Skip to content

GitLab

Unverified Commit 2a30cfdd authored by arai713's avatar arai713 Committed by GitHub
Browse files

Merge branch 'develop' into codegen-enable-hiprtc

parents 9533a172 78195ccc
Hide whitespace changes
Inline Side-by-side
Showing with 345 additions and 121 deletions
example/ck_tile/12_smoothquant/instances/smoothquant_instance_common.hpp
View file @ 2a30cfdd
// 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 @ 2a30cfdd
......@@ -33,7 +33,8 @@ auto create_args(int argc, char* argv[])
ck_tile::ArgParser arg_parser;
arg_parser.insert("m", "3328", "m dimension")
.insert("n", "4096", "n dimension")
.insert("stride", "-1", "stride per row, if -1 then equal to n")
.insert("x_stride", "-1", "input stride per row, if -1 then equal to n")
.insert("y_stride", "-1", "output stride per row, if -1 then equal to n")
.insert("v", "1", "cpu validation or not")
.insert("kname", "1", "print kernel name or not")
.insert("prec", "fp16", "precision")
......@@ -47,65 +48,70 @@ auto create_args(int argc, char* argv[])
template <typename DataType>
bool run(const ck_tile::ArgParser& arg_parser)
{
ck_tile::index_t m = arg_parser.get_int("m");
ck_tile::index_t n = arg_parser.get_int("n");
ck_tile::index_t stride = arg_parser.get_int("stride");
if(stride < 0)
stride = n;
ck_tile::index_t m = arg_parser.get_int("m");
ck_tile::index_t n = arg_parser.get_int("n");
ck_tile::index_t x_stride = arg_parser.get_int("x_stride");
if(x_stride < 0)
x_stride = n;
ck_tile::index_t y_stride = arg_parser.get_int("y_stride");
if(y_stride < 0)
y_stride = n;
std::string data_type = arg_parser.get_str("prec");
int kname = arg_parser.get_int("kname");
int do_validation = arg_parser.get_int("v");
int warmup = arg_parser.get_int("warmup");
int repeat = arg_parser.get_int("repeat");
assert(stride >= n);
assert(x_stride >= n);
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}, {stride, 1});
ck_tile::HostTensor<XScaleDataType> xscale_host({n});
ck_tile::HostTensor<XDataType> x_host({m, n}, {x_stride, 1});
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});
ck_tile::HostTensor<QYDataType> qy_host_ref({m, n}, {stride, 1});
ck_tile::HostTensor<QYDataType> qy_host_dev({m, n}, {stride, 1});
ck_tile::HostTensor<QYDataType> qy_host_ref({m, n}, {y_stride, 1});
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 << ", stride:" << stride << std::flush;
<< " m:" << m << ", n:" << n << ", x_stride:" << x_stride << ", y_stride:" << y_stride
<< std::flush;
smoothquant_traits traits{data_type};
smoothquant_args args{x_buf.GetDeviceBuffer(),
xscale_buf.GetDeviceBuffer(),
smscale_buf.GetDeviceBuffer(),
yscale_buf.GetDeviceBuffer(),
qy_buf.GetDeviceBuffer(),
m,
n,
stride};
x_stride,
y_stride};
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;
......@@ -116,20 +122,20 @@ bool run(const ck_tile::ArgParser& arg_parser)
if(do_validation)
{
using YDataType = ComputeDataType;
ck_tile::HostTensor<ComputeDataType> y_host({m, n}, {stride, 1});
ck_tile::HostTensor<ComputeDataType> y_host({m, n}, {y_stride, 1});
// 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());
}
......@@ -166,7 +172,7 @@ bool run(const ck_tile::ArgParser& arg_parser)
qy_buf.FromDevice(qy_host_dev.data());
auto [rtol, atol] = get_elimit<QYDataType>();
if(stride == n)
if(y_stride == n)
{
pass = ck_tile::check_err(qy_host_dev,
qy_host_ref,
......@@ -178,10 +184,12 @@ bool run(const ck_tile::ArgParser& arg_parser)
{
for(int i_r = 0; i_r < m; i_r++)
{
std::vector<QYDataType> qy_host_dev_row(qy_host_dev.begin() + i_r * stride,
qy_host_dev.begin() + i_r * stride + n);
std::vector<QYDataType> qy_host_ref_row(qy_host_ref.begin() + i_r * stride,
qy_host_ref.begin() + i_r * stride + n);
std::vector<QYDataType> qy_host_dev_row(qy_host_dev.begin() + i_r * y_stride,
qy_host_dev.begin() + i_r * y_stride +
n);
std::vector<QYDataType> qy_host_ref_row(qy_host_ref.begin() + i_r * y_stride,
qy_host_ref.begin() + i_r * y_stride +
n);
pass &= ck_tile::check_err(qy_host_dev_row,
qy_host_ref_row,
std::string("qy[") + std::to_string(i_r) +
......
