Skip to content

GitLab

Commit 0ef27d53 authored by Andriy Roshchenko's avatar Andriy Roshchenko
Browse files

Merge remote-tracking branch 'origin/gfx950' into andriy/lwpck-2682

parents 6778c318 74a743e2
Hide whitespace changes
Inline Side-by-side
Showing with 317 additions and 189 deletions
example/ck_tile/12_smoothquant/example_smoothquant.cpp
View file @ 0ef27d53
......@@ -63,17 +63,17 @@ bool run(const ck_tile::ArgParser& arg_parser)
int warmup = arg_parser.get_int("warmup");
int repeat = arg_parser.get_int("repeat");
assert(stride >= n);
assert(x_stride >= n);
using XDataType = DataType;
using XScaleDataType = float;
using YScaleDataType = float;
using QYDataType = ck_tile::int8_t;
using ComputeDataType = float;
using XDataType = DataType;
using SmoothScaleDataType = float;
using YScaleDataType = float;
using QYDataType = ck_tile::int8_t;
using ComputeDataType = float;
// host verify
ck_tile::HostTensor<XDataType> x_host({m, n}, {x_stride, 1});
ck_tile::HostTensor<XScaleDataType> xscale_host({n});
ck_tile::HostTensor<SmoothScaleDataType> smscale_host({n});
ck_tile::HostTensor<YScaleDataType> yscale_host_ref({m}, {1});
ck_tile::HostTensor<YScaleDataType> yscale_host_dev({m}, {1});
......@@ -82,15 +82,15 @@ bool run(const ck_tile::ArgParser& arg_parser)
ck_tile::HostTensor<QYDataType> qy_host_dev({m, n}, {y_stride, 1});
ck_tile::FillUniformDistribution<XDataType>{-.5f, .5f}(x_host);
ck_tile::FillUniformDistribution<XScaleDataType>{1e-3, .5f}(xscale_host);
ck_tile::FillUniformDistribution<SmoothScaleDataType>{1e-3, .5f}(smscale_host);
ck_tile::DeviceMem x_buf(x_host.get_element_space_size_in_bytes());
ck_tile::DeviceMem xscale_buf(xscale_host.get_element_space_size_in_bytes());
ck_tile::DeviceMem smscale_buf(smscale_host.get_element_space_size_in_bytes());
ck_tile::DeviceMem yscale_buf(yscale_host_dev.get_element_space_size_in_bytes());
ck_tile::DeviceMem qy_buf(qy_host_dev.get_element_space_size_in_bytes());
x_buf.ToDevice(x_host.data());
xscale_buf.ToDevice(xscale_host.data());
smscale_buf.ToDevice(smscale_host.data());
constexpr bool kTwoPass = true;
......@@ -101,7 +101,7 @@ bool run(const ck_tile::ArgParser& arg_parser)
using Shape = ck_tile::Generic2dBlockShape<BlockTile, BlockWarps, WarpTile, Vector>;
using Problem = ck_tile::SmoothquantPipelineProblem<XDataType,
XScaleDataType,
SmoothScaleDataType,
ComputeDataType,
YScaleDataType,
QYDataType,
......@@ -115,7 +115,7 @@ bool run(const ck_tile::ArgParser& arg_parser)
using Kernel = ck_tile::Smoothquant<Pipeline>;
ck_tile::SmoothquantHostArgs args{x_buf.GetDeviceBuffer(),
xscale_buf.GetDeviceBuffer(),
smscale_buf.GetDeviceBuffer(),
yscale_buf.GetDeviceBuffer(),
qy_buf.GetDeviceBuffer(),
m,
......@@ -142,16 +142,16 @@ bool run(const ck_tile::ArgParser& arg_parser)
// smooth outlier
{
auto f = [&](auto n_) {
auto v_xscale = ck_tile::type_convert<ComputeDataType>(xscale_host(n_));
auto v_smscale = ck_tile::type_convert<ComputeDataType>(smscale_host(n_));
for(int m_ = 0; m_ < m; ++m_)
{
auto v_x = ck_tile::type_convert<ComputeDataType>(x_host(m_, n_));
y_host(m_, n_) = v_x * v_xscale;
y_host(m_, n_) = v_x * v_smscale;
}
};
ck_tile::make_ParallelTensorFunctor(f, xscale_host.get_element_space_size())(
ck_tile::make_ParallelTensorFunctor(f, smscale_host.get_element_space_size())(
std::thread::hardware_concurrency());
}
......
example/ck_tile/12_smoothquant/instances/smoothquant_instance_common.hpp
View file @ 0ef27d53
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
// Copyright (c) 2018-2025, Advanced Micro Devices, Inc. All rights reserved.
#include <ck_tile/core.hpp>
#include "smoothquant.hpp"
......@@ -35,7 +35,7 @@ float smoothquant_(const S& s, A a)
using PipelineProblem = ck_tile::SmoothquantPipelineProblem<
typename SmoothquantTypeConfig<DataType>::XDataType,
typename SmoothquantTypeConfig<DataType>::XScaleDataType,
typename SmoothquantTypeConfig<DataType>::SmoothScaleDataType,
typename SmoothquantTypeConfig<DataType>::ComputeDataType,
typename SmoothquantTypeConfig<DataType>::YScaleDataType,
typename SmoothquantTypeConfig<DataType>::QYDataType,
......
