Skip to content

GitLab

Commit 0286b9bf authored by danyao12's avatar danyao12
Browse files

adjust parameters to verify

parent f1a49daf
Hide whitespace changes
Inline Side-by-side
Showing with 15 additions and 15 deletions
example/32_batched_gemm_scale_softmax_gemm/batched_multihead_attention_backward.cpp
View file @ 0286b9bf
......@@ -25,7 +25,7 @@ Kernel outputs:
#define PRINT_HOST 0
#define USING_MASK 0
#define DIM 64 // DIM should be a multiple of 8.
#define DIM 128 // DIM should be a multiple of 8.
#include <iostream>
#include <numeric>
......@@ -63,9 +63,9 @@ using Scale = ck::tensor_operation::element_wise::Scale;
using QKVElementOp = PassThrough;
using YElementOp = PassThrough;
using InputDataType = BF16;
using OutputDataType = F32;
using GemmDataType = BF16;
using InputDataType = F16;
using OutputDataType = F16;
using GemmDataType = F16;
using AccDataType = F32;
using ShuffleDataType = F32;
using LSEDataType = F32;
......@@ -80,7 +80,7 @@ static constexpr ck::index_t NumDimK = 1;
static constexpr ck::index_t NumDimO = 1;
// When OutputDataType == F32, CShuffleBlockTransferScalarPerVector_NPerBlock = 4
// When OutputDataType == F16/BF16, CShuffleBlockTransferScalarPerVector_NPerBlock = 8
static constexpr ck::index_t CShuffleBlockTransferScalarPerVector_NPerBlock = 4;
static constexpr ck::index_t CShuffleBlockTransferScalarPerVector_NPerBlock = 8;
static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKOPadding;
#if USING_MASK
......@@ -95,7 +95,7 @@ static constexpr auto TensorSpecQ = ck::tensor_operation::device::TensorSpecia
static constexpr auto TensorSpecK = ck::tensor_operation::device::TensorSpecialization::Default;
static constexpr auto TensorSpecV = ck::tensor_operation::device::TensorSpecialization::Default;
static constexpr auto TensorSpecY = ck::tensor_operation::device::TensorSpecialization::Default;
static constexpr bool Deterministic = true;
static constexpr bool Deterministic = false;
// DIM should be a multiple of 8.
// If DIM <= 32 , ues prototype1 1st template.
......
example/32_batched_gemm_scale_softmax_gemm/batched_multihead_attention_train.cpp
View file @ 0286b9bf
......@@ -32,7 +32,7 @@ Kernel outputs:
#define PRINT_HOST 0
#define USING_MASK 0
#define DIM 64 // DIM should be a multiple of 8.
#define DIM 128 // DIM should be a multiple of 8.
#include <iostream>
#include <numeric>
......@@ -89,7 +89,7 @@ static constexpr ck::index_t NumDimK = 1;
static constexpr ck::index_t NumDimO = 1;
// When OutputDataType == F32, bwd CShuffleBlockTransferScalarPerVector_NPerBlock = 4
// When OutputDataType == F16/BF16, bwd CShuffleBlockTransferScalarPerVector_NPerBlock = 8
static constexpr ck::index_t CShuffleBlockTransferScalarPerVector_NPerBlock = 4;
static constexpr ck::index_t CShuffleBlockTransferScalarPerVector_NPerBlock = 8;
static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKOPadding;
#if USING_MASK
......@@ -104,7 +104,7 @@ static constexpr auto TensorSpecQ = ck::tensor_operation::device::TensorSpecia
static constexpr auto TensorSpecK = ck::tensor_operation::device::TensorSpecialization::Default;
static constexpr auto TensorSpecV = ck::tensor_operation::device::TensorSpecialization::Default;
static constexpr auto TensorSpecY = ck::tensor_operation::device::TensorSpecialization::Default;
static constexpr bool Deterministic = true;
static constexpr bool Deterministic = false;
// DIM should be a multiple of 8.
// If DIM <= 32 , ues prototype1 1st template.
......
example/32_batched_gemm_scale_softmax_gemm/grouped_multihead_attention_train.cpp
View file @ 0286b9bf
......@@ -31,7 +31,7 @@ Kernel outputs:
*/
#define USING_MASK 0
#define DIM 64 // DIM should be a multiple of 8.
#define DIM 128 // DIM should be a multiple of 8.
#include <iostream>
#include <numeric>
......@@ -71,9 +71,9 @@ using Scale = ck::tensor_operation::element_wise::Scale;
using QKVElementOp = PassThrough;
using YElementOp = PassThrough;
using InputDataType = BF16;
using OutputDataType = F32;
using GemmDataType = BF16;
using InputDataType = F16;
using OutputDataType = F16;
using GemmDataType = F16;
using AccDataType = F32;
using ShuffleDataType = F32;
using LSEDataType = F32;
......@@ -88,7 +88,7 @@ static constexpr ck::index_t NumDimK = 1;
static constexpr ck::index_t NumDimO = 1;
// When OutputDataType == F32, bwd CShuffleBlockTransferScalarPerVector_NPerBlock = 4
// When OutputDataType == F16/BF16, bwd CShuffleBlockTransferScalarPerVector_NPerBlock = 8
static constexpr ck::index_t CShuffleBlockTransferScalarPerVector_NPerBlock = 4;
static constexpr ck::index_t CShuffleBlockTransferScalarPerVector_NPerBlock = 8;
static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKOPadding;
#if USING_MASK
......@@ -103,7 +103,7 @@ static constexpr auto TensorSpecQ = ck::tensor_operation::device::TensorSpecia
static constexpr auto TensorSpecK = ck::tensor_operation::device::TensorSpecialization::Default;
static constexpr auto TensorSpecV = ck::tensor_operation::device::TensorSpecialization::Default;
static constexpr auto TensorSpecY = ck::tensor_operation::device::TensorSpecialization::Default;
static constexpr bool Deterministic = true;
static constexpr bool Deterministic = false;
// DIM should be a multiple of 8.
// If DIM <= 32 , ues prototype1 1st template.
......
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