mirror of
https://github.com/ROCm/composable_kernel.git
synced 2026-05-14 10:09:41 +00:00
Dev/a8w4 and a8w8splitk (#3447)
* Ck moe bs splitk pr (#3440)
* splitk kick-off. Compilation fail
* splitk hack pass
* fix scale offset calc.
* clang-format for a8w8_moe_blk_gemm1 splitk change
* fix testcase error
---------
Co-authored-by: oscar <huaiguxu@amd.com>
Co-authored-by: huaiguxu <145733371+huaiguxu@users.noreply.github.com>
* Zan/moe a8w4 (#3441)
* update
* update
* update ck moe a8w4
* update
* update
* update
* compile pass
* update
* update
* python3 op_tests/test_moe_2stage.py -t 16 -e 1 -k 1 -dim 256,256 ready
* support new a8w4 kernel
* update
* update ck_tile
* re format
* update
* update
* fix conflict
* fix build
* update ck_tile moe
* fix clang format
* fix the problem
* fix accruacy issue
* fix
---------
Co-authored-by: oscar <huaiguxu@amd.com>
Co-authored-by: huaiguxu <145733371+huaiguxu@users.noreply.github.com>
Co-authored-by: Zzz9990 <zanzhang@amd.com>
Co-authored-by: felix <felix.li@amd.com>
[ROCm/composable_kernel commit: c0ee71d735]
This commit is contained in:
yadaish
@@ -18,6 +18,7 @@ add_example_executable(example_moe_gemm1_xdl_fp8 moe_gemm1_xdl_fp8.cpp)
|
||||
add_example_executable(example_moe_gemm2_xdl_fp8 moe_gemm2_xdl_fp8.cpp)
|
||||
add_example_executable(example_moe_gemm2_xdl_fp8_blockscale moe_gemm2_xdl_fp8_blockscale.cpp)
|
||||
add_example_executable(example_moe_gemm1_xdl_fp8_blockscale moe_gemm1_xdl_fp8_blockscale.cpp)
|
||||
add_example_executable(example_moe_gemm1_xdl_fp8_blockscale_splitk moe_gemm1_xdl_fp8_blockscale_splitk.cpp)
|
||||
|
||||
list(APPEND gpu_list gfx942 gfx950 gfx1100 gfx1101 gfx1102 gfx1103 gfx1150 gfx1151 gfx1152 gfx1153 gfx1200 gfx1201 gfx11-generic gfx12-generic)
|
||||
set(target 0)
|
||||
|
||||
@@ -171,7 +171,7 @@ using DeviceOpInstance = ck::tensor_operation::device::DeviceM
|
||||
// MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector|
|
||||
// PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl|
|
||||
CShuffleMXDLPerWave, CShuffleNXDLPerWave, S<1, 32, 1, 8>, S<EVec, D0Vec, D1Vec, 1>,
|
||||
ck::BlockGemmPipelineScheduler::Intrawave, ck::BlockGemmPipelineVersion::v1, ActOP, Nswizzle, true, MulRoutedWeight, int32_t, A0DataType>;
|
||||
ck::BlockGemmPipelineScheduler::Intrawave, ck::BlockGemmPipelineVersion::v1, ActOP, Nswizzle, true, false, MulRoutedWeight, int32_t, A0DataType>;
|
||||
#else
|
||||
static constexpr ck::index_t MPerBlock = 64; using DeviceOpInstance = ck::tensor_operation::device::DeviceMoeGemmBlockScale<
|
||||
Row, Col, DsLayout, ELayout,
|
||||
@@ -185,7 +185,7 @@ static constexpr ck::index_t MPerBlock = 64; using DeviceOpInstance = ck::tensor
|
||||
S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0,
|
||||
S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0,
|
||||
4, 2, S<1, 32, 1, 8>, S<2, 1, 1, 1>,
|
||||
ck::BlockGemmPipelineScheduler::Intrawave, ck::BlockGemmPipelineVersion::v3, ActOP, Nswizzle, true, MulRoutedWeight, int32_t, A0DataType>;
|
||||
ck::BlockGemmPipelineScheduler::Intrawave, ck::BlockGemmPipelineVersion::v3, ActOP, Nswizzle, true, false, MulRoutedWeight, int32_t, A0DataType>;
|
||||
#endif
|
||||
// clang-format on
|
||||
|
||||
|
||||
@@ -0,0 +1,539 @@
|
||||
// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
|
||||
// SPDX-License-Identifier: MIT
|
||||
|
||||
#include <iostream>
|
||||
#include <numeric>
|
||||
#include <initializer_list>
|
||||
#include <cstdlib>
|
||||
|
||||
#include "ck/ck.hpp"
|
||||
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
|
||||
#include "ck/tensor_operation/gpu/device/impl/device_moe_gemm_blockscale.hpp"