example/ck_tile/12_smoothquant/smoothquant.cpp
View file @ 0ef27d53
......@@ -66,15 +66,15 @@ bool run(const ck_tile::ArgParser& arg_parser)
using TypeConfig = SmoothquantTypeConfig<DataType>;
using XDataType = typename TypeConfig::XDataType;
using XScaleDataType = typename TypeConfig::XScaleDataType;
using YScaleDataType = typename TypeConfig::YScaleDataType;
using QYDataType = typename TypeConfig::QYDataType;
using ComputeDataType = typename TypeConfig::ComputeDataType;
using XDataType = typename TypeConfig::XDataType;
using SmoothScaleDataType = typename TypeConfig::SmoothScaleDataType;
using YScaleDataType = typename TypeConfig::YScaleDataType;
using QYDataType = typename TypeConfig::QYDataType;
using ComputeDataType = typename TypeConfig::ComputeDataType;
// host verify
ck_tile::HostTensor<XDataType> x_host({m, n}, {x_stride, 1});
ck_tile::HostTensor<XScaleDataType> xscale_host({n});
ck_tile::HostTensor<SmoothScaleDataType> smscale_host({n});
ck_tile::HostTensor<YScaleDataType> yscale_host_ref({m}, {1});
ck_tile::HostTensor<YScaleDataType> yscale_host_dev({m}, {1});
......@@ -83,15 +83,15 @@ bool run(const ck_tile::ArgParser& arg_parser)
ck_tile::HostTensor<QYDataType> qy_host_dev({m, n}, {y_stride, 1});
ck_tile::FillUniformDistribution<XDataType>{-.5f, .5f}(x_host);
ck_tile::FillUniformDistribution<XScaleDataType>{1e-3, .5f}(xscale_host);
ck_tile::FillUniformDistribution<SmoothScaleDataType>{1e-3, .5f}(smscale_host);
ck_tile::DeviceMem x_buf(x_host.get_element_space_size_in_bytes());
ck_tile::DeviceMem xscale_buf(xscale_host.get_element_space_size_in_bytes());
ck_tile::DeviceMem smscale_buf(smscale_host.get_element_space_size_in_bytes());
ck_tile::DeviceMem yscale_buf(yscale_host_dev.get_element_space_size_in_bytes());
ck_tile::DeviceMem qy_buf(qy_host_dev.get_element_space_size_in_bytes());
x_buf.ToDevice(x_host.data());
xscale_buf.ToDevice(xscale_host.data());
smscale_buf.ToDevice(smscale_host.data());
std::cout << "[" << data_type << "]"
<< " m:" << m << ", n:" << n << ", x_stride:" << x_stride << ", y_stride:" << y_stride
......@@ -100,7 +100,7 @@ bool run(const ck_tile::ArgParser& arg_parser)
smoothquant_traits traits{data_type};
smoothquant_args args{x_buf.GetDeviceBuffer(),
xscale_buf.GetDeviceBuffer(),
smscale_buf.GetDeviceBuffer(),
yscale_buf.GetDeviceBuffer(),
qy_buf.GetDeviceBuffer(),
m,
......@@ -111,7 +111,7 @@ bool run(const ck_tile::ArgParser& arg_parser)
float ave_time = smoothquant(
traits, args, ck_tile::stream_config{nullptr, true, kname ? 1 : 0, warmup, repeat});
std::size_t num_byte = sizeof(XDataType) * m * n + sizeof(XScaleDataType) * n +
std::size_t num_byte = sizeof(XDataType) * m * n + sizeof(SmoothScaleDataType) * n +
sizeof(YScaleDataType) * m + sizeof(QYDataType) * m * n;
float gb_per_sec = num_byte / 1.E6 / ave_time;
......@@ -126,16 +126,16 @@ bool run(const ck_tile::ArgParser& arg_parser)
// smooth outlier
{
auto f = [&](auto n_) {
auto v_xscale = ck_tile::type_convert<ComputeDataType>(xscale_host(n_));
auto v_smscale = ck_tile::type_convert<ComputeDataType>(smscale_host(n_));
for(int m_ = 0; m_ < m; ++m_)
{
auto v_x = ck_tile::type_convert<ComputeDataType>(x_host(m_, n_));
y_host(m_, n_) = v_x * v_xscale;
y_host(m_, n_) = v_x * v_smscale;
}
};
ck_tile::make_ParallelTensorFunctor(f, xscale_host.get_element_space_size())(
ck_tile::make_ParallelTensorFunctor(f, smscale_host.get_element_space_size())(
std::thread::hardware_concurrency());
}
......
example/ck_tile/12_smoothquant/smoothquant.hpp
View file @ 0ef27d53
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
// Copyright (c) 2018-2025, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
......@@ -14,21 +14,21 @@ struct SmoothquantTypeConfig;
template <>
struct SmoothquantTypeConfig<ck_tile::half_t>
{
using XDataType = ck_tile::half_t;
using XScaleDataType = float;
using YScaleDataType = float;
using QYDataType = ck_tile::int8_t;
using ComputeDataType = float;
using XDataType = ck_tile::half_t;
using SmoothScaleDataType = float;
using YScaleDataType = float;
using QYDataType = ck_tile::int8_t;
using ComputeDataType = float;
};
template <>
struct SmoothquantTypeConfig<ck_tile::bf16_t>
{
using XDataType = ck_tile::bf16_t;
using XScaleDataType = float;
using YScaleDataType = float;
using QYDataType = ck_tile::int8_t;
using ComputeDataType = float;
using XDataType = ck_tile::bf16_t;
using SmoothScaleDataType = float;
using YScaleDataType = float;
using QYDataType = ck_tile::int8_t;
using ComputeDataType = float;
};
// runtime args
......
