Skip to content

GitLab

Unverified Commit 7ec8ed67 authored by Masaki Kozuki's avatar Masaki Kozuki Committed by GitHub
Browse files

Faster `--fast_multihead_attn` build (#1245) 

* merge .so files

* odr

* fix build

* update import

* apply psf/black with max line length of 120

* update

* fix

* update

* build fixed again but undefined symbol again

* fix 2, still layer norm grad is undefined

* remove unused cpp files

* without layer_norm.cuh, import works

* import fast_multihead_attn works...

but why? Was unnecessary `#include "layer_norm.cuh"` was the culprit
causing .shared objects not to be able to link `HostApplyLayerNorm` and
`HostLayerNormGradient`?

* clean up layer norm
parent ed94d0bb
Expand all Hide whitespace changes
Inline Side-by-side
Showing with 876 additions and 994 deletions
apex/contrib/csrc/fmha/src/fmha_kernel.h
View file @ 7ec8ed67
......@@ -27,7 +27,7 @@
#pragma once
#include <multihead_attn/philox.h>
#include <multihead_attn/philox.cuh>
#include <fmha.h>
#include <fmha/utils.h>
......
apex/contrib/csrc/multihead_attn/additive_masked_softmax_dropout.cpp deleted 100644 → 0
View file @ ed94d0bb
#include <cuda_fp16.h>
#include <torch/extension.h>
#include <vector>
namespace multihead_attn {
namespace fused_softmax {
namespace additive_mask_softmax_dropout {
std::vector<torch::Tensor> fwd_cuda(bool is_training, int heads,
torch::Tensor const &input,
const half *pad_mask, float dropout_prob);
torch::Tensor bwd_cuda(int heads, torch::Tensor const &output_grads,
torch::Tensor const &softmax_results,
torch::Tensor const &dropout_mask, float dropout_prob);
// C++ interface
#define CHECK_CUDA(x) \
AT_ASSERTM(x.type().is_cuda(), #x " must be a CUDA tensor")
#define CHECK_CONTIGUOUS(x) \
AT_ASSERTM(x.is_contiguous(), #x " must be contiguous")
#define CHECK_INPUT(x) \
CHECK_CUDA(x); \
CHECK_CONTIGUOUS(x)
std::vector<torch::Tensor> fwd(bool use_mask, bool is_training, int heads,
torch::Tensor const &input,
torch::Tensor const &pad_mask,
float dropout_prob) {
AT_ASSERTM(input.dim() == 3, "expected 3D tensor");
AT_ASSERTM(input.type().scalarType() == at::ScalarType::Half,
"Only HALF is supported");
if (use_mask) {
AT_ASSERTM(pad_mask.dim() == 2, "expected 2D tensor");
AT_ASSERTM(pad_mask.type().scalarType() == at::ScalarType::Half,
"Only BYTE is supported");
}
return fwd_cuda(is_training, heads, input,
use_mask ? static_cast<const half *>(pad_mask.data_ptr())
: nullptr,
dropout_prob);
}
torch::Tensor bwd(bool use_mask, int heads, torch::Tensor const &output_grads,
torch::Tensor const &softmax_results,
torch::Tensor const &dropout_mask, float dropout_prob) {
AT_ASSERTM(output_grads.dim() == 3, "expected 3D tensor");
AT_ASSERTM(softmax_results.dim() == 3, "expected 3D tensor");
AT_ASSERTM(dropout_mask.dim() == 3, "expected 3D tensor");
AT_ASSERTM(output_grads.type().scalarType() == at::ScalarType::Half,
"Only HALF is supported");
AT_ASSERTM(softmax_results.type().scalarType() == at::ScalarType::Half,
"Only HALF is supported");
// AT_ASSERTM(dropout_mask.type().scalarType() == at::ScalarType::Byte,
// "Only BYTE is supported");
return bwd_cuda(heads, output_grads, softmax_results, dropout_mask,
dropout_prob);
}
} // namespace additive_mask_softmax_dropout
} // end namespace fused_softmax
} // end namespace multihead_attn
PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
m.def("forward",
&multihead_attn::fused_softmax::additive_mask_softmax_dropout::fwd,
"Self Multihead Attention masked softmax dropout -- Forward.");
m.def("backward",
&multihead_attn::fused_softmax::additive_mask_softmax_dropout::bwd,
"Self Multihead Attention masked softmax dropout -- Backward.");
}
apex/contrib/csrc/multihead_attn/additive_masked_softmax_dropout_cuda.cu
View file @ 7ec8ed67
......@@ -11,8 +11,8 @@
#include <ATen/cuda/CUDAContext.h>
#include <torch/extension.h>
#include "dropout.h"
#include "softmax.h"
#include "dropout.cuh"
#include "softmax.cuh"
// symbol to be automatically resolved by PyTorch libs
......@@ -27,7 +27,7 @@ std::vector<torch::Tensor> fwd_cuda(bool is_training, int heads,
const int sequences = attn_batches / heads;
const int q_seq_len = input.size(1);
const int k_seq_len = q_seq_len;
const int dropout_elems = attn_batches * q_seq_len * k_seq_len;
// const int dropout_elems = attn_batches * q_seq_len * k_seq_len;
// There is no reason to use more than one stream as every kernel is
// sequentially dependent
......@@ -86,7 +86,7 @@ torch::Tensor bwd_cuda(int heads, torch::Tensor const &output_grads,
const int attn_batches = output_grads.size(0);
const int q_seq_len = output_grads.size(1);
const int k_seq_len = q_seq_len;
const int dropout_elems = attn_batches * q_seq_len * k_seq_len;
// const int dropout_elems = attn_batches * q_seq_len * k_seq_len;
// TODO: Streams can be used in Backprop but I haven't added more than one
// in my first attempt to create the code
cublasHandle_t handle = at::cuda::getCurrentCUDABlasHandle();
......
apex/contrib/csrc/multihead_attn/dropout.h → apex/contrib/csrc/multihead_attn/dropout.cuh
View file @ 7ec8ed67
#pragma once
#include <ATen/ATen.h>
#ifdef OLD_GENERATOR
......@@ -9,7 +10,9 @@
#include <ATen/cuda/CUDAContext.h>
#include <curand_kernel.h>
const int UNROLL = 4;
namespace {
constexpr int UNROLL = 4;
} // namespace
template <typename scalar_t, typename accscalar_t, typename IndexType>
__global__ void
......
apex/contrib/csrc/multihead_attn/encdec_multihead_attn.cpp deleted 100644 → 0
View file @ ed94d0bb
#include <torch/extension.h>
#include <vector>
namespace multihead_attn {
namespace encdec {
namespace cublas_gemmex {
std::vector<torch::Tensor> fwd_cuda(bool use_time_mask, bool is_training,
int heads, torch::Tensor const &inputs_q,
torch::Tensor const &inputs_kv,
torch::Tensor const &input_weights_q,
torch::Tensor const &input_weights_kv,
torch::Tensor const &output_weights,
const uint8_t *pad_mask,
float dropout_prob);
std::vector<torch::Tensor> bwd_cuda(
int heads, torch::Tensor const &output_grads,