|
||||
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
|
||||
#include "ck/tensor_operation/gpu/element/unary_element_wise_operation.hpp"
|
||||
|
||||
#include "ck/library/utility/device_memory.hpp"
|
||||
#include "ck/library/utility/host_tensor.hpp"
|
||||
#include "ck/library/utility/host_tensor_generator.hpp"
|
||||
#include "ck/library/utility/literals.hpp"
|
||||
#include "ck/library/reference_tensor_operation/cpu/reference_moe_gemm1_blockscale_splitk.hpp"
|
||||
#include "ck/library/utility/check_err.hpp"
|
||||
|
||||
#include "ck/utility/blkgemmpipe_scheduler.hpp"
|
||||
|
||||
using ::ck::DeviceMem;
|
||||
using ::ck::HostTensorDescriptor;
|
||||
using ::ck::Tensor;
|
||||
|
||||
template <ck::index_t... Is>
|
||||
using S = ck::Sequence<Is...>;
|
||||
|
||||
using F16 = ck::half_t;
|
||||
using BF16 = ck::bhalf_t;
|
||||
using F8 = ck::f8_t;
|
||||
using F32 = float;
|
||||
using I64 = int64_t;
|
||||
|
||||
using Row = ck::tensor_layout::gemm::RowMajor;
|
||||
using Col = ck::tensor_layout::gemm::ColumnMajor;
|
||||
using Bypass = ck::tensor_layout::BypassLayoutVerification;
|
||||
|
||||
using A0DataType = F8;
|
||||
using A1DataType = F32;
|
||||
using B0DataType = F8;
|
||||
using B1DataType = F32;
|
||||
using EDataType = F32;
|
||||
using AccDataType = F32;
|
||||
using CShuffleDataType = EDataType;
|
||||
using D2DataType = F32;
|
||||
using DsDataType = ck::Tuple<D2DataType>;
|
||||
|
||||
using A0Layout = Row;
|
||||
using B0Layout = Col;
|
||||
using ELayout = Row;
|
||||
using D0Layout = Row;
|
||||
using D1Layout = Col;
|
||||
using D2Layout = ELayout;
|
||||
using DsLayout = ck::Tuple<D2Layout>;
|
||||
|
||||
struct MulABScaleExpertWeight
|
||||
{
|
||||
template <typename E, typename C, typename D2>
|
||||
__host__ __device__ constexpr void operator()(E& e, const C& c, const D2& d2) const;
|
||||
// for real kernel use
|
||||
template <>
|
||||
__host__ __device__ constexpr void
|
||||
operator()<EDataType, EDataType, float>(EDataType& e, const EDataType& c, const float& d2) const
|
||||
{
|
||||
(void)d2;
|
||||
e = ck::type_convert<EDataType>(c);
|
||||
}
|
||||
};
|
||||
|
||||
void preShuffleBuffer(const B0DataType* src, B0DataType* dst, int N, int K, int NXdl)
|
||||
{
|
||||
int KPack = 16 / sizeof(B0DataType);
|
||||
int NLane = NXdl;
|
||||
int KLane = 64 / NLane;
|
||||
|
||||
int K0 = K / (KLane * KPack);
|
||||
// K -> K0 KLane KPack
|
||||
// N -> N0 NLane
|
||||
// N, K -> N0 K0 KLane NLane KPack
|
||||
int tempk;
|
||||
for(I64 n = 0; n < N; ++n)
|
||||
{
|
||||
for(I64 k = 0; k < K; ++k)
|
||||
{
|
||||
I64 n0 = n / NLane;
|
||||
I64 n1 = n % NLane;
|
||||
|
||||
I64 k0 = k / (KLane * KPack);
|
||||
tempk = k % (KLane * KPack);
|
||||
I64 k1 = tempk / KPack;
|
||||
I64 k2 = tempk % KPack;
|
||||
|
||||
I64 outputIndex = n0 * KPack * NLane * KLane * K0 + k0 * KPack * NLane * KLane +
|
||||
k1 * KPack * NLane + n1 * KPack + k2;
|
||||
|
||||
dst[outputIndex] = src[n * static_cast<I64>(K) + k];
|
||||
}
|
||||
}
|
||||
}
|
||||
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
|
||||
|
||||
using AElementOp = PassThrough;
|
||||
using BElementOp = PassThrough;
|
||||
using CDEElementOp = MulABScaleExpertWeight;
|
||||
|
||||
static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::Default;
|
||||
|
||||
static constexpr ck::index_t Scale_Block_M = 1;
|
||||
static constexpr ck::index_t Scale_Block_N = 128;
|
||||
static constexpr ck::index_t Scale_Block_K = 128;
|
||||
|
||||
static constexpr ck::index_t Nswizzle = false;
|
||||
static constexpr ck::index_t IsInputGemm = true; // splitk gemm1 goes to gemm2 pipeline.