example/ck_tile/13_moe_sorting/moe_sorting_api.cpp
View file @ 0ef27d53
......@@ -3,18 +3,42 @@
#include "moe_sorting_api.hpp"
#define MOE_SORTING_DISPATCH(unroll_num_) \
constexpr ck_tile::index_t unroll_num = unroll_num_; \
using ms_problem = ck_tile::MoeSortingProblem<index_t, ms_weight_type, unroll_num>; \
using kernel = ck_tile::MoeSortingKernel<ms_problem>; \
auto kargs = kernel::MakeKargs(a); \
const dim3 grids = kernel::GridSize(a); \
const dim3 blocks = kernel::BlockSize(a); \
const auto lds_bytes = kernel::GetSmemSize(a); \
float ave_time = ck_tile::launch_kernel( \
s, ck_tile::make_kernel(kernel{}, grids, blocks, lds_bytes, kargs)); \
#define MOE_SORTING_DISPATCH_ETILE(unroll_num_, expert_tile_) \
constexpr ck_tile::index_t unroll_num = unroll_num_; \
constexpr ck_tile::index_t expert_tile = expert_tile_; \
using ms_problem = \
ck_tile::MoeSortingProblem<index_t, ms_weight_type, unroll_num, expert_tile>; \
using kernel = ck_tile::MoeSortingKernel<ms_problem>; \
auto kargs = kernel::MakeKargs(a); \
const dim3 grids = kernel::GridSize(a); \
const dim3 blocks = kernel::BlockSize(a); \
const auto lds_bytes = kernel::GetSmemSize(a); \
float ave_time = ck_tile::launch_kernel( \
s, ck_tile::make_kernel(kernel{}, grids, blocks, lds_bytes, kargs)); \
return ave_time;
#define MOE_SORTING_DISPATCH(unroll_num_) \
if(a.num_experts <= 8) \
{ \
MOE_SORTING_DISPATCH_ETILE(unroll_num_, 8) \
} \
else if(a.num_experts <= 16) \
{ \
MOE_SORTING_DISPATCH_ETILE(unroll_num_, 16) \
} \
else if(a.num_experts <= 32) \
{ \
MOE_SORTING_DISPATCH_ETILE(unroll_num_, 32) \
} \
else if(a.num_experts <= 64) \
{ \
MOE_SORTING_DISPATCH_ETILE(unroll_num_, 64) \
} \
else \
{ \
MOE_SORTING_DISPATCH_ETILE(unroll_num_, 0) \
}
float moe_sorting(moe_sorting_trait t, moe_sorting_args a, ck_tile::stream_config s)
{
if(t.weight_type == "fp32" && t.index_type == "int32")
......@@ -49,21 +73,12 @@ float moe_sorting(moe_sorting_trait t, moe_sorting_args a, ck_tile::stream_confi
case(6): {
MOE_SORTING_DISPATCH(6);
}
case(7): {
MOE_SORTING_DISPATCH(7);
}
case(8): {
MOE_SORTING_DISPATCH(8);
}
case(9): {
MOE_SORTING_DISPATCH(9);
}
case(10): {
MOE_SORTING_DISPATCH(10);
}
case(11): {
MOE_SORTING_DISPATCH(11);
}
default: {
MOE_SORTING_DISPATCH(4);
}
......
example/ck_tile/13_moe_sorting/script/smoke_test.sh
View file @ 0ef27d53
......@@ -16,4 +16,5 @@ $EXE -t=127 -e=99 -k=19
$EXE -t=71 -e=11 -k=11
$EXE -t=1 -e=1 -k=1
$EXE -t=99 -e=2 -k=1
$EXE -t=333 -e=99 -k=13
\ No newline at end of file
$EXE -t=333 -e=99 -k=13
$EXE -t=128 -e=32 -k=5 -moe_buf_size=262144
example/ck_tile/14_moe_smoothquant/instances/moe_smoothquant_instance_common.hpp
View file @ 0ef27d53
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
// Copyright (c) 2018-2025, Advanced Micro Devices, Inc. All rights reserved.
#include <ck_tile/core.hpp>
#include "moe_smoothquant.hpp"
......@@ -35,7 +35,7 @@ float moe_smoothquant_(const S& s, A a)
using PipelineProblem = ck_tile::SmoothquantPipelineProblem<
typename MoeSmoothquantTypeConfig<DataType>::XDataType,
typename MoeSmoothquantTypeConfig<DataType>::XScaleDataType,
typename MoeSmoothquantTypeConfig<DataType>::SmoothScaleDataType,
typename MoeSmoothquantTypeConfig<DataType>::ComputeDataType,
typename MoeSmoothquantTypeConfig<DataType>::YScaleDataType,
typename MoeSmoothquantTypeConfig<DataType>::QYDataType,
......
example/ck_tile/14_moe_smoothquant/moe_smoothquant.cpp
View file @ 0ef27d53
......@@ -91,15 +91,15 @@ bool run(const ck_tile::ArgParser& arg_parser)
using TypeConfig = MoeSmoothquantTypeConfig<DataType>;
using XDataType = typename TypeConfig::XDataType;
using XScaleDataType = typename TypeConfig::XScaleDataType;
using YScaleDataType = typename TypeConfig::YScaleDataType;
using QYDataType = typename TypeConfig::QYDataType;
using ComputeDataType = typename TypeConfig::ComputeDataType;
using XDataType = typename TypeConfig::XDataType;
using SmoothScaleDataType = typename TypeConfig::SmoothScaleDataType;
using YScaleDataType = typename TypeConfig::YScaleDataType;
using QYDataType = typename TypeConfig::QYDataType;
using ComputeDataType = typename TypeConfig::ComputeDataType;
// host verify
ck_tile::HostTensor<XDataType> x_host({tokens, hidden_size}, {stride, 1});
ck_tile::HostTensor<XScaleDataType> xscale_host({experts * hidden_size});
ck_tile::HostTensor<SmoothScaleDataType> smscale_host({experts * hidden_size});
ck_tile::HostTensor<ck_tile::index_t> topk_ids_host({tokens, topk});
ck_tile::HostTensor<YScaleDataType> yscale_host_ref({topk * tokens}, {1});
......@@ -110,16 +110,16 @@ bool run(const ck_tile::ArgParser& arg_parser)
topid_unique_gen<ck_tile::index_t>(topk_ids_host.mData, tokens, topk, experts, 11937);
ck_tile::FillUniformDistribution<XDataType>{-.5f, .5f}(x_host);
ck_tile::FillUniformDistribution<XScaleDataType>{1e-3, .5f}(xscale_host);
ck_tile::FillUniformDistribution<SmoothScaleDataType>{1e-3, .5f}(smscale_host);
ck_tile::DeviceMem x_buf(x_host.get_element_space_size_in_bytes());
ck_tile::DeviceMem xscale_buf(xscale_host.get_element_space_size_in_bytes());
ck_tile::DeviceMem smscale_buf(smscale_host.get_element_space_size_in_bytes());
ck_tile::DeviceMem topk_ids_buf(topk_ids_host.get_element_space_size_in_bytes());