example/ck_tile/12_smoothquant/smoothquant.hpp
View file @ 2a30cfdd
// 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.cpp
View file @ 2a30cfdd
......@@ -26,6 +26,10 @@ auto create_args(int argc, char* argv[])
.insert("k", "4", "topk")
.insert("unit", "32", "unit_size")
.insert("moe_buf_size", "0", "moe_buf_size")
.insert("local_eid",
"-1",
"a list of experts enabled as local expert. e.g. \"0,1,4,5\"\n"
"please make sure eid is in ascending order!")
.insert("seed", "-1", "seed to be used, -1 means random every time")
.insert("kname", "0", "when set to 1 it will print kernel name")
.insert("warmup", "5", "number of iterations before benchmark the kernel")
......@@ -74,6 +78,7 @@ bool test_moe_sorting(ck_tile::ArgParser args)
int kname = args.get_int("kname");
int warmup = args.get_int("warmup");
int repeat = args.get_int("repeat");
int max_output_ids =
ck_tile::integer_least_multiple(topk * tokens + num_experts * unit_size - topk, unit_size);
......@@ -90,6 +95,30 @@ bool test_moe_sorting(ck_tile::ArgParser args)
return false;
}
bool local_expert_masking = args.get_str("local_eid") != "-1";
auto local_expert_masking_host = [&]() {
if(local_expert_masking)
{
auto local_eid = args.get_int_vec("local_eid");
// std::vector<int> v_ {num_experts, 0};
ck_tile::HostTensor<IndexType> v_{{num_experts}};
v_.SetZero();
for(auto eid : local_eid)
{
if(eid >= num_experts)
{
throw std::runtime_error(
"local_eid larger than number of expert, please check");
}
v_.mData[eid] = 1;
}
return v_;
}
else
// return std::vector<int>{};
return ck_tile::HostTensor<IndexType>{{1}};
}();
// tokens already considered batch size
ck_tile::HostTensor<IndexType> topk_ids_host({tokens, topk}, {topk, 1});
ck_tile::HostTensor<WeightType> weights_host({tokens, topk}, {topk, 1});
......@@ -111,6 +140,8 @@ bool test_moe_sorting(ck_tile::ArgParser args)
sorted_expert_ids_host.get_element_space_size_in_bytes());
ck_tile::DeviceMem sorted_id_cnt_dev(sorted_id_cnt_host.get_element_space_size_in_bytes());
ck_tile::DeviceMem moe_buf_dev(moe_buf_host.get_element_space_size_in_bytes());
ck_tile::DeviceMem local_expert_masking_dev(
local_expert_masking_host.get_element_space_size_in_bytes());
topk_ids_dev.ToDevice(topk_ids_host.data());
weights_dev.ToDevice(weights_host.data());
......@@ -118,11 +149,15 @@ bool test_moe_sorting(ck_tile::ArgParser args)
{
moe_buf_dev.ToDevice(moe_buf_host.data());
}
if(local_expert_masking)
local_expert_masking_dev.ToDevice(local_expert_masking_host.data());
moe_sorting_trait trait{index_prec, weight_prec};
moe_sorting_trait trait{index_prec, weight_prec, local_expert_masking};
moe_sorting_args karg{topk_ids_dev.GetDeviceBuffer(),
weights_dev.GetDeviceBuffer(),
local_expert_masking ? local_expert_masking_dev.GetDeviceBuffer()
: nullptr,
sorted_ids_dev.GetDeviceBuffer(),
sorted_weights_dev.GetDeviceBuffer(),
sorted_expert_ids_dev.GetDeviceBuffer(),
......@@ -140,15 +175,22 @@ bool test_moe_sorting(ck_tile::ArgParser args)
warmup,
repeat};
auto ms = moe_sorting(trait, karg, sc);
printf("[%s|%s]tokens:%d, num_experts:%d, topk:%d, ms:%f , ",
printf("[%s|%s]tokens:%d, num_experts:%d, topk:%d, ",
index_prec.c_str(),
weight_prec.c_str(),
tokens,
num_experts,
topk,
ms);
topk);
if(local_expert_masking)
{
printf("local_eid:%s, ", args.get_str("local_eid").c_str());
}
if(ms < 0)
printf("not supported\n");
else
printf("ms:%f, ", ms);
fflush(stdout);
if(ms < 0)
{
......@@ -174,12 +216,14 @@ bool test_moe_sorting(ck_tile::ArgParser args)
int32_t ref_total_tokens_post_pad = 0;
ck_tile::reference_moe_sorting<WeightType, IndexType>(topk_ids_host,
weights_host,
local_expert_masking_host,
sorted_ids_ref,
sorted_weights_ref,
sorted_expert_ids_ref,
ref_total_tokens_post_pad,
num_experts,
unit_size);
unit_size,
local_expert_masking);
rtn &= ck_tile::check_err(
sorted_ids_host, sorted_ids_ref, std::string("OUT Error: Incorrect ids!"), 1e-6, 1e-6);
rtn &= ck_tile::check_err(sorted_weights_host,