torch::Tensor const &matmul2_results, torch::Tensor const &dropout_results,
torch::Tensor const &softmax_results,
torch::Tensor const &input_lin_q_results,
torch::Tensor const &input_lin_kv_results, torch::Tensor const &inputs_q,
torch::Tensor const &inputs_kv, torch::Tensor const &input_weights_q,
torch::Tensor const &input_weights_kv, torch::Tensor const &output_weights,
torch::Tensor const &dropout_mask, float dropout_prob);
// C++ interface
#define CHECK_CUDA(x) \
AT_ASSERTM(x.type().is_cuda(), #x " must be a CUDA tensor")
#define CHECK_CONTIGUOUS(x) \
AT_ASSERTM(x.is_contiguous(), #x " must be contiguous")
#define CHECK_INPUT(x) \
CHECK_CUDA(x); \
CHECK_CONTIGUOUS(x)
std::vector<torch::Tensor>
fwd(bool use_mask, bool use_time_mask, bool is_training, int heads,
torch::Tensor const &inputs_q, torch::Tensor const &inputs_kv,
torch::Tensor const &input_weights_q, torch::Tensor const &input_weights_kv,
torch::Tensor const &output_weights, torch::Tensor const &pad_mask,
float dropout_prob) {
AT_ASSERTM(inputs_q.dim() == 3, "expected 3D tensor");
AT_ASSERTM(inputs_kv.dim() == 3, "expected 3D tensor");
AT_ASSERTM(input_weights_q.dim() == 2, "expected 2D tensor");
AT_ASSERTM(input_weights_kv.dim() == 2, "expected 2D tensor");
AT_ASSERTM(output_weights.dim() == 2, "expected 2D tensor");
AT_ASSERTM(inputs_q.type().scalarType() == at::ScalarType::Half,
"Only HALF is supported");
AT_ASSERTM(inputs_kv.type().scalarType() == at::ScalarType::Half,
"Only HALF is supported");
AT_ASSERTM(input_weights_q.type().scalarType() == at::ScalarType::Half,
"Only HALF is supported");
AT_ASSERTM(input_weights_kv.type().scalarType() == at::ScalarType::Half,
"Only HALF is supported");
AT_ASSERTM(output_weights.type().scalarType() == at::ScalarType::Half,
"Only HALF is supported");
if (use_mask) {
AT_ASSERTM(pad_mask.dim() == 2, "expected 2D tensor");
AT_ASSERTM(pad_mask.type().scalarType() == at::ScalarType::Byte,
"Only BYTE is supported");
}
return fwd_cuda(use_time_mask, is_training, heads, inputs_q, inputs_kv,
input_weights_q, input_weights_kv, output_weights,
use_mask ? static_cast<const uint8_t *>(pad_mask.data_ptr())
: nullptr,
dropout_prob);
}
std::vector<torch::Tensor>
bwd(int heads, torch::Tensor const &output_grads,
torch::Tensor const &matmul2_results, torch::Tensor const &dropout_results,
torch::Tensor const &softmax_results,
torch::Tensor const &input_lin_q_results,
torch::Tensor const &input_lin_kv_results, torch::Tensor const &inputs_q,
torch::Tensor const &inputs_kv, torch::Tensor const &input_weights_q,
torch::Tensor const &input_weights_kv, torch::Tensor const &output_weights,
torch::Tensor const &dropout_mask, float dropout_prob) {
AT_ASSERTM(output_grads.dim() == 3, "expected 3D tensor");
AT_ASSERTM(matmul2_results.dim() == 3, "expected 3D tensor");
AT_ASSERTM(dropout_results.dim() == 3, "expected 3D tensor");
AT_ASSERTM(softmax_results.dim() == 3, "expected 3D tensor");
AT_ASSERTM(input_lin_q_results.dim() == 3, "expected 3D tensor");
AT_ASSERTM(input_lin_kv_results.dim() == 3, "expected 3D tensor");
AT_ASSERTM(inputs_q.dim() == 3, "expected 3D tensor");
AT_ASSERTM(inputs_kv.dim() == 3, "expected 3D tensor");
AT_ASSERTM(input_weights_q.dim() == 2, "expected 2D tensor");
AT_ASSERTM(input_weights_kv.dim() == 2, "expected 2D tensor");
AT_ASSERTM(output_weights.dim() == 2, "expected 2D tensor");
AT_ASSERTM(dropout_mask.dim() == 3, "expected 3D tensor");
AT_ASSERTM(output_grads.type().scalarType() == at::ScalarType::Half,
"Only HALF is supported");
AT_ASSERTM(matmul2_results.type().scalarType() == at::ScalarType::Half,
"Only HALF is supported");
AT_ASSERTM(dropout_results.type().scalarType() == at::ScalarType::Half,
"Only HALF is supported");
AT_ASSERTM(softmax_results.type().scalarType() == at::ScalarType::Half,
"Only HALF is supported");
AT_ASSERTM(input_lin_q_results.type().scalarType() == at::ScalarType::Half,
"Only HALF is supported");
AT_ASSERTM(input_lin_kv_results.type().scalarType() == at::ScalarType::Half,
"Only HALF is supported");
AT_ASSERTM(inputs_q.type().scalarType() == at::ScalarType::Half,
"Only HALF is supported");
AT_ASSERTM(inputs_kv.type().scalarType() == at::ScalarType::Half,
"Only HALF is supported");
AT_ASSERTM(input_weights_q.type().scalarType() == at::ScalarType::Half,
"Only HALF is supported");
AT_ASSERTM(input_weights_kv.type().scalarType() == at::ScalarType::Half,
"Only HALF is supported");
AT_ASSERTM(output_weights.type().scalarType() == at::ScalarType::Half,
"Only HALF is supported");
AT_ASSERTM(dropout_mask.type().scalarType() == at::ScalarType::Byte,
"Only BYTE is supported");
return bwd_cuda(heads, output_grads, matmul2_results, dropout_results,
softmax_results, input_lin_q_results, input_lin_kv_results,
inputs_q, inputs_kv, input_weights_q, input_weights_kv,
output_weights, dropout_mask, dropout_prob);
}
} // end namespace cublas_gemmex
} // end namespace encdec
} // end namespace multihead_attn
PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
m.def("forward", &multihead_attn::encdec::cublas_gemmex::fwd,
"Encdec Multihead Attention Forward.");
m.def("backward", &multihead_attn::encdec::cublas_gemmex::bwd,
"Encdec Multihead Attention Backward.");
}
apex/contrib/csrc/multihead_attn/encdec_multihead_attn_cuda.cu
View file @ 7ec8ed67
......@@ -11,10 +11,9 @@
#include <ATen/cuda/CUDAContext.h>
#include <torch/extension.h>
#include "dropout.h"
#include "layer_norm.h"
#include "softmax.h"
#include "strided_batched_gemm.h"
#include "dropout.cuh"
#include "softmax.cuh"
#include "strided_batched_gemm.cuh"
namespace multihead_attn {
namespace encdec {
......
apex/contrib/csrc/multihead_attn/encdec_multihead_attn_norm_add.cpp deleted 100644 → 0
View file @ ed94d0bb
#include <torch/extension.h>
#include <vector>
namespace multihead_attn {
namespace encdec_norm_add {
namespace cublas_gemmex {
std::vector<torch::Tensor> fwd_cuda(bool use_time_mask, bool is_training,
int heads, torch::Tensor const &inputs_q,
torch::Tensor const &inputs_kv,