ck_tile::DeviceMem yscale_buf(yscale_host_dev.get_element_space_size_in_bytes());
ck_tile::DeviceMem qy_buf(qy_host_dev.get_element_space_size_in_bytes());
x_buf.ToDevice(x_host.data());
xscale_buf.ToDevice(xscale_host.data());
smscale_buf.ToDevice(smscale_host.data());
topk_ids_buf.ToDevice(topk_ids_host.data());
std::cout << "[" << data_type << "]"
......@@ -129,7 +129,7 @@ bool run(const ck_tile::ArgParser& arg_parser)
moe_smoothquant_traits traits{data_type};
moe_smoothquant_args args{x_buf.GetDeviceBuffer(),
xscale_buf.GetDeviceBuffer(),
smscale_buf.GetDeviceBuffer(),
topk_ids_buf.GetDeviceBuffer(),
yscale_buf.GetDeviceBuffer(),
qy_buf.GetDeviceBuffer(),
......@@ -143,9 +143,10 @@ bool run(const ck_tile::ArgParser& arg_parser)
float ave_time = moe_smoothquant(
traits, args, ck_tile::stream_config{nullptr, true, kname ? 1 : 0, warmup, repeat});
std::size_t num_byte =
sizeof(XDataType) * tokens * hidden_size + sizeof(XScaleDataType) * topk * hidden_size +
sizeof(YScaleDataType) * topk * tokens + sizeof(QYDataType) * topk * tokens * hidden_size;
std::size_t num_byte = sizeof(XDataType) * tokens * hidden_size +
sizeof(SmoothScaleDataType) * topk * hidden_size +
sizeof(YScaleDataType) * topk * tokens +
sizeof(QYDataType) * topk * tokens * hidden_size;
float gb_per_sec = num_byte / 1.E6 / ave_time;
std::cout << ", " << ave_time * 1.E3 << " us, " << gb_per_sec << " GB/s" << std::flush;
......@@ -165,11 +166,11 @@ bool run(const ck_tile::ArgParser& arg_parser)
for(int i_h = 0; i_h < hidden_size; ++i_h)
{
auto v_xscale = ck_tile::type_convert<ComputeDataType>(
xscale_host(i_expert * hidden_size + i_h));
auto v_smscale = ck_tile::type_convert<ComputeDataType>(
smscale_host(i_expert * hidden_size + i_h));
auto v_x = ck_tile::type_convert<ComputeDataType>(x_host(i_token, i_h));
// y_host(i_token * topk + i_topk, i_h) = v_x * v_xscale;
y_host(i_topk * tokens + i_token, i_h) = v_x * v_xscale;
// y_host(i_token * topk + i_topk, i_h) = v_x * v_smscale;
y_host(i_topk * tokens + i_token, i_h) = v_x * v_smscale;
}
}
};
......
example/ck_tile/14_moe_smoothquant/moe_smoothquant.hpp
View file @ 0ef27d53
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
// Copyright (c) 2018-2025, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
......@@ -14,21 +14,21 @@ struct MoeSmoothquantTypeConfig;
template <>
struct MoeSmoothquantTypeConfig<ck_tile::half_t>
{
using XDataType = ck_tile::half_t;
using XScaleDataType = float;
using YScaleDataType = float;
using QYDataType = ck_tile::int8_t;
using ComputeDataType = float;
using XDataType = ck_tile::half_t;
using SmoothScaleDataType = float;
using YScaleDataType = float;
using QYDataType = ck_tile::int8_t;
using ComputeDataType = float;
};
template <>
struct MoeSmoothquantTypeConfig<ck_tile::bf16_t>
{
using XDataType = ck_tile::bf16_t;
using XScaleDataType = float;
using YScaleDataType = float;
using QYDataType = ck_tile::int8_t;
using ComputeDataType = float;
using XDataType = ck_tile::bf16_t;
using SmoothScaleDataType = float;
using YScaleDataType = float;
using QYDataType = ck_tile::int8_t;
using ComputeDataType = float;
};
// runtime args
......
example/ck_tile/15_fused_moe/README.md
View file @ 0ef27d53
......@@ -8,6 +8,9 @@ The benifit of this fused-moe:
* much less kernel instance, easy to maintain
# Implementation and feature support
## NOTES:
currently gate+up in fp16 case will very easily cause accumulator overflow the fp16 max(65504), hence result in INF. Please use BF16 for gate+up case, API side will have no check for this.
## moe-sorting
this is a common pre-process step before the actual moe-gemm. The purpose is to transform the moe loop over from token-by-token to expert-by-expert, make sure very workgroup is working for a single expert (B matrix). Besides, we extend this op to do the zeroing of the output buffer(to be used for reduce buffer with atomic)
......
example/ck_tile/15_fused_moe/fused_moe.hpp
View file @ 0ef27d53
......@@ -26,7 +26,7 @@ struct fused_moe_args
ck_tile::index_t block_m; // block_m, used to devide the input
ck_tile::index_t hidden_size; // k
ck_tile::index_t intermediate_size; // n / TP, for Gate. if Gate+Up, Down need divide by 2
ck_tile::index_t intermediate_size; // n / TP, for Gate. and Up, Down is also this value
ck_tile::index_t num_tokens; // input number of tokens for current iteration
ck_tile::index_t num_experts; // number of groups
ck_tile::index_t topk; // need this?
......@@ -45,7 +45,8 @@ struct fused_moe_traits
std::string prec_sq; // smooth quant scale
std::string prec_kw; // topk-weight data type
int block_m;
int gate_only;
int activation; // 0:gelu, 1:silu
int gate_only; // 0:g1u0, 1:g1u1
int fused_quant; // 0:no-sweep, 1:smooth-dynamic-quant, 2:dynamic-quant
};
......
example/ck_tile/15_fused_moe/fused_moegemm.hpp
View file @ 0ef27d53
......@@ -77,7 +77,8 @@ struct fused_moegemm_traits
std::string prec_sq; // smooth quant scale
std::string prec_kw; // topk-weight data type
int block_m;
int gate_only;
int activation; // 0:gelu, 1:silu
int gate_only; // 0:g1u0, 1:g1u1
int fused_quant; // 0:no-sweep, 1:smooth-dynamic-quant, 2:dynamic-quant
};
......