......@@ -199,9 +243,16 @@ bool test_moe_sorting(ck_tile::ArgParser args)
moe_buf_host, moe_buf_ref, std::string("OUT Error: Incorrect zero buf!"), 0, 0);
}
rtn &= ref_total_tokens_post_pad == sorted_id_cnt_host.mData[0];
printf("total_tokens_post_pad:%d(%d), ",
ref_total_tokens_post_pad,
sorted_id_cnt_host.mData[0]);
}
printf("valid:%s\n", rtn ? "y" : "n");
printf("valid:%s", rtn ? "y" : "n");
fflush(stdout);
if(!rtn)
printf(", (%d)", seed);
printf("\n");
fflush(stdout);
return rtn;
}
......
example/ck_tile/13_moe_sorting/moe_sorting_api.cpp
View file @ 2a30cfdd
......@@ -3,22 +3,115 @@
#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)); \
#ifndef MOE_SORTING_USE_EX_KERNEL
#define MOE_SORTING_USE_EX_KERNEL 1
#endif
#if !MOE_SORTING_USE_EX_KERNEL
#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;
#else
#define MOE_SORTING_DISPATCH_(sub_token_tile_, sub_token_onshot_, local_expert_masking_) \
constexpr ck_tile::index_t sub_token_tile = sub_token_tile_; \
constexpr bool sub_token_onshot = sub_token_onshot_; \
constexpr bool local_expert_masking = local_expert_masking_; \
using ms_problem = ck_tile::MoeSortingProblemEx<index_t, \
ms_weight_type, \
sub_token_tile, \
sub_token_onshot, \
local_expert_masking>; \
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_SUB_TOKEN_(row_, sub_token_onshot_, local_expert_masking_) \
if(row_ % 8 == 0) \
{ \
MOE_SORTING_DISPATCH_(8, sub_token_onshot_, local_expert_masking_); \
} \
else if(row_ % 4 == 0) \
{ \
MOE_SORTING_DISPATCH_(4, sub_token_onshot_, local_expert_masking_); \
} \
else if(row_ % 2 == 0) \
{ \
MOE_SORTING_DISPATCH_(2, sub_token_onshot_, local_expert_masking_); \
} \
else \
{ \
MOE_SORTING_DISPATCH_(1, sub_token_onshot_, local_expert_masking_); \
}
#define MOE_SORTING_DISPATCH_SUBTO_(row_, local_expert_masking_) \
if(is_sub_token_onshot) \
{ \
MOE_SORTING_DISPATCH_SUB_TOKEN_(row_, true, local_expert_masking_) \
} \
else \
{ \
MOE_SORTING_DISPATCH_SUB_TOKEN_(row_, false, local_expert_masking_) \
}
#define MOE_SORTING_DISPATCH_EMASK_(row_) \
if(is_local_expert_masking) \
{ \
MOE_SORTING_DISPATCH_SUBTO_(row_, true) \
} \
else \
{ \
MOE_SORTING_DISPATCH_SUBTO_(row_, false) \
}
#endif
#if !MOE_SORTING_USE_EX_KERNEL
#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) \
}
#endif
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")
{
#if !MOE_SORTING_USE_EX_KERNEL
if(a.num_experts > 127)
{
printf("lds size exceed, only support experts <127 \n");
......@@ -49,25 +142,29 @@ 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);
}
}
#else
using index_t = ck_tile::index_t;
using ms_weight_type = float;
auto [r_, c_] = ck_tile::moe_sorting_get_smem_row_col(a.tokens, a.num_experts);
auto sub_token_ = r_ - 2;
r_ = (r_ - 2) / 8;
bool is_sub_token_onshot = a.tokens <= sub_token_;
bool is_local_expert_masking = t.local_expert_masking;
(void)c_;
MOE_SORTING_DISPATCH_EMASK_(r_);
// MOE_SORTING_DISPATCH_ETILE(0, 0);
#endif
}
return -1;
}
example/ck_tile/13_moe_sorting/moe_sorting_api.hpp
View file @ 2a30cfdd
......@@ -10,7 +10,8 @@
struct moe_sorting_trait
{
std::string index_type;
std::string weight_type; // currently always float
std::string weight_type; // currently always float
bool local_expert_masking; // if mask experts as local expert
};
struct moe_sorting_args : public ck_tile::MoeSortingHostArgs
......
example/ck_tile/13_moe_sorting/script/smoke_test.sh
View file @ 2a30cfdd
......@@ -16,4 +16,13 @@ $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=11 -e=256 -k=5
$EXE -t=64 -e=455 -k=8
$EXE -t=777 -e=802 -k=99
$EXE -t=4097 -e=906 -k=51
$EXE -t=128 -e=32 -k=5 -moe_buf_size=262144
$EXE -t=13 -e=64 -k=3 -local_eid=4,5,6,7,8,9,10,11
$EXE -t=99 -e=33 -k=9 -local_eid=6,10,11,15,19
$EXE -t=80 -e=99 -k=10 -local_eid=0,8,12,33
$EXE -t=11 -e=256 -k=5 -local_eid=99,110,129
example/ck_tile/14_moe_smoothquant/instances/moe_smoothquant_bf16_n1024_instance.cpp
View file @ 2a30cfdd