torch::Tensor const &lyr_nrm_gamma_weights,
torch::Tensor const &lyr_nrm_beta_weights,
torch::Tensor const &input_weights_q,
torch::Tensor const &input_weights_kv,
torch::Tensor const &output_weights,
const uint8_t *pad_mask,
float dropout_prob);
std::vector<torch::Tensor> bwd_cuda(
int heads, torch::Tensor const &output_grads,
torch::Tensor const &matmul2_results, torch::Tensor const &dropout_results,
torch::Tensor const &softmax_results,
torch::Tensor const &input_lin_q_results,
torch::Tensor const &input_lin_kv_results,
torch::Tensor const &lyr_nrm_results, torch::Tensor const &lyr_nrm_mean,
torch::Tensor const &lyr_nrm_invvar, torch::Tensor const &inputs_q,
torch::Tensor const &inputs_kv, torch::Tensor const &lyr_nrm_gamma_weights,
torch::Tensor const &lyr_nrm_beta_weights,
torch::Tensor const &input_weights_q, torch::Tensor const &input_weights_kv,
torch::Tensor const &output_weights, torch::Tensor const &dropout_mask,
torch::Tensor const &dropout_add_mask, float dropout_prob);
// C++ interface
#define CHECK_CUDA(x) \
AT_ASSERTM(x.type().is_cuda(), #x " must be a CUDA tensor")
#define CHECK_CONTIGUOUS(x) \
AT_ASSERTM(x.is_contiguous(), #x " must be contiguous")
#define CHECK_INPUT(x) \
CHECK_CUDA(x); \
CHECK_CONTIGUOUS(x)
std::vector<torch::Tensor>
fwd(bool use_mask, bool use_time_mask, bool is_training, int heads,
torch::Tensor const &inputs_q, torch::Tensor const &inputs_kv,
torch::Tensor const &lyr_nrm_gamma_weights,
torch::Tensor const &lyr_nrm_beta_weights,
torch::Tensor const &input_weights_q, torch::Tensor const &input_weights_kv,
torch::Tensor const &output_weights, torch::Tensor const &pad_mask,
float dropout_prob) {
AT_ASSERTM(inputs_q.dim() == 3, "expected 3D tensor");
AT_ASSERTM(inputs_kv.dim() == 3, "expected 3D tensor");
AT_ASSERTM(lyr_nrm_gamma_weights.dim() == 1, "expected 1D tensor");
AT_ASSERTM(lyr_nrm_beta_weights.dim() == 1, "expected 1D tensor");
AT_ASSERTM(input_weights_q.dim() == 2, "expected 2D tensor");
AT_ASSERTM(input_weights_kv.dim() == 2, "expected 2D tensor");
AT_ASSERTM(output_weights.dim() == 2, "expected 2D tensor");
AT_ASSERTM(inputs_q.type().scalarType() == at::ScalarType::Half,
"Only HALF is supported");
AT_ASSERTM(inputs_kv.type().scalarType() == at::ScalarType::Half,
"Only HALF is supported");
AT_ASSERTM(lyr_nrm_gamma_weights.type().scalarType() == at::ScalarType::Half,
"Only HALF is supported");
AT_ASSERTM(lyr_nrm_beta_weights.type().scalarType() == at::ScalarType::Half,
"Only HALF is supported");
AT_ASSERTM(input_weights_q.type().scalarType() == at::ScalarType::Half,
"Only HALF is supported");
AT_ASSERTM(input_weights_kv.type().scalarType() == at::ScalarType::Half,
"Only HALF is supported");
AT_ASSERTM(output_weights.type().scalarType() == at::ScalarType::Half,
"Only HALF is supported");
if (use_mask) {
AT_ASSERTM(pad_mask.dim() == 2, "expected 2D tensor");
AT_ASSERTM(pad_mask.type().scalarType() == at::ScalarType::Byte,
"Only BYTE is supported");
}
return fwd_cuda(use_time_mask, is_training, heads, inputs_q, inputs_kv,
lyr_nrm_gamma_weights, lyr_nrm_beta_weights, input_weights_q,
input_weights_kv, output_weights,
use_mask ? static_cast<const uint8_t *>(pad_mask.data_ptr())
: nullptr,
dropout_prob);
}
std::vector<torch::Tensor>
bwd(int heads, torch::Tensor const &output_grads,
torch::Tensor const &matmul2_results, torch::Tensor const &dropout_results,
torch::Tensor const &softmax_results,
torch::Tensor const &input_lin_q_results,
torch::Tensor const &input_lin_kv_results,
torch::Tensor const &lyr_nrm_results, torch::Tensor const &lyr_nrm_mean,
torch::Tensor const &lyr_nrm_invvar, torch::Tensor const &inputs_q,
torch::Tensor const &inputs_kv, torch::Tensor const &lyr_nrm_gamma_weights,
torch::Tensor const &lyr_nrm_beta_weights,
torch::Tensor const &input_weights_q, torch::Tensor const &input_weights_kv,
torch::Tensor const &output_weights, torch::Tensor const &dropout_mask,
torch::Tensor const &dropout_add_mask, float dropout_prob) {
AT_ASSERTM(output_grads.dim() == 3, "expected 3D tensor");
AT_ASSERTM(matmul2_results.dim() == 3, "expected 3D tensor");
AT_ASSERTM(dropout_results.dim() == 3, "expected 3D tensor");
AT_ASSERTM(softmax_results.dim() == 3, "expected 3D tensor");
AT_ASSERTM(input_lin_q_results.dim() == 3, "expected 3D tensor");
AT_ASSERTM(input_lin_kv_results.dim() == 3, "expected 3D tensor");
AT_ASSERTM(lyr_nrm_results.dim() == 3, "expected 3D tensor");
AT_ASSERTM(lyr_nrm_mean.dim() == 1, "expected 1D tensor");
AT_ASSERTM(lyr_nrm_invvar.dim() == 1, "expected 1D tensor");
AT_ASSERTM(inputs_q.dim() == 3, "expected 3D tensor");
AT_ASSERTM(inputs_kv.dim() == 3, "expected 3D tensor");
AT_ASSERTM(lyr_nrm_gamma_weights.dim() == 1, "expected 1D tensor");
AT_ASSERTM(lyr_nrm_beta_weights.dim() == 1, "expected 1D tensor");
AT_ASSERTM(input_weights_q.dim() == 2, "expected 2D tensor");
AT_ASSERTM(input_weights_kv.dim() == 2, "expected 2D tensor");
AT_ASSERTM(output_weights.dim() == 2, "expected 2D tensor");
AT_ASSERTM(dropout_mask.dim() == 3, "expected 3D tensor");
AT_ASSERTM(dropout_add_mask.dim() == 3, "expected 3D tensor");
AT_ASSERTM(output_grads.type().scalarType() == at::ScalarType::Half,
"Only HALF is supported");
AT_ASSERTM(matmul2_results.type().scalarType() == at::ScalarType::Half,
"Only HALF is supported");
AT_ASSERTM(dropout_results.type().scalarType() == at::ScalarType::Half,
"Only HALF is supported");
AT_ASSERTM(softmax_results.type().scalarType() == at::ScalarType::Half,
"Only HALF is supported");
AT_ASSERTM(input_lin_q_results.type().scalarType() == at::ScalarType::Half,
"Only HALF is supported");
AT_ASSERTM(input_lin_kv_results.type().scalarType() == at::ScalarType::Half,
"Only HALF is supported");
AT_ASSERTM(lyr_nrm_results.type().scalarType() == at::ScalarType::Half,
"Only HALF is supported");
AT_ASSERTM(lyr_nrm_mean.type().scalarType() == at::ScalarType::Float,
"Only FLOAT is supported");
AT_ASSERTM(lyr_nrm_invvar.type().scalarType() == at::ScalarType::Float,
"Only FLOAT is supported");
AT_ASSERTM(inputs_q.type().scalarType() == at::ScalarType::Half,