example/ck_tile/15_fused_moe/instances/fused_moe_api.cpp
View file @ 0ef27d53
......@@ -41,6 +41,7 @@ float fused_moe(fused_moe_traits t, fused_moe_args a, const ck_tile::stream_conf
t.prec_sq,
t.prec_kw,
t.block_m,
t.activation,
t.gate_only,
t.fused_quant};
auto a1 = fused_moegemm_args{
......
example/ck_tile/15_fused_moe/instances/fused_moegemm_api.cpp
View file @ 0ef27d53
......@@ -17,15 +17,67 @@ float fused_moegemm(fused_moegemm_traits t, fused_moegemm_args a, const ck_tile:
// clang-format off
float r = -1;
if(t.prec_i == "bf16" && t.prec_w == "bf16" && t.prec_o == "bf16" && t.prec_st == "fp32" &&
t.prec_sw == "fp32" && t.prec_sq == "fp32" && t.prec_kw == "fp32" && t.block_m == 32 && t.gate_only == 1)
t.prec_sw == "fp32" && t.prec_sq == "fp32" && t.prec_kw == "fp32" && t.block_m == 32 && t.gate_only == 1 && t.activation == 0)
{
using t_ = fmoe_<ck_tile::bf16_t, ck_tile::bf16_t, ck_tile::bf16_t, float, float, float, float, S<32, 512, 128, 128>, S<1, 4, 1>, S<16, 16, 32>, 1, 0>;
constexpr ck_tile::index_t act_ = 0;
constexpr ck_tile::index_t go_ = 1;
using t_ = fmoe_<ck_tile::bf16_t, ck_tile::bf16_t, ck_tile::bf16_t, float, float, float, float, S<32, 512, 128, 128>, S<1, 4, 1>, S<16, 16, 32>, act_, go_, 0>;
r = fused_moegemm_<t_>(s, a);
}
else if(t.prec_i == "bf16" && t.prec_w == "bf16" && t.prec_o == "bf16" && t.prec_st == "fp32" &&
t.prec_sw == "fp32" && t.prec_sq == "fp32" && t.prec_kw == "fp32" && t.block_m == 32 && t.gate_only == 0 && t.activation == 0)
{
constexpr ck_tile::index_t act_ = 0;
constexpr ck_tile::index_t go_ = 0;
using t_ = fmoe_<ck_tile::bf16_t, ck_tile::bf16_t, ck_tile::bf16_t, float, float, float, float, S<32, 512, 128, 128>, S<1, 4, 1>, S<16, 16, 32>, act_, go_, 0>;
r = fused_moegemm_<t_>(s, a);
}
else if(t.prec_i == "fp16" && t.prec_w == "fp16" && t.prec_o == "fp16" && t.prec_st == "fp32" &&
t.prec_sw == "fp32" && t.prec_sq == "fp32" && t.prec_kw == "fp32" && t.block_m == 32 && t.gate_only == 1 && t.activation == 0)
{
constexpr ck_tile::index_t act_ = 0;
constexpr ck_tile::index_t go_ = 1;
using t_ = fmoe_<ck_tile::fp16_t, ck_tile::fp16_t, ck_tile::fp16_t, float, float, float, float, S<32, 512, 128, 128>, S<1, 4, 1>, S<16, 16, 32>, act_, go_, 0>;
r = fused_moegemm_<t_>(s, a);
}
else if(t.prec_i == "fp16" && t.prec_w == "fp16" && t.prec_o == "fp16" && t.prec_st == "fp32" &&
t.prec_sw == "fp32" && t.prec_sq == "fp32" && t.prec_kw == "fp32" && t.block_m == 32 && t.gate_only == 0 && t.activation == 0)
{
constexpr ck_tile::index_t act_ = 0;
constexpr ck_tile::index_t go_ = 0;
using t_ = fmoe_<ck_tile::fp16_t, ck_tile::fp16_t, ck_tile::fp16_t, float, float, float, float, S<32, 512, 128, 128>, S<1, 4, 1>, S<16, 16, 32>, act_, go_, 0>;
r = fused_moegemm_<t_>(s, a);
}
else if(t.prec_i == "bf16" && t.prec_w == "bf16" && t.prec_o == "bf16" && t.prec_st == "fp32" &&
t.prec_sw == "fp32" && t.prec_sq == "fp32" && t.prec_kw == "fp32" && t.block_m == 32 && t.gate_only == 1 && t.activation == 1)
{
constexpr ck_tile::index_t act_ = 1;
constexpr ck_tile::index_t go_ = 1;
using t_ = fmoe_<ck_tile::bf16_t, ck_tile::bf16_t, ck_tile::bf16_t, float, float, float, float, S<32, 512, 128, 128>, S<1, 4, 1>, S<16, 16, 32>, act_, go_, 0>;
r = fused_moegemm_<t_>(s, a);
}
else if(t.prec_i == "bf16" && t.prec_w == "bf16" && t.prec_o == "bf16" && t.prec_st == "fp32" &&
t.prec_sw == "fp32" && t.prec_sq == "fp32" && t.prec_kw == "fp32" && t.block_m == 32 && t.gate_only == 0 && t.activation == 1)
{
constexpr ck_tile::index_t act_ = 1;
constexpr ck_tile::index_t go_ = 0;
using t_ = fmoe_<ck_tile::bf16_t, ck_tile::bf16_t, ck_tile::bf16_t, float, float, float, float, S<32, 512, 128, 128>, S<1, 4, 1>, S<16, 16, 32>, act_, go_, 0>;
r = fused_moegemm_<t_>(s, a);
}
else if(t.prec_i == "fp16" && t.prec_w == "fp16" && t.prec_o == "fp16" && t.prec_st == "fp32" &&
t.prec_sw == "fp32" && t.prec_sq == "fp32" && t.prec_kw == "fp32" && t.block_m == 32 && t.gate_only == 1 && t.activation == 1)
{
constexpr ck_tile::index_t act_ = 1;
constexpr ck_tile::index_t go_ = 1;
using t_ = fmoe_<ck_tile::fp16_t, ck_tile::fp16_t, ck_tile::fp16_t, float, float, float, float, S<32, 512, 128, 128>, S<1, 4, 1>, S<16, 16, 32>, act_, go_, 0>;
r = fused_moegemm_<t_>(s, a);
}
else if(t.prec_i == "fp16" && t.prec_w == "fp16" && t.prec_o == "fp16" && t.prec_st == "fp32" &&
t.prec_sw == "fp32" && t.prec_sq == "fp32" && t.prec_kw == "fp32" && t.block_m == 32 && t.gate_only == 1)
t.prec_sw == "fp32" && t.prec_sq == "fp32" && t.prec_kw == "fp32" && t.block_m == 32 && t.gate_only == 0 && t.activation == 1)
{
using t_ = fmoe_<ck_tile::fp16_t, ck_tile::fp16_t, ck_tile::fp16_t, float, float, float, float, S<32, 512, 128, 128>, S<1, 4, 1>, S<16, 16, 32>, 1, 0>;
constexpr ck_tile::index_t act_ = 1;
constexpr ck_tile::index_t go_ = 0;
using t_ = fmoe_<ck_tile::fp16_t, ck_tile::fp16_t, ck_tile::fp16_t, float, float, float, float, S<32, 512, 128, 128>, S<1, 4, 1>, S<16, 16, 32>, act_, go_, 0>;
r = fused_moegemm_<t_>(s, a);
}
// clang-format on
......