......@@ -15,8 +15,13 @@ template float moe_smoothquant_<trait_<ck_tile::bf16_t, 1, 16, 4, 64, 1, true
template float moe_smoothquant_<trait_<ck_tile::bf16_t, 1, 1, 1, 256, 4, true, false>>(const S&, A);
#endif
template float moe_smoothquant_<trait_<ck_tile::bf16_t, 1, 1, 2, 128, 8, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, 1, 2, 2, 128, 4, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, 1, 4, 2, 128, 2, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, 1, 4, 1, 256, 1, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, ck_tile::int8_t, 1, 1, 2, 128, 8, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, ck_tile::int8_t, 1, 2, 2, 128, 4, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, ck_tile::int8_t, 1, 4, 2, 128, 2, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, ck_tile::int8_t, 1, 4, 1, 256, 1, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, ck_tile::fp8_t, 1, 1, 2, 128, 8, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, ck_tile::fp8_t, 1, 2, 2, 128, 4, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, ck_tile::fp8_t, 1, 4, 2, 128, 2, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, ck_tile::fp8_t, 1, 4, 1, 256, 1, true, false>>(const S&, A);
// clang-format on
example/ck_tile/14_moe_smoothquant/instances/moe_smoothquant_bf16_n1536_instance.cpp
View file @ 2a30cfdd
......@@ -6,8 +6,13 @@
// clang-format off
// rm rn tm tn vn pd 2p
template float moe_smoothquant_<trait_<ck_tile::bf16_t, 1, 3, 4, 64, 8, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, 1, 3, 2, 128, 4, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, 1, 3, 1, 256, 2, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, 1, 6, 1, 256, 1, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, ck_tile::int8_t, 1, 3, 4, 64, 8, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, ck_tile::int8_t, 1, 3, 2, 128, 4, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, ck_tile::int8_t, 1, 3, 1, 256, 2, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, ck_tile::int8_t, 1, 6, 1, 256, 1, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, ck_tile::fp8_t, 1, 3, 4, 64, 8, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, ck_tile::fp8_t, 1, 3, 2, 128, 4, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, ck_tile::fp8_t, 1, 3, 1, 256, 2, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, ck_tile::fp8_t, 1, 6, 1, 256, 1, true, false>>(const S&, A);
// clang-format on
example/ck_tile/14_moe_smoothquant/instances/moe_smoothquant_bf16_n2048_instance.cpp
View file @ 2a30cfdd
......@@ -6,9 +6,14 @@
// clang-format off
// rm rn tm tn vn pd 2p
template float moe_smoothquant_<trait_<ck_tile::bf16_t, 1, 1, 1, 256, 8, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, 1, 2, 1, 256, 4, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, 1, 4, 1, 256, 2, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, 1, 8, 1, 256, 1, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, ck_tile::int8_t, 1, 1, 1, 256, 8, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, ck_tile::int8_t, 1, 2, 1, 256, 4, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, ck_tile::int8_t, 1, 4, 1, 256, 2, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, ck_tile::int8_t, 1, 8, 1, 256, 1, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, ck_tile::fp8_t, 1, 1, 1, 256, 8, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, ck_tile::fp8_t, 1, 2, 1, 256, 4, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, ck_tile::fp8_t, 1, 4, 1, 256, 2, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, ck_tile::fp8_t, 1, 8, 1, 256, 1, true, false>>(const S&, A);
// clang-format on
example/ck_tile/14_moe_smoothquant/instances/moe_smoothquant_bf16_n256_instance.cpp
View file @ 2a30cfdd
......@@ -6,7 +6,11 @@
// clang-format off
// rm rn tm tn vn pd 2p