"Only HALF is supported");
AT_ASSERTM(inputs_kv.type().scalarType() == at::ScalarType::Half,
"Only HALF is supported");
AT_ASSERTM(lyr_nrm_gamma_weights.type().scalarType() == at::ScalarType::Half,
"Only HALF is supported");
AT_ASSERTM(lyr_nrm_beta_weights.type().scalarType() == at::ScalarType::Half,
"Only HALF is supported");
AT_ASSERTM(input_weights_q.type().scalarType() == at::ScalarType::Half,
"Only HALF is supported");
AT_ASSERTM(input_weights_kv.type().scalarType() == at::ScalarType::Half,
"Only HALF is supported");
AT_ASSERTM(output_weights.type().scalarType() == at::ScalarType::Half,
"Only HALF is supported");
AT_ASSERTM(dropout_mask.type().scalarType() == at::ScalarType::Byte,
"Only BYTE is supported");
AT_ASSERTM(dropout_add_mask.type().scalarType() == at::ScalarType::Byte,
"Only BYTE is supported");
return bwd_cuda(heads, output_grads, matmul2_results, dropout_results,
softmax_results, input_lin_q_results, input_lin_kv_results,
lyr_nrm_results, lyr_nrm_mean, lyr_nrm_invvar, inputs_q,
inputs_kv, lyr_nrm_gamma_weights, lyr_nrm_beta_weights,
input_weights_q, input_weights_kv, output_weights,
dropout_mask, dropout_add_mask, dropout_prob);
}
} // end namespace cublas_gemmex
} // end namespace encdec_norm_add
} // end namespace multihead_attn
PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
m.def("forward", &multihead_attn::encdec_norm_add::cublas_gemmex::fwd,
"Encdec Multihead Attention Plus Layer Norm and Residual Add Forward.");
m.def(
"backward", &multihead_attn::encdec_norm_add::cublas_gemmex::bwd,
"Encdec Multihead Attention Plus Layer Norm and Residual Add Backward.");
}
apex/contrib/csrc/multihead_attn/encdec_multihead_attn_norm_add_cuda.cu
View file @ 7ec8ed67
......@@ -11,10 +11,10 @@
#include <ATen/cuda/CUDAContext.h>
#include <torch/extension.h>
#include "dropout.h"
#include "layer_norm.h"
#include "softmax.h"
#include "strided_batched_gemm.h"
#include "dropout.cuh"
#include "layer_norm.cuh"
#include "softmax.cuh"
#include "strided_batched_gemm.cuh"
namespace multihead_attn {
namespace encdec_norm_add {
......
apex/contrib/csrc/multihead_attn/layer_norm.h → apex/contrib/csrc/multihead_attn/layer_norm.cuh
View file @ 7ec8ed67
#include "ATen/ATen.h"
#include "ATen/cuda/DeviceUtils.cuh"
#pragma once
#include <cuda.h>
#include <cuda_runtime.h>
#include <ATen/ATen.h>
#include <ATen/cuda/DeviceUtils.cuh>
namespace {
template <typename U>
__device__ void cuWelfordOnlineSum(const U curr, U &mu, U &sigma2, U &count) {
count = count + U(1);
......@@ -203,11 +204,10 @@ __device__ void cuWelfordMuSigma2(const at::Half *__restrict__ vals,
}
}
template <typename U> U rsqrt(U v) { return U(1) / sqrt(v); }
template <> float rsqrt(float v) { return rsqrtf(v); }
template <> double rsqrt(double v) { return rsqrt(v); }
template <typename U> __device__ U rsqrt(U v) { return U(1) / sqrt(v); }
template <> __device__ float rsqrt(float v) { return rsqrtf(v); }
template <> __device__ double rsqrt(double v) { return rsqrt(v); }
namespace {
// This is the un-specialized struct. Note that we prevent instantiation of
// this struct by putting an undefined symbol in the function body so it won't
// compile.
......@@ -224,7 +224,6 @@ namespace {
// };
// https://github.com/NVIDIA/apex/issues/246
template <typename T> struct SharedMemory;
template <> struct SharedMemory<float> {
__device__ float *getPointer() {
extern __shared__ float s_float[];
......@@ -238,7 +237,6 @@ template <> struct SharedMemory<double> {
return s_double;
}
};
} // namespace
template <typename T, typename U>
__global__ void
......@@ -457,6 +455,7 @@ cuComputeGradGammaBeta(const U *part_grad_gamma, const U *part_grad_beta,
}
}
template <typename T, typename U>
__global__ void
cuComputeGradInput(const T *__restrict__ dout, const T *__restrict__ dout_resid,
......@@ -634,3 +633,4 @@ void HostLayerNormGradient(const T *dout, const T *dout_resid, const U *mean,
dout, dout_resid, static_cast<T *>(input.data_ptr()), n1, n2, mean,
invvar, U(epsilon), gamma, grad_input);
}
} // namespace
apex/contrib/csrc/multihead_attn/masked_softmax_dropout.cpp deleted 100644 → 0
View file @ ed94d0bb
#include <torch/extension.h>
#include <vector>
namespace multihead_attn {
namespace fused_softmax {
namespace mask_softmax_dropout {
std::vector<torch::Tensor> fwd_cuda(bool is_training, int heads,
torch::Tensor const &input,
const uint8_t *pad_mask,
float dropout_prob);
torch::Tensor bwd_cuda(int heads, torch::Tensor const &output_grads,
torch::Tensor const &softmax_results,
torch::Tensor const &dropout_mask,
const uint8_t *padding_mask, float dropout_prob);
// C++ interface
#define CHECK_CUDA(x) \
AT_ASSERTM(x.type().is_cuda(), #x " must be a CUDA tensor")
#define CHECK_CONTIGUOUS(x) \
AT_ASSERTM(x.is_contiguous(), #x " must be contiguous")
#define CHECK_INPUT(x) \
CHECK_CUDA(x); \
CHECK_CONTIGUOUS(x)
std::vector<torch::Tensor> fwd(bool use_mask, bool is_training, int heads,
torch::Tensor const &input,
torch::Tensor const &pad_mask,
float dropout_prob) {
AT_ASSERTM(input.dim() == 3, "expected 3D tensor");
AT_ASSERTM(input.type().scalarType() == at::ScalarType::Half,
"Only HALF is supported");
if (use_mask) {
AT_ASSERTM(pad_mask.dim() == 2, "expected 2D tensor");
AT_ASSERTM(pad_mask.type().scalarType() == at::ScalarType::Byte,
"Only BYTE is supported");
}
return fwd_cuda(is_training, heads, input,
use_mask ? static_cast<const uint8_t *>(pad_mask.data_ptr())
: nullptr,
dropout_prob);
}
torch::Tensor bwd(bool use_mask, int heads, torch::Tensor const &output_grads,
torch::Tensor const &softmax_results,
torch::Tensor const &dropout_mask,
torch::Tensor const &padding_mask, float dropout_prob) {
AT_ASSERTM(output_grads.dim() == 3, "expected 3D tensor");
AT_ASSERTM(softmax_results.dim() == 3, "expected 3D tensor");
AT_ASSERTM(dropout_mask.dim() == 3, "expected 3D tensor");
AT_ASSERTM(output_grads.type().scalarType() == at::ScalarType::Half,
"Only HALF is supported");
AT_ASSERTM(softmax_results.type().scalarType() == at::ScalarType::Half,
