Skip to content

GitLab

Commit 487826b3 authored by aska-0096's avatar aska-0096
Browse files

tempsave, fp8 sanity error

parent 50e10656
Hide whitespace changes
Inline Side-by-side
Showing with 71 additions and 30 deletions
example/ck_tile/11_add_rmsnorm2d_rdquant/instances/add_rmsnorm2d_rdquant_fwd_fp16_n512_instance.cpp
View file @ 487826b3
......@@ -6,8 +6,12 @@
// clang-format off
// rm rn tm tn vn pd x 3p
template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t, 1, 1, 4, 64, 8, true , true, false>>(const S&, A);
template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t, 1, 2, 4, 64, 4, true , true, false>>(const S&, A);
template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t, 1, 4, 4, 64, 2, true , true, false>>(const S&, A);
template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t, 1, 8, 4, 64, 1, true , true, false>>(const S&, A);
template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t, ck_tile::int8_t, 1, 1, 4, 64, 8, true , true, false>>(const S&, A);
template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t, ck_tile::int8_t, 1, 2, 4, 64, 4, true , true, false>>(const S&, A);
template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t, ck_tile::int8_t, 1, 4, 4, 64, 2, true , true, false>>(const S&, A);
template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t, ck_tile::int8_t, 1, 8, 4, 64, 1, true , true, false>>(const S&, A);
template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t, ck_tile::fp8_t, 1, 1, 4, 64, 8, true , true, false>>(const S&, A);
template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t, ck_tile::fp8_t, 1, 2, 4, 64, 4, true , true, false>>(const S&, A);
template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t, ck_tile::fp8_t, 1, 4, 4, 64, 2, true , true, false>>(const S&, A);
template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t, ck_tile::fp8_t, 1, 8, 4, 64, 1, true , true, false>>(const S&, A);
// clang-format on
example/ck_tile/11_add_rmsnorm2d_rdquant/instances/add_rmsnorm2d_rdquant_fwd_fp16_n64_n128_instance.cpp
View file @ 487826b3
......@@ -5,8 +5,11 @@
#include "add_rmsnorm2d_rdquant_fwd_instance_common.hpp"
// clang-format off
// rm rn tm tn vn pd x 3p
template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t, 1, 1, 4, 64, 1, true , true, false>>(const S&, A);
template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t, 1, 1, 4, 64, 2, true , true, false>>(const S&, A);
template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t, 1, 2, 4, 64, 1, true , true, false>>(const S&, A);
// rm rn tm tn vn pd x 3p
template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t, ck_tile::int8_t, 1, 1, 4, 64, 1, true , true, false>>(const S&, A);
template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t, ck_tile::int8_t, 1, 1, 4, 64, 2, true , true, false>>(const S&, A);
template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t, ck_tile::int8_t, 1, 2, 4, 64, 1, true , true, false>>(const S&, A);
template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t, ck_tile::fp8_t, 1, 1, 4, 64, 1, true , true, false>>(const S&, A);
template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t, ck_tile::fp8_t, 1, 1, 4, 64, 2, true , true, false>>(const S&, A);
template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t, ck_tile::fp8_t, 1, 2, 4, 64, 1, true , true, false>>(const S&, A);
// clang-format on
example/ck_tile/11_add_rmsnorm2d_rdquant/instances/add_rmsnorm2d_rdquant_fwd_fp16_n768_instance.cpp
View file @ 487826b3
......@@ -6,7 +6,10 @@
// clang-format off
// rm rn tm tn vn pd x 3p
template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t, 1, 3, 4, 64, 4, true , true, false>>(const S&, A);
template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t, 1, 6, 4, 64, 2, true , true, false>>(const S&, A);
template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t, 1, 12, 4, 64, 1, true , true, false>>(const S&, A);
template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t, ck_tile::int8_t, 1, 3, 4, 64, 4, true , true, false>>(const S&, A);
template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t, ck_tile::int8_t, 1, 6, 4, 64, 2, true , true, false>>(const S&, A);
template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t, ck_tile::int8_t, 1, 12, 4, 64, 1, true , true, false>>(const S&, A);