template float moe_smoothquant_<trait_<ck_tile::bf16_t, 1, 1, 4, 64, 4, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, 1, 2, 4, 64, 2, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, 1, 4, 4, 64, 1, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, ck_tile::int8_t, 1, 1, 4, 64, 4, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, ck_tile::int8_t, 1, 2, 4, 64, 2, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, ck_tile::int8_t, 1, 4, 4, 64, 1, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, ck_tile::fp8_t, 1, 1, 4, 64, 4, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, ck_tile::fp8_t, 1, 2, 4, 64, 2, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, ck_tile::fp8_t, 1, 4, 4, 64, 1, true, false>>(const S&, A);
// clang-format on
example/ck_tile/14_moe_smoothquant/instances/moe_smoothquant_bf16_n3072_instance.cpp
View file @ 2a30cfdd
......@@ -6,9 +6,13 @@
// clang-format off
// rm rn tm tn vn pd 2p
template float moe_smoothquant_<trait_<ck_tile::bf16_t, 1, 3, 1, 128, 8, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, 1, 3, 1, 256, 4, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, 1, 6, 1, 256, 2, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, 1, 3, 1, 1024, 1, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, ck_tile::int8_t, 1, 3, 1, 128, 8, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, ck_tile::int8_t, 1, 3, 1, 256, 4, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, ck_tile::int8_t, 1, 6, 1, 256, 2, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, ck_tile::int8_t, 1, 3, 1, 1024, 1, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, ck_tile::fp8_t, 1, 3, 1, 128, 8, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, ck_tile::fp8_t, 1, 3, 1, 256, 4, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, ck_tile::fp8_t, 1, 6, 1, 256, 2, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, ck_tile::fp8_t, 1, 3, 1, 1024, 1, true, false>>(const S&, A);
// clang-format on
example/ck_tile/14_moe_smoothquant/instances/moe_smoothquant_bf16_n4096_instance.cpp
View file @ 2a30cfdd
......@@ -6,9 +6,13 @@
// clang-format off
// rm rn tm tn vn pd 2p
template float moe_smoothquant_<trait_<ck_tile::bf16_t, 1, 2, 1, 256, 8, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, 1, 4, 1, 256, 4, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, 1, 2, 1, 1024, 2, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, 1, 4, 1, 1024, 1, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, ck_tile::int8_t, 1, 2, 1, 256, 8, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, ck_tile::int8_t, 1, 4, 1, 256, 4, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, ck_tile::int8_t, 1, 2, 1, 1024, 2, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, ck_tile::int8_t, 1, 4, 1, 1024, 1, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, ck_tile::fp8_t, 1, 2, 1, 256, 8, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, ck_tile::fp8_t, 1, 4, 1, 256, 4, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, ck_tile::fp8_t, 1, 2, 1, 1024, 2, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, ck_tile::fp8_t, 1, 4, 1, 1024, 1, true, false>>(const S&, A);
// clang-format on
example/ck_tile/14_moe_smoothquant/instances/moe_smoothquant_bf16_n4096_tp_instance.cpp
View file @ 2a30cfdd
......@@ -6,9 +6,13 @@
// clang-format off
// rm rn tm tn vn pd 2p
template float moe_smoothquant_<trait_<ck_tile::bf16_t, 1, 2, 1, 256, 8, true, true>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, 1, 4, 1, 256, 4, true, true>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, 1, 2, 1, 1024, 2, true, true>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, 1, 4, 1, 1024, 1, true, true>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, ck_tile::int8_t, 1, 2, 1, 256, 8, true, true>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, ck_tile::int8_t, 1, 4, 1, 256, 4, true, true>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, ck_tile::int8_t, 1, 2, 1, 1024, 2, true, true>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, ck_tile::int8_t, 1, 4, 1, 1024, 1, true, true>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, ck_tile::fp8_t, 1, 2, 1, 256, 8, true, true>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, ck_tile::fp8_t, 1, 4, 1, 256, 4, true, true>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, ck_tile::fp8_t, 1, 2, 1, 1024, 2, true, true>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, ck_tile::fp8_t, 1, 4, 1, 1024, 1, true, true>>(const S&, A);