example/ck_tile/15_fused_moe/instances/fused_moegemm_api_internal.hpp
View file @ 0ef27d53
......@@ -21,21 +21,31 @@ float fused_moegemm_(const ck_tile::stream_config& s, fused_moegemm_args a)
typename Ts_::BlockTile_1,
typename Ts_::WarpPerBlock_0,
typename Ts_::WarpTile_0>;
using f_problem =
ck_tile::FusedMoeGemmPipelineProblem<typename Ts_::ADataType,
typename Ts_::GDataType,
typename Ts_::DDataType,
typename Ts_::AccDataType,
typename Ts_::ODataType,
typename Ts_::AScaleDataType,
typename Ts_::GScaleDataType,
typename Ts_::DScaleDataType,
typename Ts_::YSmoothScaleDataType,
typename Ts_::TopkWeightDataType,
typename Ts_::IndexDataType,
ck_tile::element_wise::FastGeluAsm, // TODO: hardcoded
f_shape,
f_traits>;
constexpr auto get_activation_ = []() {
if constexpr(Ts_::Activation == 0)
{
return ck_tile::element_wise::FastGeluAsm{};
}
else
return ck_tile::element_wise::Silu{};
};
using f_act_ = ck_tile::remove_cvref_t<decltype(get_activation_())>;
using f_problem = ck_tile::FusedMoeGemmPipelineProblem<typename Ts_::ADataType,
typename Ts_::GDataType,
typename Ts_::DDataType,
typename Ts_::AccDataType,
typename Ts_::ODataType,
typename Ts_::AScaleDataType,
typename Ts_::GScaleDataType,
typename Ts_::DScaleDataType,
typename Ts_::YSmoothScaleDataType,
typename Ts_::TopkWeightDataType,
typename Ts_::IndexDataType,
f_act_, // TODO: hardcoded
f_shape,
f_traits>;
// using f_pipeline = ck_tile::FusedMoeGemmPipeline_FlatmmEx<f_problem>;
using f_pipeline = ck_tile::FusedMoeGemmPipeline_FlatmmUk<f_problem>;
......
example/ck_tile/15_fused_moe/instances/fused_moegemm_api_traits.hpp
View file @ 0ef27d53
......@@ -15,7 +15,8 @@ template <typename I,
typename KW,
typename BlockTIle_, // seq<b_token, b_interm, b_hidden, b_down>
typename WarpPerBlock_,
typename WarpTile_, // seq<*,*,*>, used to select mfma
typename WarpTile_, // seq<*,*,*>, used to select mfma
ck_tile::index_t Activation_ = 0, // 0: Gelu 1: Silu
ck_tile::index_t GateOnly_ = 0,
ck_tile::index_t FusedQuant_ = 0>
struct fmoe_ // traits, ugly name, only used for internal
......@@ -44,10 +45,11 @@ struct fmoe_ // traits, ugly name, only used for internal
using WarpPerBlock_0 = ck_tile::remove_cvref_t<WarpPerBlock_>;
using WarpTile_0 = ck_tile::remove_cvref_t<WarpTile_>;
using BlockTile_1 = ck_tile::sequence<BT_, BD_, BI_ / (GateOnly_ ? 1 : 2)>;
using BlockTile_1 = ck_tile::sequence<BT_, BD_, BI_>;
using WarpPerBlock_1 = ck_tile::remove_cvref_t<WarpPerBlock_>;
using WarpTile_1 = ck_tile::remove_cvref_t<WarpTile_>;
static constexpr ck_tile::index_t Activation = Activation_; // 0: Gelu 1: Silu
static constexpr ck_tile::index_t GateOnly = GateOnly_;
static constexpr ck_tile::index_t FusedQuant = FusedQuant_;
};
example/ck_tile/15_fused_moe/instances/fused_moegemm_bf16_m32.cpp
View file @ 0ef27d53
......@@ -8,7 +8,18 @@
// clang-format off
template float fused_moegemm_<
fmoe_<ck_tile::bf16_t, ck_tile::bf16_t, ck_tile::bf16_t, float, float, float, float, S<32, 512, 128, 128>, S<1, 4, 1>, S<16, 16, 32>, 1, 0>
fmoe_<ck_tile::bf16_t, ck_tile::bf16_t, ck_tile::bf16_t, float, float, float, float, S<32, 512, 128, 128>, S<1, 4, 1>, S<16, 16, 32>, 0, 0, 0>
>(const ck_tile::stream_config& s, fused_moegemm_args a);
template float fused_moegemm_<
fmoe_<ck_tile::bf16_t, ck_tile::bf16_t, ck_tile::bf16_t, float, float, float, float, S<32, 512, 128, 128>, S<1, 4, 1>, S<16, 16, 32>, 0, 1, 0>
>(const ck_tile::stream_config& s, fused_moegemm_args a);
template float fused_moegemm_<
fmoe_<ck_tile::bf16_t, ck_tile::bf16_t, ck_tile::bf16_t, float, float, float, float, S<32, 512, 128, 128>, S<1, 4, 1>, S<16, 16, 32>, 1, 0, 0>
>(const ck_tile::stream_config& s, fused_moegemm_args a);
template float fused_moegemm_<