template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t, ck_tile::fp8_t, 1, 3, 4, 64, 4, true , true, false>>(const S&, A);
template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t, ck_tile::fp8_t, 1, 6, 4, 64, 2, true , true, false>>(const S&, A);
template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t, ck_tile::fp8_t, 1, 12, 4, 64, 1, true , true, false>>(const S&, A);
// clang-format on
example/ck_tile/11_add_rmsnorm2d_rdquant/instances/add_rmsnorm2d_rdquant_fwd_fp16_n8192_instance.cpp
View file @ 487826b3
......@@ -6,9 +6,29 @@
// clang-format off
// rm rn tm tn vn pd x 3p
template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t, 1, 4, 1, 256, 8, true, true, false>>(const S&, A);
template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t, 1, 8, 1, 256, 4, true, true, false>>(const S&, A);
template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t, 1, 4, 1, 1024, 2, true, true, false>>(const S&, A);
template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t, 1, 8, 1, 1024, 1, true, true, false>>(const S&, A);
template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t, ck_tile::int8_t, 1, 4, 1, 256, 8, true, true, false>>(const S&, A);
template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t, ck_tile::int8_t, 1, 8, 1, 256, 4, true, true, false>>(const S&, A);
template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t, ck_tile::int8_t, 1, 4, 1, 1024, 2, true, true, false>>(const S&, A);
template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t, ck_tile::int8_t, 1, 8, 1, 1024, 1, true, true, false>>(const S&, A);
template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t, ck_tile::int8_t, 1, 4, 1, 256, 8, true, false, false>>(const S&, A);
template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t, ck_tile::int8_t, 1, 8, 1, 256, 4, true, false, false>>(const S&, A);
template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t, ck_tile::int8_t, 1, 4, 1, 1024, 2, true, false, false>>(const S&, A);
template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t, ck_tile::int8_t, 1, 8, 1, 1024, 1, true, false, false>>(const S&, A);
template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t, ck_tile::int8_t, 1, 4, 1, 256, 8, false, false, false>>(const S&, A);
template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t, ck_tile::int8_t, 1, 8, 1, 256, 4, false, false, false>>(const S&, A);
template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t, ck_tile::int8_t, 1, 4, 1, 1024, 2, false, false, false>>(const S&, A);
template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t, ck_tile::int8_t, 1, 8, 1, 1024, 1, false, false, false>>(const S&, A);
template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t, ck_tile::fp8_t, 1, 4, 1, 256, 8, true, true, false>>(const S&, A);
template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t, ck_tile::fp8_t, 1, 8, 1, 256, 4, true, true, false>>(const S&, A);
template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t, ck_tile::fp8_t, 1, 4, 1, 1024, 2, true, true, false>>(const S&, A);
template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t, ck_tile::fp8_t, 1, 8, 1, 1024, 1, true, true, false>>(const S&, A);
template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t, ck_tile::fp8_t, 1, 4, 1, 256, 8, true, false, false>>(const S&, A);
template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t, ck_tile::fp8_t, 1, 8, 1, 256, 4, true, false, false>>(const S&, A);
template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t, ck_tile::fp8_t, 1, 4, 1, 1024, 2, true, false, false>>(const S&, A);
template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t, ck_tile::fp8_t, 1, 8, 1, 1024, 1, true, false, false>>(const S&, A);
template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t, ck_tile::fp8_t, 1, 4, 1, 256, 8, false, false, false>>(const S&, A);
template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t, ck_tile::fp8_t, 1, 8, 1, 256, 4, false, false, false>>(const S&, A);
template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t, ck_tile::fp8_t, 1, 4, 1, 1024, 2, false, false, false>>(const S&, A);
template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t, ck_tile::fp8_t, 1, 8, 1, 1024, 1, false, false, false>>(const S&, A);
// clang-format on
example/ck_tile/11_add_rmsnorm2d_rdquant/instances/add_rmsnorm2d_rdquant_fwd_fp16_n8192_tp_instance.cpp
View file @ 487826b3