// clang-format on
example/ck_tile/14_moe_smoothquant/instances/moe_smoothquant_bf16_n512_instance.cpp
View file @ 2a30cfdd
......@@ -6,8 +6,13 @@
// clang-format off
// rm rn tm tn vn pd 2p
template float moe_smoothquant_<trait_<ck_tile::bf16_t, 1, 1, 4, 64, 8, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, 1, 2, 4, 64, 4, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, 1, 4, 4, 64, 2, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, 1, 8, 4, 64, 1, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, ck_tile::int8_t, 1, 1, 4, 64, 8, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, ck_tile::int8_t, 1, 2, 4, 64, 4, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, ck_tile::int8_t, 1, 4, 4, 64, 2, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, ck_tile::int8_t, 1, 8, 4, 64, 1, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, ck_tile::fp8_t, 1, 1, 4, 64, 8, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, ck_tile::fp8_t, 1, 2, 4, 64, 4, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, ck_tile::fp8_t, 1, 4, 4, 64, 2, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, ck_tile::fp8_t, 1, 8, 4, 64, 1, true, false>>(const S&, A);
// clang-format on
example/ck_tile/14_moe_smoothquant/instances/moe_smoothquant_bf16_n64_n128_instance.cpp
View file @ 2a30cfdd
......@@ -6,7 +6,11 @@
// clang-format off
// rm rn tm tn vn pd 2p
template float moe_smoothquant_<trait_<ck_tile::bf16_t, 1, 1, 4, 64, 1, true , false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, 1, 1, 4, 64, 2, true , false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, 1, 2, 4, 64, 1, true , false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, ck_tile::int8_t, 1, 1, 4, 64, 1, true , false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, ck_tile::int8_t, 1, 1, 4, 64, 2, true , false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, ck_tile::int8_t, 1, 2, 4, 64, 1, true , false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, ck_tile::fp8_t, 1, 1, 4, 64, 1, true , false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, ck_tile::fp8_t, 1, 1, 4, 64, 2, true , false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, ck_tile::fp8_t, 1, 2, 4, 64, 1, true , false>>(const S&, A);
// clang-format on
example/ck_tile/14_moe_smoothquant/instances/moe_smoothquant_bf16_n768_instance.cpp
View file @ 2a30cfdd
......@@ -6,7 +6,11 @@
// clang-format off
// rm rn tm tn vn pd 2p
template float moe_smoothquant_<trait_<ck_tile::bf16_t, 1, 3, 4, 64, 4, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, 1, 6, 4, 64, 2, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, 1, 12, 4, 64, 1, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, ck_tile::int8_t, 1, 3, 4, 64, 4, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, ck_tile::int8_t, 1, 6, 4, 64, 2, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, ck_tile::int8_t, 1, 12, 4, 64, 1, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, ck_tile::fp8_t, 1, 3, 4, 64, 4, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, ck_tile::fp8_t, 1, 6, 4, 64, 2, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, ck_tile::fp8_t, 1, 12, 4, 64, 1, true, false>>(const S&, A);
// clang-format on
example/ck_tile/14_moe_smoothquant/instances/moe_smoothquant_fp16_n1024_instance.cpp
View file @ 2a30cfdd
......@@ -15,8 +15,13 @@ template float moe_smoothquant_<trait_<ck_tile::fp16_t, 1, 16, 4, 64, 1, true