fmoe_<ck_tile::bf16_t, ck_tile::bf16_t, ck_tile::bf16_t, float, float, float, float, S<32, 512, 128, 128>, S<1, 4, 1>, S<16, 16, 32>, 1, 1, 0>
>(const ck_tile::stream_config& s, fused_moegemm_args a);
// clang-format on
example/ck_tile/15_fused_moe/instances/fused_moegemm_fp16_m32.cpp
View file @ 0ef27d53
......@@ -8,7 +8,19 @@
// clang-format off
template float fused_moegemm_<
fmoe_<ck_tile::fp16_t, ck_tile::fp16_t, ck_tile::fp16_t, float, float, float, float, S<32, 512, 128, 128>, S<1, 4, 1>, S<16, 16, 32>, 1, 0>
fmoe_<ck_tile::fp16_t, ck_tile::fp16_t, ck_tile::fp16_t, float, float, float, float, S<32, 512, 128, 128>, S<1, 4, 1>, S<16, 16, 32>, 0, 0, 0>
>(const ck_tile::stream_config& s, fused_moegemm_args a);
template float fused_moegemm_<
fmoe_<ck_tile::fp16_t, ck_tile::fp16_t, ck_tile::fp16_t, float, float, float, float, S<32, 512, 128, 128>, S<1, 4, 1>, S<16, 16, 32>, 0, 1, 0>
>(const ck_tile::stream_config& s, fused_moegemm_args a);
template float fused_moegemm_<
fmoe_<ck_tile::fp16_t, ck_tile::fp16_t, ck_tile::fp16_t, float, float, float, float, S<32, 512, 128, 128>, S<1, 4, 1>, S<16, 16, 32>, 1, 0, 0>
>(const ck_tile::stream_config& s, fused_moegemm_args a);
template float fused_moegemm_<
fmoe_<ck_tile::fp16_t, ck_tile::fp16_t, ck_tile::fp16_t, float, float, float, float, S<32, 512, 128, 128>, S<1, 4, 1>, S<16, 16, 32>, 1, 1, 0>
>(const ck_tile::stream_config& s, fused_moegemm_args a);
// clang-format on
example/ck_tile/15_fused_moe/instances/fused_moesorting_api.cpp
View file @ 0ef27d53
......@@ -3,18 +3,42 @@
#include "fused_moesorting.hpp"
#define MOE_SORTING_DISPATCH(unroll_num_) \
constexpr ck_tile::index_t unroll_num = unroll_num_; \
using ms_problem = ck_tile::MoeSortingProblem<index_t, ms_weight_type, unroll_num>; \
using kernel = ck_tile::MoeSortingKernel<ms_problem>; \
auto kargs = kernel::MakeKargs(a); \
const dim3 grids = kernel::GridSize(a); \
const dim3 blocks = kernel::BlockSize(a); \
const auto lds_bytes = kernel::GetSmemSize(a); \
float ave_time = ck_tile::launch_kernel( \
s, ck_tile::make_kernel(kernel{}, grids, blocks, lds_bytes, kargs)); \
#define MOE_SORTING_DISPATCH_ETILE(unroll_num_, expert_tile_) \
constexpr ck_tile::index_t unroll_num = unroll_num_; \
constexpr ck_tile::index_t expert_tile = expert_tile_; \
using ms_problem = \
ck_tile::MoeSortingProblem<index_t, ms_weight_type, unroll_num, expert_tile>; \
using kernel = ck_tile::MoeSortingKernel<ms_problem>; \
auto kargs = kernel::MakeKargs(a); \
const dim3 grids = kernel::GridSize(a); \
const dim3 blocks = kernel::BlockSize(a); \
const auto lds_bytes = kernel::GetSmemSize(a); \
float ave_time = ck_tile::launch_kernel( \
s, ck_tile::make_kernel(kernel{}, grids, blocks, lds_bytes, kargs)); \
return ave_time;
#define MOE_SORTING_DISPATCH(unroll_num_) \
if(a.num_experts <= 8) \
{ \
MOE_SORTING_DISPATCH_ETILE(unroll_num_, 8) \
} \
else if(a.num_experts <= 16) \
{ \
MOE_SORTING_DISPATCH_ETILE(unroll_num_, 16) \
} \
else if(a.num_experts <= 32) \
{ \
MOE_SORTING_DISPATCH_ETILE(unroll_num_, 32) \
} \
else if(a.num_experts <= 64) \
{ \
MOE_SORTING_DISPATCH_ETILE(unroll_num_, 64) \
} \
else \
{ \
MOE_SORTING_DISPATCH_ETILE(unroll_num_, 0) \
}
float fused_moesorting(fused_moesorting_trait t, fused_moesorting_args a, ck_tile::stream_config s)
{
if(t.weight_type == "fp32" && t.index_type == "int32")
......@@ -49,21 +73,12 @@ float fused_moesorting(fused_moesorting_trait t, fused_moesorting_args a, ck_til
case(6): {
MOE_SORTING_DISPATCH(6);
}
case(7): {
MOE_SORTING_DISPATCH(7);
}
case(8): {
MOE_SORTING_DISPATCH(8);
}
case(9): {
MOE_SORTING_DISPATCH(9);
}
case(10): {
MOE_SORTING_DISPATCH(10);
}
case(11): {
MOE_SORTING_DISPATCH(11);
}
default: {
MOE_SORTING_DISPATCH(4);
}
......