......@@ -6,9 +6,12 @@
// clang-format off
// rm rn tm tn vn pd x 3p
template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t, 1, 4, 1, 256, 8, true, true, true>>(const S&, A);
template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t, 1, 8, 1, 256, 4, true, true, true>>(const S&, A);
template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t, 1, 4, 1, 1024, 2, true, true, true>>(const S&, A);
template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t, 1, 8, 1, 1024, 1, true, true, true>>(const S&, A);
template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t, ck_tile::int8_t, 1, 4, 1, 256, 8, true, true, true>>(const S&, A);
template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t, ck_tile::int8_t, 1, 8, 1, 256, 4, true, true, true>>(const S&, A);
template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t, ck_tile::int8_t, 1, 4, 1, 1024, 2, true, true, true>>(const S&, A);
template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t, ck_tile::int8_t, 1, 8, 1, 1024, 1, true, true, true>>(const S&, A);
template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t, ck_tile::fp8_t, 1, 4, 1, 256, 8, true, true, true>>(const S&, A);
template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t, ck_tile::fp8_t, 1, 8, 1, 256, 4, true, true, true>>(const S&, A);
template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t, ck_tile::fp8_t, 1, 4, 1, 1024, 2, true, true, true>>(const S&, A);
template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t, ck_tile::fp8_t, 1, 8, 1, 1024, 1, true, true, true>>(const S&, A);
// clang-format on
example/ck_tile/11_add_rmsnorm2d_rdquant/instances/add_rmsnorm2d_rdquant_fwd_instance_common.hpp
View file @ 487826b3
......@@ -11,7 +11,8 @@
using S = ck_tile::stream_config;
using A = add_rmsnorm2d_rdquant_fwd_args;
template <typename DataType_,
template <typename InputDataType_,
typename QuantizedDataType_,
ck_tile::index_t Repeat_M_, // each thread repeat along M
ck_tile::index_t Repeat_N_, // each thread repeat along N
ck_tile::index_t ThreadPerBlock_M_, // num threads along M
......@@ -20,7 +21,8 @@ template <typename DataType_,
bool kPadN_,
bool kSaveInvRms_,
bool kTwoPass_>
using trait_ = add_rmsnorm2d_rdquant_fwd_traits_<DataType_,
using trait_ = add_rmsnorm2d_rdquant_fwd_traits_<InputDataType_,
QuantizedDataType_,
Repeat_M_,
Repeat_N_,
ThreadPerBlock_M_,
......@@ -33,16 +35,17 @@ using trait_ = add_rmsnorm2d_rdquant_fwd_traits_<DataType_,
template <typename Traits_>
float add_rmsnorm2d_rdquant_fwd_(const S& s, A a)
{
using DataType = typename Traits_::DataType;
using InputDataType = typename Traits_::InputDataType;
using QuantizedDataType = typename Traits_::QuantizedDataType;
using PipelineProblem = ck_tile::AddRmsnorm2dRdquantFwdPipelineProblem<
typename AddRmsnormRdquantTypeConfig<DataType>::ADataType,
typename AddRmsnormRdquantTypeConfig<DataType>::BDataType,
typename AddRmsnormRdquantTypeConfig<DataType>::GammaDataType,
typename AddRmsnormRdquantTypeConfig<DataType>::ComputeDataType,
typename AddRmsnormRdquantTypeConfig<DataType>::XDataType,
typename AddRmsnormRdquantTypeConfig<DataType>::YScaleDataType,
typename AddRmsnormRdquantTypeConfig<DataType>::QYDataType,
typename AddRmsnormRdquantTypeConfig<InputDataType, QuantizedDataType>::ADataType,
typename AddRmsnormRdquantTypeConfig<InputDataType, QuantizedDataType>::BDataType,
typename AddRmsnormRdquantTypeConfig<InputDataType, QuantizedDataType>::GammaDataType,
typename AddRmsnormRdquantTypeConfig<InputDataType, QuantizedDataType>::ComputeDataType,
typename AddRmsnormRdquantTypeConfig<InputDataType, QuantizedDataType>::XDataType,
typename AddRmsnormRdquantTypeConfig<InputDataType, QuantizedDataType>::YScaleDataType,
typename AddRmsnormRdquantTypeConfig<InputDataType, QuantizedDataType>::QYDataType,
typename Traits_::Shape,
Traits_::kPadN,
Traits_::kSaveX,
......
include/ck_tile/host/reference/reference_rowwise_quantization2d.hpp
View file @ 487826b3
......@@ -23,6 +23,11 @@ CK_TILE_HOST void reference_rowwise_quantization2d(const HostTensor<XDataType>&
auto v_scale = type_convert<XDataType>(scale_m(m));
auto v_qx = v_x / v_scale;
qx_m_n(m, n) = saturates<QXDataType>{}(v_qx);
if(m == 0 && n == 4)
printf("Qy: %lf, Satruates Qy: %lf\n",
type_convert<float>(v_qx),
type_convert<float>(qx_m_n(m, n)));
}
};
......
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