|
||||
static constexpr ck::index_t IsSplitK = true; // splitk gemm1
|
||||
static constexpr ck::index_t ActOP = 0; // 0: gelu_and_mul, 1: silu_and_mul
|
||||
static constexpr bool MulRoutedWeight = false; // splitk gemm1 does not do routedWeight.
|
||||
|
||||
#if 1
|
||||
static constexpr ck::index_t MPerBlock = 32;
|
||||
static constexpr ck::index_t NPerBlock = 128;
|
||||
static constexpr ck::index_t MNPerXDL = 16;
|
||||
static constexpr ck::index_t MXDLPerWave = MPerBlock / (MNPerXDL * 1);
|
||||
static constexpr ck::index_t NXDLPerWave = NPerBlock / (MNPerXDL * 4);
|
||||
static constexpr ck::index_t CShuffleMXDLPerWave = MXDLPerWave;
|
||||
static constexpr ck::index_t CShuffleNXDLPerWave = NXDLPerWave;
|
||||
static constexpr ck::index_t BLOCKSIZE = 256;
|
||||
|
||||
static constexpr ck::index_t KPerBlock = 128 / sizeof(A0DataType);
|
||||
static constexpr ck::index_t AK1 = 16 / sizeof(A0DataType);
|
||||
static constexpr ck::index_t BK1 = 16 / sizeof(B0DataType);
|
||||
static constexpr ck::index_t EVec = 16 / sizeof(EDataType);
|
||||
static constexpr ck::index_t D0Vec = 1;
|
||||
static constexpr ck::index_t D1Vec = 1;
|
||||
|
||||
using DeviceOpInstance = ck::tensor_operation::device::DeviceMoeGemmBlockScale
|
||||
// clang-format off
|
||||
< Row, Col, DsLayout, ELayout,
|
||||
A0DataType, A1DataType, B0DataType, B1DataType, DsDataType, EDataType, AccDataType, CShuffleDataType,
|
||||
AElementOp, BElementOp, CDEElementOp, GemmSpec,
|
||||
//threadnum, mblock, nblock, kblock
|
||||
BLOCKSIZE, Scale_Block_M, Scale_Block_N, Scale_Block_K,
|
||||
MPerBlock, NPerBlock, KPerBlock,
|
||||
// ak1, bk1
|
||||
AK1, BK1,
|
||||
// mn_perxdl
|
||||
MNPerXDL, MNPerXDL,
|
||||
// mn_xdlperwave
|
||||
MXDLPerWave, NXDLPerWave,
|
||||
// a,b: loadtranfer cluster, cluster order, srcorder,VECDIM, srcpervec, dstpervec, lds_extra
|
||||
S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, AK1, AK1, 0,
|
||||
S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, BK1, BK1, 0,
|
||||
// CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
|
||||
// MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector|
|
||||
// PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl|
|
||||
CShuffleMXDLPerWave, CShuffleNXDLPerWave, S<1, 32, 1, 8>, S<EVec, D0Vec, D1Vec, 1>,
|
||||
ck::BlockGemmPipelineScheduler::Intrawave, ck::BlockGemmPipelineVersion::v1, ActOP, Nswizzle, IsInputGemm, IsSplitK, MulRoutedWeight, int32_t, A0DataType>;
|
||||
#else
|
||||
|
||||
static constexpr ck::index_t MPerBlock = 64; using DeviceOpInstance = ck::tensor_operation::device::DeviceMoeGemmBlockScale<
|
||||
Row, Col, DsLayout, ELayout,
|
||||
A0DataType, A1DataType, B0DataType, B1DataType, DsDataType, EDataType, AccDataType, CShuffleDataType,
|
||||
AElementOp, BElementOp, CDEElementOp, GemmSpec,
|
||||
256, Scale_Block_M, Scale_Block_N, Scale_Block_K,
|
||||
MPerBlock, 128, 128,
|
||||
16, 16,
|
||||
16, 16,
|
||||
4, 2,
|
||||
S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0,