example/ck_tile/15_fused_moe/main.cpp
View file @ 0ef27d53
......@@ -108,12 +108,14 @@ auto create_args(int argc, char* argv[])
.insert(
"gate_only", "1", "w0(gate/up) style, 0:gate+up will double interm size, 1:only gate")
.insert("api", "0", "benchmark api set: 0:fused-moe(moe-gemm+moe-sorting), 1:moe-gemm")
.insert("act", "0", "activation after first gemm. 0:gelu, 1:silu")
.insert("balance",
"0",
"if set to 1, will try balance the expert in topk-ids(convenient for testing)")
.insert("init",
"2",
"init method. 0:random stepped float(fast). 1: random uniform, 2:rand normalized"
"1",
"init method. 0:random stepped float(fast). 1: random uniform[-0.5, 0.5], 2:rand "
"normalized[0, 1]"
"normalized(slow)")
.insert("seed", "11939", "seed used to do random")
.insert("warmup", "5", "cold iter")
......@@ -135,6 +137,7 @@ bool run(const ck_tile::ArgParser& arg_parser)
ck_tile::index_t intermediate_size = arg_parser.get_int("i");
ck_tile::index_t stride = arg_parser.get_int("stride");
ck_tile::index_t block_m = arg_parser.get_int("bm");
ck_tile::index_t activation = arg_parser.get_int("act");
if(stride < 0)
stride = hidden_size;
std::string prec_i = arg_parser.get_str("prec_i");
......@@ -194,11 +197,14 @@ bool run(const ck_tile::ArgParser& arg_parser)
return std::string(", st:") + std::to_string(stride);
}();
std::cout << "[" << api_str << "|" << prec_str << "]"
<< " t:" << tokens << ", e:" << experts << ", k:" << topk << stride_str
<< ", hidden:" << hidden_size << ", interm:" << intermediate_size << ", tp:" << tp
<< ", shrd_interm:" << shared_intermediate_size_0 << "|" << shared_intermediate_size_1
<< ", go:" << gate_only << ", q:" << fused_quant << std::flush;
std::cout
<< "[" << api_str << "|" << prec_str << "]"
<< " t:" << tokens << ", e:" << experts << ", k:" << topk << stride_str
<< ", hidden:" << hidden_size << ", interm:" << intermediate_size << ", tp:" << tp
<< ", act:"
<< activation
// << ", shrd_interm:" << shared_intermediate_size_0 << "|" << shared_intermediate_size_1
<< (gate_only ? ", g1u0" : ", g1u1") << ", q:" << fused_quant << std::flush;
using TypeConfig = FusedMoeGemmTypeConfig<I, W, O, ST, SW, SQ, KW>;
using ADataType = typename TypeConfig::ADataType;
......@@ -370,6 +376,7 @@ bool run(const ck_tile::ArgParser& arg_parser)
prec_sq,
prec_kw,
block_m,
activation,
gate_only,
fused_quant};
......@@ -389,7 +396,7 @@ bool run(const ck_tile::ArgParser& arg_parser)
num_sorted_tiles_buf.GetDeviceBuffer(),
block_m,
hidden_size,
shared_intermediate_size_0,
intermediate_size / tp,
tokens,
experts,
topk,
......@@ -408,6 +415,28 @@ bool run(const ck_tile::ArgParser& arg_parser)
<< cal_tbps(ave_time) << " TB/s" << std::flush;
bool pass = true;
#define CPU_FUSED_MOE(act_type_) \
ck_tile::reference_fused_moe<AccDataType, act_type_>(a_host, \
g_host, \
d_host, \
sa_host, \
sg_host, \
sd_host, \
sy_host, \
o_host, \
sorted_token_ids_host, \
sorted_weight_host, \
sorted_expert_ids_host, \
num_sorted_tiles_host, \
topk_ids_host, \
block_m, \
tokens, \
experts, \
hidden_size, \
intermediate_size / tp, \
topk, \
gate_only)
if(do_validation)
{
ck_tile::reference_moe_sorting<TopkWeightDataType, IndexDataType>(
......@@ -419,28 +448,14 @@ bool run(const ck_tile::ArgParser& arg_parser)
num_sorted_tiles_host.mData[0],
experts,
block_m);
ck_tile::reference_fused_moe<AccDataType, ck_tile::element_wise::Gelu>(
a_host,
g_host,
d_host,
sa_host,
sg_host,
sd_host,
sy_host,
o_host,
sorted_token_ids_host,
sorted_weight_host,
sorted_expert_ids_host,
num_sorted_tiles_host,
topk_ids_host,
block_m,
tokens,
experts,
hidden_size,
shared_intermediate_size_0,
topk,
gate_only);
if(activation == 0)
{
CPU_FUSED_MOE(ck_tile::element_wise::Gelu);
}
else
{
CPU_FUSED_MOE(ck_tile::element_wise::Silu);
}
auto o_dev = o_buf.ToHost<ODataType>();
// o_dev.savetxt("gpu-out.txt", "float");
......@@ -491,6 +506,7 @@ bool run(const ck_tile::ArgParser& arg_parser)
prec_sq,
prec_kw,
block_m,
activation,
gate_only,
fused_quant};
......@@ -507,7 +523,7 @@ bool run(const ck_tile::ArgParser& arg_parser)
sorted_expert_ids_buf.GetDeviceBuffer(),
num_sorted_tiles_buf.GetDeviceBuffer(),
hidden_size,
shared_intermediate_size_0,
intermediate_size / tp,
tokens,
experts,
topk,
......@@ -529,27 +545,14 @@ bool run(const ck_tile::ArgParser& arg_parser)
if(do_validation)
{
ck_tile::reference_fused_moe<AccDataType, ck_tile::element_wise::Gelu>(
a_host,
g_host,
d_host,
sa_host,
sg_host,
sd_host,
sy_host,
o_host,
sorted_token_ids_host,
sorted_weight_host,
sorted_expert_ids_host,
num_sorted_tiles_host,
topk_ids_host,
block_m,
tokens,
experts,
hidden_size,
shared_intermediate_size_0,
topk,
gate_only);
if(activation == 0)
{
CPU_FUSED_MOE(ck_tile::element_wise::Gelu);
}
else
{
CPU_FUSED_MOE(ck_tile::element_wise::Silu);
}
auto o_dev = o_buf.ToHost<ODataType>();
// o_dev.savetxt("gpu-out.txt", "float");
......
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment