormat(context, nodeName, quantMode),
OPS_LOG_E(nodeName, "params format is invalid."),
return ge::GRAPH_FAILED);
return ge::GRAPH_SUCCESS;
}
static void CalTilingKey(uint64_t &tilingKey, const uint32_t quantMode, const uint32_t tpWorldSize)
{
tilingKey += static_cast<uint64_t>(quantMode);
if (tpWorldSize == TP_WORLD_SIZE_TWO) {
tilingKey += static_cast<uint64_t>(TILINGKEY_TP_WORLD_SIZE);
}
return;
}
static void SetHcommCfg(const gert::TilingContext *context, MoeDispatchNormalTilingData *tiling, const std::string groupEp,
const std::string groupTp)
{
const char *nodeName = context->GetNodeName();
OPS_LOG_D(nodeName, "MoeDispatchNormal groupEp = %s, groupTp = %s", groupEp.c_str(), groupTp.c_str());
uint32_t opType1 = OP_TYPE_ALL_TO_ALL;
uint32_t opType2 = OP_TYPE_ALL_GATHER;
std::string algConfigAllToAllStr = "AlltoAll=level0:fullmesh;level1:pairwise";
std::string algConfigAllGatherStr = "AllGather=level0:ring";
AscendC::Mc2CcTilingConfig mc2CcTilingConfig(groupEp, opType1, algConfigAllToAllStr);
mc2CcTilingConfig.GetTiling(tiling->mc2InitTiling);
mc2CcTilingConfig.GetTiling(tiling->mc2CcTiling1);
mc2CcTilingConfig.SetGroupName(groupTp);
mc2CcTilingConfig.SetOpType(opType2);
mc2CcTilingConfig.SetAlgConfig(algConfigAllGatherStr);
mc2CcTilingConfig.GetTiling(tiling->mc2CcTiling2);
}
static ge::graphStatus SetWorkSpace(gert::TilingContext *context, const char *nodeName)
{
size_t *workSpaces = context->GetWorkspaceSizes(1);
OPS_CHECK(workSpaces == nullptr, OPS_LOG_E(nodeName, "workSpaces is nullptr."), return ge::GRAPH_FAILED);
auto ascendcPlatform = platform_ascendc::PlatformAscendC(context->GetPlatformInfo());
uint32_t aivNum = ascendcPlatform.GetCoreNumAiv();
workSpaces[0] = static_cast<uint64_t>(SYSTEM_NEED_WORKSPACE + WORKSPACE_ELEMENT_OFFSET * aivNum * aivNum);
return ge::GRAPH_SUCCESS;
}
static ge::graphStatus MoeDispatchNormalA3TilingFuncImpl(gert::TilingContext *context)
{
const char *nodeName = context->GetNodeName();
MoeDispatchNormalTilingData *tilingData = context->GetTilingData<MoeDispatchNormalTilingData>();
OPS_CHECK(tilingData == nullptr, OPS_LOG_E(nodeName, "tilingData is nullptr."), return ge::GRAPH_FAILED);
std::string groupEp = "";
std::string groupTp = "";
uint32_t quantMode = NO_SCALES;
uint32_t localMoeExpertNum = 1;
OPS_LOG_I(nodeName, "Enter MoeDispatchNormal tiling check func.");
// Get input parameter attributes
OPS_CHECK(GetAttrAndSetTilingData(context, nodeName, *tilingData, groupEp, groupTp) != ge::GRAPH_SUCCESS,
OPS_LOG_E(nodeName, "Get attr and set tiling data failed."),
return ge::GRAPH_FAILED);
quantMode = tilingData->moeDispatchNormalInfo.quantMode;
// Check input/output dim, format, dataType
OPS_CHECK(TilingCheckMoeDispatchNormal(context, nodeName, quantMode) != ge::GRAPH_SUCCESS,
OPS_LOG_E(nodeName, "Tiling check param failed."),
return ge::GRAPH_FAILED);
// Check if attribute values are valid
OPS_CHECK(CheckAttrs(context, nodeName, *tilingData, localMoeExpertNum) != ge::GRAPH_SUCCESS,
OPS_LOG_E(nodeName, "Check attr failed."),
return ge::GRAPH_FAILED);
uint32_t epRankId = tilingData->moeDispatchNormalInfo.epRankId;
// Check shape dimensions and assign h, k
OPS_CHECK(
CheckTensorShape(context, nodeName, *tilingData, quantMode, static_cast<int64_t>(localMoeExpertNum)) !=
ge::GRAPH_SUCCESS,
OPS_LOG_E(nodeName, "Check tensor shape failed."),
return ge::GRAPH_FAILED);
// Validate win area size
uint64_t maxWindowSize = Mc2TilingUtils::GetMaxWindowSize();
uint64_t h = static_cast<uint64_t>(tilingData->moeDispatchNormalInfo.h);
uint64_t k = static_cast<uint64_t>(tilingData->moeDispatchNormalInfo.k);
uint64_t epWorldSize = static_cast<uint64_t>(tilingData->moeDispatchNormalInfo.epWorldSize);
uint64_t maxBs = static_cast<uint64_t>(tilingData->moeDispatchNormalInfo.globalBs) / epWorldSize;
// Dispatch data area: token start aligned to 512, valid token length h_align_32b + scale(32b) + triplet(3*4b)
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;
// Not considering dual stream size
uint64_t actualSize = maxBs * k * tokenNeedSizeDispatch * DOUBLE_DATA_BUFFER;
OPS_CHECK((actualSize > maxWindowSize),
OPS_LOG_E(nodeName,
"HCCL_BUFFSIZE is too SMALL, maxBs = %lu, h = %lu, epWorldSize = %lu,"
" 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->moeDispatchNormalInfo.totalWinSize = maxWindowSize;
OPS_LOG_D(nodeName, "windowSize = %lu", maxWindowSize);
OPS_CHECK(SetWorkSpace(context, nodeName) != ge::GRAPH_SUCCESS,
OPS_LOG_E(nodeName, "Tiling set workspace failed."),
return ge::GRAPH_FAILED);
SetHcommCfg(context, tilingData, groupEp, groupTp);
uint32_t tpWorldSize = tilingData->moeDispatchNormalInfo.tpWorldSize;
uint64_t tilingKey = INIT_TILINGKEY;
CalTilingKey(tilingKey, quantMode, tpWorldSize);
OPS_LOG_D(nodeName, "tilingKey is %lu", tilingKey);
context->SetTilingKey(tilingKey);
uint32_t blockDim = 1U;
auto ascendcPlatform = platform_ascendc::PlatformAscendC(context->GetPlatformInfo());
uint32_t aivNum = ascendcPlatform.GetCoreNumAiv();
uint64_t ubSize = 0UL;
ascendcPlatform.GetCoreMemSize(platform_ascendc::CoreMemType::UB, ubSize);
blockDim = ascendcPlatform.CalcTschBlockDim(aivNum, 0, aivNum);
context->SetBlockDim(blockDim);
context->SetScheduleMode(1); // Set to batch mode, all cores start simultaneously
tilingData->moeDispatchNormalInfo.totalUbSize = ubSize;
tilingData->moeDispatchNormalInfo.aivNum = aivNum;
OPS_LOG_D(nodeName, "blockDim=%u, aivNum=%u, ubSize=%lu", blockDim, aivNum, ubSize);
PrintTilingDataInfo(nodeName, *tilingData);
return ge::GRAPH_SUCCESS;
}
static ge::graphStatus MoeDispatchNormalTilingFunc(gert::TilingContext *context)
{
ge::graphStatus ret = MoeDispatchNormalA3TilingFuncImpl(context);
return ret;
}
struct MoeDispatchNormalCompileInfo {};
ge::graphStatus TilingParseForMoeDispatchNormal(gert::TilingParseContext *context)
{
(void)context;
return ge::GRAPH_SUCCESS;
}
IMPL_OP_OPTILING(MoeDispatchNormal)
.Tiling(MoeDispatchNormalTilingFunc)
.TilingParse<MoeDispatchNormalCompileInfo>(TilingParseForMoeDispatchNormal);
} // namespace optiling

56
csrc/moe_dispatch_normal/op_kernel/moe_dispatch_normal.cpp Normal file
View 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
View 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

30
csrc/moe_dispatch_normal/op_kernel/moe_dispatch_normal_tiling.h Normal file
View File

@@ -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