|
||||
S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0,
|
||||
4, 2, S<1, 32, 1, 8>, S<2, 1, 1, 1>,
|
||||
ck::BlockGemmPipelineScheduler::Intrawave, ck::BlockGemmPipelineVersion::v3, ActOP, Nswizzle, IsInputGemm, IsSplitK, MulRoutedWeight, int32_t, A0DataType>;
|
||||
#endif
|
||||
// clang-format on
|
||||
|
||||
int main(int argc, char* argv[])
|
||||
{
|
||||
bool do_verification = true;
|
||||
int init_method = 1;
|
||||
bool time_kernel = true;
|
||||
#if 1
|
||||
// GEMM shape
|
||||
ck::index_t N = 4096;
|
||||
ck::index_t K = 6144;
|
||||
// ck::index_t N = 128;
|
||||
// ck::index_t K = 512;
|
||||
ck::index_t experts = 8;
|
||||
ck::index_t topk = 2;
|
||||
// ck::index_t sorted_tile_num = 515;
|
||||
// ck::index_t valid_tile_num = 512;
|
||||
// ck::index_t tokens = 208;
|
||||
// ck::index_t sorted_tile_num = 15;
|
||||
// ck::index_t valid_tile_num = 13;
|
||||
// ck::index_t sorted_tile_num = 259;
|
||||
// ck::index_t valid_tile_num = 256;
|
||||
// ck::index_t tokens = 4096;
|
||||
ck::index_t sorted_tile_num = 2;
|
||||
ck::index_t valid_tile_num = 2;
|
||||
ck::index_t tokens = 32;
|
||||
#else
|
||||
// deepseek
|
||||
ck::index_t N = 2048;
|
||||
ck::index_t K = 7168;
|
||||
ck::index_t experts = 256;
|
||||
ck::index_t topk = 8;
|
||||
ck::index_t tokens = 4096;
|
||||
ck::index_t sorted_tile_num = 261;
|
||||
ck::index_t valid_tile_num = 256;
|
||||
#endif
|
||||
ck::index_t KBatch = 6;
|
||||
if(argc == 1)
|
||||
{
|
||||
// use default case
|
||||
}
|
||||
else if(argc == 2)
|
||||
{
|
||||
KBatch = std::stoi(argv[1]);
|
||||
}
|
||||
else if(argc == 4)
|
||||
{
|
||||
// use default case
|
||||
do_verification = std::stoi(argv[1]);
|
||||
init_method = std::stoi(argv[2]);
|
||||
time_kernel = std::stoi(argv[3]);
|
||||
}
|
||||
else if(argc == 7)
|
||||
{
|
||||
do_verification = std::stoi(argv[1]);
|
||||
init_method = std::stoi(argv[2]);
|
||||
time_kernel = std::stoi(argv[3]);
|
||||
N = std::stoi(argv[4]);
|
||||
K = std::stoi(argv[5]);
|
||||
tokens = std::stoi(argv[6]);
|
||||
}
|
||||
else if(argc == 9)
|
||||
{
|
||||
|
||||
do_verification = std::stoi(argv[1]);
|
||||
init_method = std::stoi(argv[2]);
|
||||
time_kernel = std::stoi(argv[3]);
|
||||
N = std::stoi(argv[4]);
|
||||
K = std::stoi(argv[5]);
|
||||
tokens = std::stoi(argv[6]);
|
||||
sorted_tile_num = std::stoi(argv[7]);
|
||||
valid_tile_num = std::stoi(argv[8]);
|
||||
}
|
||||
else
|
||||
{
|
||||
printf("arg1: verification (0=no, 1=yes)\n");
|
||||
printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
|
||||
printf("arg3: time kernel (0=no, 1=yes)\n");
|
||||
printf("arg4 to 6: N, K, tokens\n");
|
||||
exit(0);
|
||||
}
|
||||
|
||||
ck::index_t sorted_size = sorted_tile_num * MPerBlock;
|
||||
ck::index_t valid_size = valid_tile_num * MPerBlock;
|
||||
if(tokens * topk > valid_size)
|
||||
{
|
||||
printf("err config, tokens * topk > valid_size\n");
|
||||
exit(-1);
|
||||
}
|
||||
ck::index_t StrideA = K;