"Only HALF is supported");
// AT_ASSERTM(dropout_mask.type().scalarType() == at::ScalarType::Byte,
// "Only BYTE is supported");
return bwd_cuda(heads, output_grads, softmax_results, dropout_mask,
use_mask
? static_cast<const uint8_t *>(padding_mask.data_ptr())
: nullptr,
dropout_prob);
}
} // end namespace mask_softmax_dropout
} // end namespace fused_softmax
} // end namespace multihead_attn
PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
m.def("forward", &multihead_attn::fused_softmax::mask_softmax_dropout::fwd,
"Self Multihead Attention masked softmax dropout -- Forward.");
m.def("backward", &multihead_attn::fused_softmax::mask_softmax_dropout::bwd,
"Self Multihead Attention masked softmax dropout -- Backward.");
}
apex/contrib/csrc/multihead_attn/masked_softmax_dropout_cuda.cu
View file @ 7ec8ed67
......@@ -11,8 +11,8 @@
#include <ATen/cuda/CUDAContext.h>
#include <torch/extension.h>
#include "dropout.h"
#include "softmax.h"
#include "dropout.cuh"
#include "softmax.cuh"
namespace multihead_attn {
namespace fused_softmax {
......
apex/contrib/csrc/multihead_attn/multihead_attn_frontend.cpp 0 → 100644
View file @ 7ec8ed67
This diff is collapsed.
apex/contrib/csrc/multihead_attn/philox.h → apex/contrib/csrc/multihead_attn/philox.cuh
View file @ 7ec8ed67
#pragma once
// Philox CUDA.
namespace {
class Philox {
public:
__device__ inline Philox(unsigned long long seed,
......@@ -85,8 +87,10 @@ private:
static const unsigned long kPhiloxSB = 0xCD9E8D57;
};
// Inverse of 2^32.
#define M_RAN_INVM32 2.3283064e-10f
constexpr float M_RAN_INVM32 = 2.3283064e-10f;
__device__ __inline__ float4 uniform4(uint4 x) {
return make_float4(x.x * M_RAN_INVM32, x.y * M_RAN_INVM32, x.z * M_RAN_INVM32,
x.w * M_RAN_INVM32);
}
} // namespace
apex/contrib/csrc/multihead_attn/self_multihead_attn.cpp deleted 100644 → 0
View file @ ed94d0bb
#include <torch/extension.h>
#include <vector>
namespace multihead_attn {
namespace self {
namespace cublas_gemmex {
std::vector<torch::Tensor> fwd_cuda(bool use_time_mask, bool is_training,
int heads, torch::Tensor const &inputs,
torch::Tensor const &input_weights,
torch::Tensor const &output_weights,
const uint8_t *pad_mask,
float dropout_prob);
std::vector<torch::Tensor> bwd_cuda(
int heads, torch::Tensor const &output_grads,
torch::Tensor const &matmul2_results, torch::Tensor const &dropout_results,
torch::Tensor const &softmax_results,
torch::Tensor const &input_lin_results, torch::Tensor const &inputs,
torch::Tensor const &input_weights, torch::Tensor const &output_weights,
torch::Tensor const &dropout_mask, float dropout_prob);
// C++ interface
#define CHECK_CUDA(x) \
AT_ASSERTM(x.type().is_cuda(), #x " must be a CUDA tensor")
#define CHECK_CONTIGUOUS(x) \
AT_ASSERTM(x.is_contiguous(), #x " must be contiguous")
#define CHECK_INPUT(x) \
CHECK_CUDA(x); \
CHECK_CONTIGUOUS(x)
std::vector<torch::Tensor>
fwd(bool use_mask, bool use_time_mask, bool is_training, int heads,
torch::Tensor const &inputs, torch::Tensor const &input_weights,
torch::Tensor const &output_weights, torch::Tensor const &pad_mask,
float dropout_prob) {
AT_ASSERTM(inputs.dim() == 3, "expected 3D tensor");
AT_ASSERTM(input_weights.dim() == 2, "expected 2D tensor");
AT_ASSERTM(output_weights.dim() == 2, "expected 2D tensor");
AT_ASSERTM(inputs.type().scalarType() == at::ScalarType::Half,
"Only HALF is supported");
AT_ASSERTM(input_weights.type().scalarType() == at::ScalarType::Half,
"Only HALF is supported");
AT_ASSERTM(output_weights.type().scalarType() == at::ScalarType::Half,
"Only HALF is supported");
if (use_mask) {
AT_ASSERTM(pad_mask.dim() == 2, "expected 2D tensor");
AT_ASSERTM(pad_mask.type().scalarType() == at::ScalarType::Byte,
"Only BYTE is supported");
}
return fwd_cuda(
use_time_mask, is_training, heads, inputs, input_weights, output_weights,
use_mask ? static_cast<const uint8_t *>(pad_mask.data_ptr()) : nullptr,
dropout_prob);
}
std::vector<torch::Tensor>
bwd(int heads, torch::Tensor const &output_grads,
torch::Tensor const &matmul2_results, torch::Tensor const &dropout_results,
torch::Tensor const &softmax_results,
torch::Tensor const &input_lin_results, torch::Tensor const &inputs,
torch::Tensor const &input_weights, torch::Tensor const &output_weights,
torch::Tensor const &dropout_mask, float dropout_prob) {
AT_ASSERTM(output_grads.dim() == 3, "expected 3D tensor");
AT_ASSERTM(matmul2_results.dim() == 3, "expected 3D tensor");
AT_ASSERTM(dropout_results.dim() == 3, "expected 3D tensor");
AT_ASSERTM(softmax_results.dim() == 3, "expected 3D tensor");
AT_ASSERTM(input_lin_results.dim() == 3, "expected 3D tensor");
AT_ASSERTM(inputs.dim() == 3, "expected 3D tensor");
AT_ASSERTM(input_weights.dim() == 2, "expected 2D tensor");
AT_ASSERTM(output_weights.dim() == 2, "expected 2D tensor");
AT_ASSERTM(dropout_mask.dim() == 3, "expected 3D tensor");
AT_ASSERTM(output_grads.type().scalarType() == at::ScalarType::Half,
"Only HALF is supported");
AT_ASSERTM(matmul2_results.type().scalarType() == at::ScalarType::Half,
"Only HALF is supported");
AT_ASSERTM(dropout_results.type().scalarType() == at::ScalarType::Half,
"Only HALF is supported");
AT_ASSERTM(softmax_results.type().scalarType() == at::ScalarType::Half,
"Only HALF is supported");
AT_ASSERTM(input_lin_results.type().scalarType() == at::ScalarType::Half,
"Only HALF is supported");
AT_ASSERTM(inputs.type().scalarType() == at::ScalarType::Half,
"Only HALF is supported");
AT_ASSERTM(input_weights.type().scalarType() == at::ScalarType::Half,
"Only HALF is supported");
AT_ASSERTM(output_weights.type().scalarType() == at::ScalarType::Half,
"Only HALF is supported");
AT_ASSERTM(dropout_mask.type().scalarType() == at::ScalarType::Byte,
"Only BYTE is supported");
return bwd_cuda(heads, output_grads, matmul2_results, dropout_results,
softmax_results, input_lin_results, inputs, input_weights,
output_weights, dropout_mask, dropout_prob);
}
} // end namespace cublas_gemmex
} // end namespace self
} // end namespace multihead_attn
PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
m.def("forward", &multihead_attn::self::cublas_gemmex::fwd,
"Self Multihead Attention Forward.");
m.def("backward", &multihead_attn::self::cublas_gemmex::bwd,
"Self Multihead Attention Backward.");
}
apex/contrib/csrc/multihead_attn/self_multihead_attn_bias.cpp deleted 100644 → 0
View file @ ed94d0bb
#include <torch/extension.h>
#include <vector>
namespace multihead_attn {
namespace self_bias {
namespace cublas_gemmex {
std::vector<torch::Tensor>
fwd_cuda(bool use_time_mask, bool is_training, int heads,
torch::Tensor const &inputs, torch::Tensor const &input_weights,
torch::Tensor const &output_weights, torch::Tensor const &input_biases,
torch::Tensor const &output_biases, const uint8_t *pad_mask,
float dropout_prob);
std::vector<torch::Tensor> bwd_cuda(
int heads, torch::Tensor const &output_grads,
torch::Tensor const &matmul2_results, torch::Tensor const &dropout_results,
torch::Tensor const &softmax_results,
torch::Tensor const &input_lin_results, torch::Tensor const &inputs,
torch::Tensor const &input_weights, torch::Tensor const &output_weights,
// torch::Tensor const& input_biases,
// torch::Tensor const& output_biases,
torch::Tensor const &dropout_mask, float dropout_prob);
// C++ interface
#define CHECK_CUDA(x) \
AT_ASSERTM(x.type().is_cuda(), #x " must be a CUDA tensor")
#define CHECK_CONTIGUOUS(x) \
AT_ASSERTM(x.is_contiguous(), #x " must be contiguous")
#define CHECK_INPUT(x) \
CHECK_CUDA(x); \
CHECK_CONTIGUOUS(x)
std::vector<torch::Tensor>
fwd(bool use_mask, bool use_time_mask, bool is_training, int heads,
torch::Tensor const &inputs, torch::Tensor const &input_weights,
torch::Tensor const &output_weights, torch::Tensor const &input_biases,
torch::Tensor const &output_biases, torch::Tensor const &pad_mask,
float dropout_prob) {
AT_ASSERTM(inputs.dim() == 3, "expected 3D tensor");
AT_ASSERTM(input_weights.dim() == 2, "expected 2D tensor");
AT_ASSERTM(output_weights.dim() == 2, "expected 2D tensor");
AT_ASSERTM(inputs.type().scalarType() == at::ScalarType::Half,
"Only HALF is supported");
AT_ASSERTM(input_weights.type().scalarType() == at::ScalarType::Half,
"Only HALF is supported");
AT_ASSERTM(output_weights.type().scalarType() == at::ScalarType::Half,
"Only HALF is supported");
if (use_mask) {
AT_ASSERTM(pad_mask.dim() == 2, "expected 2D tensor");
AT_ASSERTM(pad_mask.type().scalarType() == at::ScalarType::Byte,
"Only BYTE is supported");
}
return fwd_cuda(use_time_mask, is_training, heads, inputs, input_weights,
output_weights, input_biases, output_biases,
use_mask ? static_cast<const uint8_t *>(pad_mask.data_ptr())
: nullptr,
dropout_prob);
}
std::vector<torch::Tensor>
bwd(int heads, torch::Tensor const &output_grads,
torch::Tensor const &matmul2_results, torch::Tensor const &dropout_results,
torch::Tensor const &softmax_results,
torch::Tensor const &input_lin_results, torch::Tensor const &inputs,
torch::Tensor const &input_weights, torch::Tensor const &output_weights,
torch::Tensor const &dropout_mask, float dropout_prob) {
AT_ASSERTM(output_grads.dim() == 3, "expected 3D tensor");
AT_ASSERTM(matmul2_results.dim() == 3, "expected 3D tensor");
AT_ASSERTM(dropout_results.dim() == 3, "expected 3D tensor");
AT_ASSERTM(softmax_results.dim() == 3, "expected 3D tensor");
AT_ASSERTM(input_lin_results.dim() == 3, "expected 3D tensor");
AT_ASSERTM(inputs.dim() == 3, "expected 3D tensor");
AT_ASSERTM(input_weights.dim() == 2, "expected 2D tensor");
AT_ASSERTM(output_weights.dim() == 2, "expected 2D tensor");
AT_ASSERTM(dropout_mask.dim() == 3, "expected 3D tensor");
AT_ASSERTM(output_grads.type().scalarType() == at::ScalarType::Half,
"Only HALF is supported");
AT_ASSERTM(matmul2_results.type().scalarType() == at::ScalarType::Half,
"Only HALF is supported");
AT_ASSERTM(dropout_results.type().scalarType() == at::ScalarType::Half,
"Only HALF is supported");
AT_ASSERTM(softmax_results.type().scalarType() == at::ScalarType::Half,
"Only HALF is supported");
AT_ASSERTM(input_lin_results.type().scalarType() == at::ScalarType::Half,
"Only HALF is supported");
AT_ASSERTM(inputs.type().scalarType() == at::ScalarType::Half,
"Only HALF is supported");
AT_ASSERTM(input_weights.type().scalarType() == at::ScalarType::Half,
"Only HALF is supported");
AT_ASSERTM(output_weights.type().scalarType() == at::ScalarType::Half,
"Only HALF is supported");
AT_ASSERTM(dropout_mask.type().scalarType() == at::ScalarType::Byte,
"Only BYTE is supported");
return bwd_cuda(heads, output_grads, matmul2_results, dropout_results,
softmax_results, input_lin_results, inputs, input_weights,
output_weights, dropout_mask, dropout_prob);
}
} // end namespace cublas_gemmex
} // namespace self_bias
} // end namespace multihead_attn
PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
m.def("forward", &multihead_attn::self_bias::cublas_gemmex::fwd,
"Self Multihead Attention with Bias -- Forward.");
m.def("backward", &multihead_attn::self_bias::cublas_gemmex::bwd,
"Self Multihead Attention with Bias -- Backward.");
}
apex/contrib/csrc/multihead_attn/self_multihead_attn_bias_additive_mask.cpp deleted 100644 → 0
View file @ ed94d0bb
#include <cuda_fp16.h>
#include <torch/extension.h>
#include <vector>
namespace multihead_attn {
namespace self_bias_additive_mask {
namespace cublas_gemmex {
std::vector<torch::Tensor> fwd_cuda(bool use_time_mask, bool is_training,
int heads, torch::Tensor const &inputs,
torch::Tensor const &input_weights,
torch::Tensor const &output_weights,
torch::Tensor const &input_biases,
torch::Tensor const &output_biases,
const half *pad_mask, float dropout_prob);
std::vector<torch::Tensor> bwd_cuda(
int heads, torch::Tensor const &output_grads,
torch::Tensor const &matmul2_results, torch::Tensor const &dropout_results,
// torch::Tensor const& softmax_results,
torch::Tensor const &bmm1_results, torch::Tensor const &pad_mask,
torch::Tensor const &input_lin_results, torch::Tensor const &inputs,
torch::Tensor const &input_weights, torch::Tensor const &output_weights,
// torch::Tensor const& input_biases,
// torch::Tensor const& output_biases,
torch::Tensor const &dropout_mask, float dropout_prob);
// C++ interface
#define CHECK_CUDA(x) \
AT_ASSERTM(x.type().is_cuda(), #x " must be a CUDA tensor")
#define CHECK_CONTIGUOUS(x) \
AT_ASSERTM(x.is_contiguous(), #x " must be contiguous")
#define CHECK_INPUT(x) \
CHECK_CUDA(x); \
CHECK_CONTIGUOUS(x)
std::vector<torch::Tensor>
fwd(bool use_mask, bool use_time_mask, bool is_training, int heads,
torch::Tensor const &inputs, torch::Tensor const &input_weights,
torch::Tensor const &output_weights, torch::Tensor const &input_biases,
torch::Tensor const &output_biases, torch::Tensor const &pad_mask,
float dropout_prob) {
AT_ASSERTM(inputs.dim() == 3, "expected 3D tensor");
AT_ASSERTM(input_weights.dim() == 2, "expected 2D tensor");
AT_ASSERTM(output_weights.dim() == 2, "expected 2D tensor");
AT_ASSERTM(inputs.type().scalarType() == at::ScalarType::Half,
"Only HALF is supported");
AT_ASSERTM(input_weights.type().scalarType() == at::ScalarType::Half,
"Only HALF is supported");
AT_ASSERTM(output_weights.type().scalarType() == at::ScalarType::Half,
"Only HALF is supported");
AT_ASSERTM(use_mask, "no mask is not supported");
if (use_mask) {
AT_ASSERTM(pad_mask.dim() == 2, "expected 2D tensor");
AT_ASSERTM(pad_mask.type().scalarType() == at::ScalarType::Half,
"Only Half is supported");
}
return fwd_cuda(use_time_mask, is_training, heads, inputs, input_weights,
output_weights, input_biases, output_biases,
use_mask ? static_cast<const half *>(pad_mask.data_ptr())
: nullptr,
dropout_prob);
}
std::vector<torch::Tensor>
bwd(int heads, torch::Tensor const &output_grads,
torch::Tensor const &matmul2_results, torch::Tensor const &dropout_results,
torch::Tensor const &bmm1_results, torch::Tensor const &pad_mask,
torch::Tensor const &input_lin_results, torch::Tensor const &inputs,
torch::Tensor const &input_weights, torch::Tensor const &output_weights,
torch::Tensor const &dropout_mask, float dropout_prob) {
AT_ASSERTM(output_grads.dim() == 3, "expected 3D tensor");
AT_ASSERTM(matmul2_results.dim() == 3, "expected 3D tensor");
AT_ASSERTM(dropout_results.dim() == 3, "expected 3D tensor");
AT_ASSERTM(input_lin_results.dim() == 3, "expected 3D tensor");
AT_ASSERTM(inputs.dim() == 3, "expected 3D tensor");
AT_ASSERTM(input_weights.dim() == 2, "expected 2D tensor");
AT_ASSERTM(output_weights.dim() == 2, "expected 2D tensor");
AT_ASSERTM(dropout_mask.dim() == 3, "expected 3D tensor");
AT_ASSERTM(output_grads.type().scalarType() == at::ScalarType::Half,
"Only HALF is supported");
AT_ASSERTM(matmul2_results.type().scalarType() == at::ScalarType::Half,
"Only HALF is supported");
AT_ASSERTM(dropout_results.type().scalarType() == at::ScalarType::Half,
"Only HALF is supported");
AT_ASSERTM(input_lin_results.type().scalarType() == at::ScalarType::Half,
"Only HALF is supported");
AT_ASSERTM(inputs.type().scalarType() == at::ScalarType::Half,
"Only HALF is supported");
AT_ASSERTM(input_weights.type().scalarType() == at::ScalarType::Half,
"Only HALF is supported");
AT_ASSERTM(output_weights.type().scalarType() == at::ScalarType::Half,
"Only HALF is supported");
AT_ASSERTM(dropout_mask.type().scalarType() == at::ScalarType::Byte,
"Only BYTE is supported");
return bwd_cuda(heads, output_grads, matmul2_results, dropout_results,
bmm1_results, pad_mask, input_lin_results, inputs,
input_weights, output_weights, dropout_mask, dropout_prob);
}
} // end namespace cublas_gemmex
} // namespace self_bias_additive_mask
} // end namespace multihead_attn
PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
m.def("forward", &multihead_attn::self_bias_additive_mask::cublas_gemmex::fwd,
"Self Multihead Attention with Bias -- Forward.");
m.def("backward",
&multihead_attn::self_bias_additive_mask::cublas_gemmex::bwd,
"Self Multihead Attention with Bias -- Backward.");
}
apex/contrib/csrc/multihead_attn/self_multihead_attn_bias_additive_mask_cuda.cu
View file @ 7ec8ed67
......@@ -11,10 +11,9 @@
#include <ATen/cuda/CUDAContext.h>
#include <torch/extension.h>
#include "dropout.h"
#include "layer_norm.h"
#include "softmax.h"
#include "strided_batched_gemm.h"
#include "dropout.cuh"
#include "softmax.cuh"
#include "strided_batched_gemm.cuh"
namespace multihead_attn {
namespace self_bias_additive_mask {
......
apex/contrib/csrc/multihead_attn/self_multihead_attn_bias_cuda.cu
View file @ 7ec8ed67
......@@ -11,10 +11,9 @@
#include <ATen/cuda/CUDAContext.h>
#include <torch/extension.h>
#include "dropout.h"
#include "layer_norm.h"
#include "softmax.h"
#include "strided_batched_gemm.h"
#include "dropout.cuh"
#include "softmax.cuh"
#include "strided_batched_gemm.cuh"
namespace multihead_attn {
namespace self_bias {
......
apex/contrib/csrc/multihead_attn/self_multihead_attn_cuda.cu
View file @ 7ec8ed67
......@@ -11,10 +11,9 @@
#include <ATen/cuda/CUDAContext.h>
#include <torch/extension.h>
#include "dropout.h"
#include "layer_norm.h"
#include "softmax.h"
#include "strided_batched_gemm.h"
#include "dropout.cuh"
#include "softmax.cuh"
#include "strided_batched_gemm.cuh"
namespace multihead_attn {
namespace self {
......
apex/contrib/csrc/multihead_attn/self_multihead_attn_norm_add.cpp deleted 100644 → 0
View file @ ed94d0bb
#include <torch/extension.h>
#include <vector>
namespace multihead_attn {
namespace self_norm_add {
namespace cublas_gemmex {
std::vector<torch::Tensor> fwd_cuda(bool use_time_mask, bool is_training,
int heads, torch::Tensor const &inputs,
torch::Tensor const &lyr_nrm_gamma_weights,
torch::Tensor const &lyr_nrm_beta_weights,
torch::Tensor const &input_weights,
torch::Tensor const &output_weights,
const uint8_t *pad_mask,
float dropout_prob);
std::vector<torch::Tensor> bwd_cuda(
int heads, torch::Tensor const &output_grads,
torch::Tensor const &matmul2_results, torch::Tensor const &dropout_results,
torch::Tensor const &softmax_results,
torch::Tensor const &input_lin_results,
torch::Tensor const &lyr_nrm_results, torch::Tensor const &lyr_nrm_mean,
torch::Tensor const &lyr_nrm_invvar, torch::Tensor const &inputs,
torch::Tensor const &lyr_nrm_gamma_weights,
torch::Tensor const &lyr_nrm_beta_weights,
torch::Tensor const &input_weights, torch::Tensor const &output_weights,
torch::Tensor const &dropout_mask, torch::Tensor const &dropout_add_mask,
float dropout_prob);
// C++ interface
#define CHECK_CUDA(x) \
AT_ASSERTM(x.type().is_cuda(), #x " must be a CUDA tensor")
#define CHECK_CONTIGUOUS(x) \
AT_ASSERTM(x.is_contiguous(), #x " must be contiguous")
#define CHECK_INPUT(x) \
CHECK_CUDA(x); \
CHECK_CONTIGUOUS(x)
std::vector<torch::Tensor>
fwd(bool use_mask, bool use_time_mask, bool is_training, int heads,
torch::Tensor const &inputs, torch::Tensor const &lyr_nrm_gamma_weights,
torch::Tensor const &lyr_nrm_beta_weights,
torch::Tensor const &input_weights, torch::Tensor const &output_weights,
torch::Tensor const &pad_mask, float dropout_prob) {
AT_ASSERTM(inputs.dim() == 3, "expected 3D tensor");
AT_ASSERTM(lyr_nrm_gamma_weights.dim() == 1, "expected 1D tensor");
AT_ASSERTM(lyr_nrm_beta_weights.dim() == 1, "expected 1D tensor");
AT_ASSERTM(input_weights.dim() == 2, "expected 2D tensor");
AT_ASSERTM(output_weights.dim() == 2, "expected 2D tensor");
AT_ASSERTM(inputs.type().scalarType() == at::ScalarType::Half,
"Only HALF is supported");
AT_ASSERTM(lyr_nrm_gamma_weights.type().scalarType() == at::ScalarType::Half,
"Only HALF is supported");
AT_ASSERTM(lyr_nrm_beta_weights.type().scalarType() == at::ScalarType::Half,
"Only HALF is supported");
AT_ASSERTM(input_weights.type().scalarType() == at::ScalarType::Half,
"Only HALF is supported");
AT_ASSERTM(output_weights.type().scalarType() == at::ScalarType::Half,
"Only HALF is supported");
if (use_mask) {
AT_ASSERTM(pad_mask.dim() == 2, "expected 2D tensor");
AT_ASSERTM(pad_mask.type().scalarType() == at::ScalarType::Byte,
"Only BYTE is supported");
}
return fwd_cuda(
use_time_mask, is_training, heads, inputs, lyr_nrm_gamma_weights,
lyr_nrm_beta_weights, input_weights, output_weights,
use_mask ? static_cast<const uint8_t *>(pad_mask.data_ptr()) : nullptr,
dropout_prob);
}
std::vector<torch::Tensor>
bwd(int heads, torch::Tensor const &output_grads,
torch::Tensor const &matmul2_results, torch::Tensor const &dropout_results,
torch::Tensor const &softmax_results,
torch::Tensor const &input_lin_results,
torch::Tensor const &lyr_nrm_results, torch::Tensor const &lyr_nrm_mean,
torch::Tensor const &lyr_nrm_invvar, torch::Tensor const &inputs,
torch::Tensor const &lyr_nrm_gamma_weights,
torch::Tensor const &lyr_nrm_beta_weights,
torch::Tensor const &input_weights, torch::Tensor const &output_weights,
torch::Tensor const &dropout_mask, torch::Tensor const &dropout_add_mask,
float dropout_prob) {
AT_ASSERTM(output_grads.dim() == 3, "expected 3D tensor");
AT_ASSERTM(matmul2_results.dim() == 3, "expected 3D tensor");
AT_ASSERTM(dropout_results.dim() == 3, "expected 3D tensor");
AT_ASSERTM(softmax_results.dim() == 3, "expected 3D tensor");
AT_ASSERTM(input_lin_results.dim() == 3, "expected 3D tensor");
AT_ASSERTM(lyr_nrm_results.dim() == 3, "expected 3D tensor");
AT_ASSERTM(lyr_nrm_mean.dim() == 1, "expected 1D tensor");
AT_ASSERTM(lyr_nrm_invvar.dim() == 1, "expected 1D tensor");
AT_ASSERTM(inputs.dim() == 3, "expected 3D tensor");
AT_ASSERTM(lyr_nrm_gamma_weights.dim() == 1, "expected 1D tensor");
AT_ASSERTM(lyr_nrm_beta_weights.dim() == 1, "expected 1D tensor");
AT_ASSERTM(input_weights.dim() == 2, "expected 2D tensor");
AT_ASSERTM(output_weights.dim() == 2, "expected 2D tensor");
AT_ASSERTM(dropout_mask.dim() == 3, "expected 3D tensor");
AT_ASSERTM(dropout_add_mask.dim() == 3, "expected 3D tensor");
AT_ASSERTM(output_grads.type().scalarType() == at::ScalarType::Half,
"Only HALF is supported");
AT_ASSERTM(matmul2_results.type().scalarType() == at::ScalarType::Half,
"Only HALF is supported");
AT_ASSERTM(dropout_results.type().scalarType() == at::ScalarType::Half,
"Only HALF is supported");
AT_ASSERTM(softmax_results.type().scalarType() == at::ScalarType::Half,
"Only HALF is supported");
AT_ASSERTM(input_lin_results.type().scalarType() == at::ScalarType::Half,
"Only HALF is supported");
AT_ASSERTM(lyr_nrm_results.type().scalarType() == at::ScalarType::Half,
"Only HALF is supported");
AT_ASSERTM(lyr_nrm_mean.type().scalarType() == at::ScalarType::Float,
"Only FLOAT is supported");
AT_ASSERTM(lyr_nrm_invvar.type().scalarType() == at::ScalarType::Float,
"Only FLOAT is supported");
AT_ASSERTM(inputs.type().scalarType() == at::ScalarType::Half,
"Only HALF is supported");
AT_ASSERTM(lyr_nrm_gamma_weights.type().scalarType() == at::ScalarType::Half,
"Only HALF is supported");
AT_ASSERTM(lyr_nrm_beta_weights.type().scalarType() == at::ScalarType::Half,
"Only HALF is supported");
AT_ASSERTM(input_weights.type().scalarType() == at::ScalarType::Half,
"Only HALF is supported");
AT_ASSERTM(output_weights.type().scalarType() == at::ScalarType::Half,
"Only HALF is supported");
AT_ASSERTM(dropout_mask.type().scalarType() == at::ScalarType::Byte,
"Only BYTE is supported");
AT_ASSERTM(dropout_add_mask.type().scalarType() == at::ScalarType::Byte,
"Only BYTE is supported");
return bwd_cuda(heads, output_grads, matmul2_results, dropout_results,
softmax_results, input_lin_results, lyr_nrm_results,
lyr_nrm_mean, lyr_nrm_invvar, inputs, lyr_nrm_gamma_weights,
lyr_nrm_beta_weights, input_weights, output_weights,
dropout_mask, dropout_add_mask, dropout_prob);
}
} // end namespace cublas_gemmex
} // end namespace self_norm_add
} // end namespace multihead_attn
PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
m.def("forward", &multihead_attn::self_norm_add::cublas_gemmex::fwd,
"Self Multihead Attention Plus Layer Norm and Residual Add Forward.");
m.def("backward", &multihead_attn::self_norm_add::cublas_gemmex::bwd,
"Self Multihead Attention Plus Layer Norm and Residual Add Backward.");
}
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