template float moe_smoothquant_<trait_<ck_tile::fp16_t, 1, 1, 1, 256, 4, true ,false>>(const S&, A);
#endif
template float moe_smoothquant_<trait_<ck_tile::fp16_t, 1, 1, 2, 128, 8, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::fp16_t, 1, 2, 2, 128, 4, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::fp16_t, 1, 4, 2, 128, 2, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::fp16_t, 1, 4, 1, 256, 1, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::fp16_t, ck_tile::int8_t, 1, 1, 2, 128, 8, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::fp16_t, ck_tile::int8_t, 1, 2, 2, 128, 4, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::fp16_t, ck_tile::int8_t, 1, 4, 2, 128, 2, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::fp16_t, ck_tile::int8_t, 1, 4, 1, 256, 1, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::fp16_t, ck_tile::fp8_t, 1, 1, 2, 128, 8, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::fp16_t, ck_tile::fp8_t, 1, 2, 2, 128, 4, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::fp16_t, ck_tile::fp8_t, 1, 4, 2, 128, 2, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::fp16_t, ck_tile::fp8_t, 1, 4, 1, 256, 1, true, false>>(const S&, A);
// clang-format on
example/ck_tile/14_moe_smoothquant/instances/moe_smoothquant_fp16_n1536_instance.cpp
View file @ 2a30cfdd
......@@ -6,8 +6,13 @@
// clang-format off
// rm rn tm tn vn pd 2p
template float moe_smoothquant_<trait_<ck_tile::fp16_t, 1, 3, 4, 64, 8, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::fp16_t, 1, 3, 2, 128, 4, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::fp16_t, 1, 3, 1, 256, 2, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::fp16_t, 1, 6, 1, 256, 1, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::fp16_t, ck_tile::int8_t, 1, 3, 4, 64, 8, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::fp16_t, ck_tile::int8_t, 1, 3, 2, 128, 4, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::fp16_t, ck_tile::int8_t, 1, 3, 1, 256, 2, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::fp16_t, ck_tile::int8_t, 1, 6, 1, 256, 1, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::fp16_t, ck_tile::fp8_t, 1, 3, 4, 64, 8, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::fp16_t, ck_tile::fp8_t, 1, 3, 2, 128, 4, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::fp16_t, ck_tile::fp8_t, 1, 3, 1, 256, 2, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::fp16_t, ck_tile::fp8_t, 1, 6, 1, 256, 1, true, false>>(const S&, A);
// clang-format on
example/ck_tile/14_moe_smoothquant/instances/moe_smoothquant_fp16_n2048_instance.cpp
View file @ 2a30cfdd
......@@ -6,9 +6,13 @@
// clang-format off
// rm rn tm tn vn pd 2p
template float moe_smoothquant_<trait_<ck_tile::fp16_t, 1, 1, 1, 256, 8, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::fp16_t, 1, 2, 1, 256, 4, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::fp16_t, 1, 4, 1, 256, 2, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::fp16_t, 1, 8, 1, 256, 1, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::fp16_t, ck_tile::int8_t, 1, 1, 1, 256, 8, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::fp16_t, ck_tile::int8_t, 1, 2, 1, 256, 4, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::fp16_t, ck_tile::int8_t, 1, 4, 1, 256, 2, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::fp16_t, ck_tile::int8_t, 1, 8, 1, 256, 1, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::fp16_t, ck_tile::fp8_t, 1, 1, 1, 256, 8, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::fp16_t, ck_tile::fp8_t, 1, 2, 1, 256, 4, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::fp16_t, ck_tile::fp8_t, 1, 4, 1, 256, 2, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::fp16_t, ck_tile::fp8_t, 1, 8, 1, 256, 1, true, false>>(const S&, A);
// clang-format on
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