|
||||
ck::index_t StrideB = K;
|
||||
ck::index_t StrideE = N * 2;
|
||||
constexpr ck::index_t NumDTensor = DsDataType::Size();
|
||||
constexpr auto StrideDs = std::array<ck::index_t, NumDTensor>{0};
|
||||
ck::index_t Scale_Stride_AM = (K + Scale_Block_K - 1) / Scale_Block_K;
|
||||
ck::index_t Scale_Stride_BN = (K + Scale_Block_K - 1) / Scale_Block_K;
|
||||
ck::index_t Scale_Stride_B = (N + Scale_Block_N - 1) / Scale_Block_N * 2;
|
||||
|
||||
Tensor<ck::index_t> expert_ids(HostTensorDescriptor({sorted_tile_num}, {1}));
|
||||
Tensor<ck::index_t> sorted_token_ids(HostTensorDescriptor({sorted_size}, {1}));
|
||||
Tensor<ck::index_t> max_token_id(HostTensorDescriptor({1 + sorted_tile_num}));
|
||||
max_token_id.mData = {valid_size};
|
||||
// int eids[] = {0, 0, 1, 2, 3, 3, 4, 4, 5, 5, 6, 6, 7, 3, 3, 3};
|
||||
for(int i = 0; i < sorted_tile_num; i++)
|
||||
{
|
||||
expert_ids.mData[i] = i / ck::math::integer_divide_ceil(valid_tile_num, experts);
|
||||
}
|
||||
|
||||
int token_per_tile = (tokens * topk + valid_tile_num - 1) / valid_tile_num;
|
||||
int tokenid = 0;
|
||||
|
||||
for(int i = 0; i < sorted_size; i++)
|
||||
{
|
||||
int tile_off = i % MPerBlock;
|
||||
if(tile_off < token_per_tile && tokenid < tokens * topk)
|
||||
{
|
||||
sorted_token_ids.mData[i] = (tokenid % tokens) | ((tokenid / tokens) << 24);
|
||||
tokenid++;
|
||||
}
|
||||
else
|
||||
{
|
||||
sorted_token_ids.mData[i] = tokens;
|
||||
}
|
||||
}
|
||||
Tensor<A0DataType> a0_t_k(HostTensorDescriptor({tokens, K}, {K, 1}));
|
||||
Tensor<A1DataType> a1_t_k(HostTensorDescriptor(
|
||||
{tokens, (K + Scale_Block_K - 1) / Scale_Block_K}, {Scale_Stride_AM, 1}, Row{}));
|
||||
Tensor<B0DataType> b0_e_n_k(
|
||||
HostTensorDescriptor({experts, K, N * 2}, {N * 2 * K, 1, K}, Col{}));
|
||||
Tensor<B1DataType> b1_e_n_k(
|
||||
HostTensorDescriptor({experts,
|
||||
(K + Scale_Block_K - 1) / Scale_Block_K,
|
||||
(N + Scale_Block_N - 1) / Scale_Block_N * 2},
|
||||
{(Scale_Stride_B * Scale_Stride_BN), 1, Scale_Stride_BN},
|
||||
Col{}));
|
||||
Tensor<B0DataType> b0_preshuffled(
|
||||
HostTensorDescriptor({experts, K, N * 2}, {N * 2 * K, 1, K}, Col{}));
|
||||
Tensor<EDataType> e_t_n_host_result(
|
||||
HostTensorDescriptor({tokens, topk, N * 2}, {topk * N * 2, N * 2, 1}, Row{}));
|
||||
Tensor<EDataType> e_t_n_device_result(
|
||||
HostTensorDescriptor({tokens, topk, N * 2}, {topk * N * 2, N * 2, 1}, Row{}));
|
||||
e_t_n_device_result.SetZero();
|
||||
std::cout << "a0_t_k: " << a0_t_k.mDesc << std::endl;
|
||||
std::cout << "a1_t_k: " << a1_t_k.mDesc << std::endl;
|
||||
std::cout << "b0_e_n_k: " << b0_e_n_k.mDesc << std::endl;
|
||||
std::cout << "b1_e_n_k: " << b1_e_n_k.mDesc << std::endl;
|
||||
std::cout << "e_t_n: " << e_t_n_host_result.mDesc << std::endl;
|
||||
std::cout << "k_batch:" << KBatch << std::endl;
|
||||
std::cout << "init_method:" << init_method << std::endl;
|
||||
|
||||
switch(init_method)
|
||||
{
|
||||
case 0: break;
|
||||
case 1:
|
||||
a0_t_k.GenerateTensorValue(GeneratorTensor_3<A0DataType>{-1.0, 1.0});
|
||||
a1_t_k.GenerateTensorValue(GeneratorTensor_3<A1DataType>{0.0, 1.0});
|
||||
b0_e_n_k.GenerateTensorValue(GeneratorTensor_3<B0DataType>{-1.0, 1.0});
|
||||
b1_e_n_k.GenerateTensorValue(GeneratorTensor_3<B1DataType>{0.0, 1.0});
|
||||
break;
|
||||
case 2:
|
||||
a0_t_k.GenerateTensorValue(GeneratorTensor_1<A0DataType>{});
|
||||
a1_t_k.GenerateTensorValue(GeneratorTensor_1<A1DataType>{});
|
||||
b0_e_n_k.GenerateTensorValue(GeneratorTensor_1<B0DataType>{});
|
||||
b1_e_n_k.GenerateTensorValue(GeneratorTensor_1<B1DataType>{});
|
||||
break;
|
||||
case 3:
|
||||
a0_t_k.GenerateTensorValue(GeneratorTensor_1<A0DataType>{});
|
||||
a1_t_k.GenerateTensorValue(GeneratorTensor_3<A1DataType>{0.0, 1.0});
|
||||
b0_e_n_k.GenerateTensorValue(GeneratorTensor_3<B0DataType>{-0.5, 0.5});
|
||||
b1_e_n_k.GenerateTensorValue(GeneratorTensor_3<B1DataType>{0, 1.0});
|
||||
break;
|
||||
case 4:
|
||||
a0_t_k.GenerateTensorValue(GeneratorTensor_3<A0DataType>{-0.5, 0.5});
|
||||
a1_t_k.GenerateTensorValue(GeneratorTensor_3<A1DataType>{0.0, 1.0});
|
||||
b0_e_n_k.GenerateTensorValue(GeneratorTensor_1<B0DataType>{});
|
||||
b1_e_n_k.GenerateTensorValue(GeneratorTensor_3<B1DataType>{0, 1.0});
|
||||
break;
|
||||
case 5:
|
||||
a0_t_k.GenerateTensorValue(GeneratorTensor_3<A0DataType>{-0.5, 0.5});
|
||||
a1_t_k.GenerateTensorValue(GeneratorTensor_1<A1DataType>{});
|
||||
b0_e_n_k.GenerateTensorValue(GeneratorTensor_3<B0DataType>{-0.5, 0.5});
|
||||
b1_e_n_k.GenerateTensorValue(GeneratorTensor_3<B1DataType>{0, 1.0});
|
||||
break;
|
||||
case 6:
|
||||
a0_t_k.GenerateTensorValue(GeneratorTensor_3<A0DataType>{-0.5, 0.5});
|
||||
a1_t_k.GenerateTensorValue(GeneratorTensor_3<A1DataType>{0.0, 1.0});
|
||||
b0_e_n_k.GenerateTensorValue(GeneratorTensor_3<B0DataType>{-0.5, 0.5});
|
||||
b1_e_n_k.GenerateTensorValue(GeneratorTensor_1<B1DataType>{});
|
||||
break;
|
||||
default:
|
||||
a0_t_k.GenerateTensorValue(GeneratorTensor_3<A0DataType>{-0.5, 0.5});
|
||||
a1_t_k.GenerateTensorValue(GeneratorTensor_3<A1DataType>{0.0, 1.0});
|
||||
b0_e_n_k.GenerateTensorValue(GeneratorTensor_3<B0DataType>{-0.5, 0.5});
|
||||
b1_e_n_k.GenerateTensorValue(GeneratorTensor_3<B1DataType>{0, 1.0});
|
||||
}
|
||||
DeviceMem sorted_token_ids_dev(sizeof(ck::index_t) *
|
||||
sorted_token_ids.mDesc.GetElementSpaceSize());
|
||||
DeviceMem expert_ids_dev(sizeof(ck::index_t) * expert_ids.mDesc.GetElementSpaceSize());
|
||||
DeviceMem max_token_id_dev(sizeof(ck::index_t) * max_token_id.mDesc.GetElementSpaceSize());
|
||||
DeviceMem a0_device_buf(sizeof(A0DataType) * a0_t_k.mDesc.GetElementSpaceSize());
|
||||
DeviceMem a1_device_buf(sizeof(A1DataType) * a1_t_k.mDesc.GetElementSpaceSize());
|
||||
DeviceMem b0_device_buf(sizeof(B0DataType) * b0_e_n_k.mDesc.GetElementSpaceSize());
|
||||
DeviceMem b1_device_buf(sizeof(B1DataType) * b1_e_n_k.mDesc.GetElementSpaceSize());
|
||||
DeviceMem e_device_buf(sizeof(EDataType) * e_t_n_device_result.mDesc.GetElementSpaceSize());
|
||||
|
||||
sorted_token_ids_dev.ToDevice(sorted_token_ids.mData.data());
|
||||
expert_ids_dev.ToDevice(expert_ids.mData.data());
|
||||
max_token_id_dev.ToDevice(max_token_id.mData.data());
|
||||
a0_device_buf.ToDevice(a0_t_k.mData.data());
|
||||
a1_device_buf.ToDevice(a1_t_k.mData.data());
|
||||
b1_device_buf.ToDevice(b1_e_n_k.mData.data());
|
||||
e_device_buf.ToDevice(e_t_n_device_result.mData.data());
|
||||
|
||||
auto a_element_op = AElementOp{};
|
||||
auto b_element_op = BElementOp{};
|
||||
auto cde_element_op = CDEElementOp{};
|
||||
|
||||
// do GEMM
|
||||
auto device_op = DeviceOpInstance{};
|
||||
|
||||
int NPerXdl = device_op.GetPreShuffleParameters();
|
||||
|
||||
preShuffleBuffer(
|
||||
b0_e_n_k.mData.data(), b0_preshuffled.mData.data(), N * 2 * experts, K, NPerXdl);
|
||||
|
||||
b0_device_buf.ToDevice(b0_preshuffled.mData.data());
|
||||
|
||||
auto invoker = device_op.MakeInvoker();
|
||||
auto argument = device_op.MakeArgument(sorted_token_ids_dev.GetDeviceBuffer(),
|
||||
expert_ids_dev.GetDeviceBuffer(),
|
||||
max_token_id_dev.GetDeviceBuffer(),
|
||||
a0_device_buf.GetDeviceBuffer(),
|
||||
b0_device_buf.GetDeviceBuffer(),
|
||||
std::array<const void*, NumDTensor>{nullptr},
|
||||
e_device_buf.GetDeviceBuffer(),
|
||||
tokens,
|
||||
topk,
|
||||
sorted_size,
|
||||
N,
|
||||
K,
|
||||
StrideA,
|
||||
StrideB,
|
||||
StrideDs,
|
||||
StrideE,
|
||||
a1_device_buf.GetDeviceBuffer(),
|
||||
b1_device_buf.GetDeviceBuffer(),
|
||||
KBatch,
|
||||
a_element_op,
|
||||
b_element_op,
|
||||
cde_element_op);
|
||||
|
||||
if(!device_op.IsSupportedArgument(argument))
|
||||
{
|
||||
throw std::runtime_error(
|
||||
"wrong! device_gemm with the specified compilation parameters does "
|
||||
"not support this GEMM problem");
|
||||
}
|
||||
if(time_kernel)
|
||||
{
|
||||
float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
|
||||
|
||||
std::size_t flop = std::size_t(2) * tokens * topk * N * 2 * K;
|
||||
std::size_t num_btype = sizeof(A0DataType) * valid_tile_num * K +
|
||||
sizeof(B0DataType) * K * N * 2 * experts +
|
||||
sizeof(EDataType) * valid_tile_num * N;
|
||||
|
||||
float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
|
||||
|
||||
float gb_per_sec = num_btype / 1.E6 / ave_time;
|
||||
|
||||
std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec
|
||||
<< " GB/s.\n"
|
||||
<< device_op.GetTypeString() << std::endl;
|
||||
}
|
||||
|
||||
if(do_verification)
|
||||
{
|
||||
// use atomic, so need to reinit outputs
|
||||
e_device_buf.ToDevice(e_t_n_device_result.mData.data());
|
||||
invoker.Run(argument, StreamConfig{nullptr, false, 0, 0, 1});
|
||||
|
||||
Tensor<float> a_t_k({tokens, K});
|
||||
Tensor<float> b_e_n_k({experts, K, N * 2});
|
||||
e_device_buf.FromDevice(e_t_n_device_result.mData.data());
|
||||
|
||||
Tensor<float> c_t_k_n({tokens, topk, N * 2}, {topk * N * 2, N * 2, 1}, Row{});
|
||||
|
||||
// handle scale before ref.
|
||||
for(int t = 0; t < tokens; ++t)
|
||||
{
|
||||
for(int k = 0; k < K; ++k)
|
||||
{
|
||||
a_t_k(t, k) = ck::type_convert<float>(a0_t_k(t, k)) * a1_t_k(t, k / Scale_Block_K);
|
||||
}
|
||||
}
|
||||
|
||||
for(int e = 0; e < experts; ++e)
|
||||
{
|
||||
for(int k = 0; k < K; ++k)
|
||||
{
|
||||
for(int n = 0; n < N * 2; ++n)
|
||||
{
|
||||
b_e_n_k(e, k, n) = ck::type_convert<float>(b0_e_n_k(e, k, n)) *
|
||||
b1_e_n_k(e, k / Scale_Block_K, n / Scale_Block_N);
|
||||
}
|
||||
}
|
||||
}
|
||||
using ReferenceGemmInstance =
|
||||
ck::tensor_operation::host::ReferenceMoeGemm1BlockScaleSplitK<float,
|
||||
float,
|
||||
float,
|
||||
AccDataType,
|
||||
PassThrough,
|
||||
PassThrough,
|
||||
PassThrough>;
|
||||
auto ref_moe_gemm = ReferenceGemmInstance{};
|
||||
auto ref_invoker = ref_moe_gemm.MakeInvoker();
|
||||
|
||||
auto ref_argument = ref_moe_gemm.MakeArgument(sorted_token_ids,
|
||||
expert_ids,
|
||||
max_token_id,
|
||||
MPerBlock,
|
||||
a_t_k,
|
||||
b_e_n_k,
|
||||
c_t_k_n,
|
||||
PassThrough{},
|
||||
PassThrough{},
|
||||
PassThrough{});
|
||||
|
||||
ref_invoker.Run(ref_argument);
|
||||
for(int m = 0; m < valid_size; ++m)
|
||||
{
|
||||
|
||||
const int fuse_t = sorted_token_ids.mData[m];
|
||||
const int t = fuse_t & 0xffffff;
|
||||
const int topk_id = (fuse_t & 0xff000000) >> 24;
|
||||
|
||||
if(t >= tokens)
|
||||
{
|
||||
continue;
|
||||
}
|
||||
for(int n = 0; n < 2 * N; ++n)
|
||||
{
|
||||
e_t_n_host_result(t, topk_id, n) =
|
||||
ck::type_convert<EDataType>(c_t_k_n(t, topk_id, n));
|
||||
}
|
||||
}
|
||||
|
||||
e_device_buf.FromDevice(e_t_n_device_result.mData.data());
|
||||
|
||||
auto status =
|
||||
ck::utils::check_err(
|
||||
e_t_n_device_result, e_t_n_host_result, "Error: Incorrect results!", 1e-3, 5e-1)
|
||||
? 0
|
||||
: 1;
|
||||
if(status == 0)
|
||||
{
|
||||
printf("Validation Pass.\n");
|
||||
}
|
||||
return status;
|
||||
}
|
||||
|
||||
return 0;
|
||||
}
|
||||
@@ -165,7 +165,7 @@ using DeviceOpInstance = ck::tensor_operation::device::DeviceMoeGemmBlockScale<
|
||||
S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, AK1, AK1, 0,
|
||||
S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, AK1, AK1, 0,
|
||||
2, 2, S<1, CShuffleMLane, 1, CShuffleNLane>, S<EVec, D0Vec, D1Vec, D2Vec>,
|
||||
ck::BlockGemmPipelineScheduler::Intrawave, ck::BlockGemmPipelineVersion::v1, 0, false, false, MulRoutedWeight, int32_t, A0DataType>;
|
||||
ck::BlockGemmPipelineScheduler::Intrawave, ck::BlockGemmPipelineVersion::v1, 0, false, false, false, MulRoutedWeight, int32_t, A0DataType>;
|
||||
|
||||
#else
|
||||
static constexpr ck::index_t MPerBlock = 64; using DeviceOpInstance = ck::tensor_operation::device::DeviceMoeGemmBlockScale<
|
||||
@@ -180,7 +180,7 @@ static constexpr ck::index_t MPerBlock = 64; using DeviceOpInstance = ck::tensor
|
||||
S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0,
|
||||
S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0,
|
||||
2, 2, S<1, 32, 1, 8>, S<2, 1, 1, 1>,
|
||||
ck::BlockGemmPipelineScheduler::Intrawave, ck::BlockGemmPipelineVersion::v3, 0, false, false, MulRoutedWeight, int32_t, A0DataType>;
|
||||
ck::BlockGemmPipelineScheduler::Intrawave, ck::BlockGemmPipelineVersion::v3, 0, false, false, false, MulRoutedWeight, int32_t, A0DataType>;
|
||||
#endif
|
||||
// clang-format on
|
||||
|
||||
|
||||
Reference in New Issue
Block a user