biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

[CI/Build] Enforce style for C++ and CUDA code with clang-format (#4722)

Browse Source
This commit is contained in:
Michael Goin
2024-05-22 03:18:41 -04:00
committed by GitHub
parent 9b9a10d6cb
commit 5f6d10c14c
64 changed files with 6398 additions and 6790 deletions
Download Patch File Download Diff File Expand all files Collapse all files

60
csrc/cpu/activation.cpp
View File

@@ -1,10 +1,10 @@
#include "cpu_types.hpp"
namespace {
template <typename scalar_t, vec_op::FP32Vec8 (*func)(const vec_op::FP32Vec8 &),
template <typename scalar_t, vec_op::FP32Vec8 (*func)(const vec_op::FP32Vec8&),
bool is_gated>
void activation_kernel(int num_tokens, int d, scalar_t *__restrict__ input,
scalar_t *__restrict__ output) {
void activation_kernel(int num_tokens, int d, scalar_t* __restrict__ input,
scalar_t* __restrict__ output) {
using scalar_vec_t = vec_op::vec_t<scalar_t>;
constexpr int VEC_ELEM_NUM = scalar_vec_t::get_elem_num();
@@ -34,13 +34,13 @@ void activation_kernel(int num_tokens, int d, scalar_t *__restrict__ input,
}
}
FORCE_INLINE vec_op::FP32Vec8 silu_act(const vec_op::FP32Vec8 &x) {
FORCE_INLINE vec_op::FP32Vec8 silu_act(const vec_op::FP32Vec8& x) {
const vec_op::FP32Vec8 zeros(0.0);
const vec_op::FP32Vec8 ones(1.0);
return x / (ones + (zeros - x).exp());
}
FORCE_INLINE vec_op::FP32Vec8 gelu_new_act(const vec_op::FP32Vec8 &x) {
FORCE_INLINE vec_op::FP32Vec8 gelu_new_act(const vec_op::FP32Vec8& x) {
const vec_op::FP32Vec8 ones(1.0);
const vec_op::FP32Vec8 w1(0.79788456f);
const vec_op::FP32Vec8 w2(0.044715f);
@@ -50,7 +50,7 @@ FORCE_INLINE vec_op::FP32Vec8 gelu_new_act(const vec_op::FP32Vec8 &x) {
return w3 * x * (ones + t);
}
FORCE_INLINE vec_op::FP32Vec8 gelu_fast_act(const vec_op::FP32Vec8 &x) {
FORCE_INLINE vec_op::FP32Vec8 gelu_fast_act(const vec_op::FP32Vec8& x) {
const vec_op::FP32Vec8 ones(1.0);
const vec_op::FP32Vec8 w1(0.79788456f);
const vec_op::FP32Vec8 w2(0.044715f);
@@ -59,14 +59,14 @@ FORCE_INLINE vec_op::FP32Vec8 gelu_fast_act(const vec_op::FP32Vec8 &x) {
return w3 * x * (ones + t);
}
FORCE_INLINE vec_op::FP32Vec8 gelu_act(const vec_op::FP32Vec8 &x) {
FORCE_INLINE vec_op::FP32Vec8 gelu_act(const vec_op::FP32Vec8& x) {
const vec_op::FP32Vec8 ones(1.0);
const vec_op::FP32Vec8 w1(M_SQRT1_2);
const vec_op::FP32Vec8 w2(0.5);
return x * w2 * (ones + (x * w1).er());
}
FORCE_INLINE vec_op::FP32Vec8 gelu_tanh_act(const vec_op::FP32Vec8 &x) {
FORCE_INLINE vec_op::FP32Vec8 gelu_tanh_act(const vec_op::FP32Vec8& x) {
const vec_op::FP32Vec8 ones(1.0);
const vec_op::FP32Vec8 w1(M_SQRT2 * M_2_SQRTPI * 0.5);
const vec_op::FP32Vec8 w2(0.5);
@@ -75,40 +75,36 @@ FORCE_INLINE vec_op::FP32Vec8 gelu_tanh_act(const vec_op::FP32Vec8 &x) {
const vec_op::FP32Vec8 inner = w1 * (x + x_3 * w3);
return x * w2 * (ones + inner.tanh());
}
}; // namespace
}; // namespace
void silu_and_mul(torch::Tensor &out, torch::Tensor &input) {
void silu_and_mul(torch::Tensor& out, torch::Tensor& input) {
int num_tokens = input.numel() / input.size(-1);
int d = input.size(-1) / 2;
VLLM_DISPATCH_FLOATING_TYPES(
input.scalar_type(), "silu_and_mul_impl", [&] {
CPU_KERNEL_GUARD_IN(silu_and_mul_impl)
activation_kernel<scalar_t, silu_act, true>(num_tokens, d,
input.data_ptr<scalar_t>(),
out.data_ptr<scalar_t>());
CPU_KERNEL_GUARD_OUT(silu_and_mul_impl)
});
VLLM_DISPATCH_FLOATING_TYPES(input.scalar_type(), "silu_and_mul_impl", [&] {
CPU_KERNEL_GUARD_IN(silu_and_mul_impl)
activation_kernel<scalar_t, silu_act, true>(
num_tokens, d, input.data_ptr<scalar_t>(), out.data_ptr<scalar_t>());
CPU_KERNEL_GUARD_OUT(silu_and_mul_impl)
});
}
void gelu_and_mul(torch::Tensor &out, // [..., d]
torch::Tensor &input) // [..., 2 * d]
void gelu_and_mul(torch::Tensor& out, // [..., d]
torch::Tensor& input) // [..., 2 * d]
{
int num_tokens = input.numel() / input.size(-1);
int d = input.size(-1) / 2;
VLLM_DISPATCH_FLOATING_TYPES(
input.scalar_type(), "gelu_and_mul_impl", [&] {
CPU_KERNEL_GUARD_IN(gelu_and_mul_impl)
activation_kernel<scalar_t, gelu_act, true>(num_tokens, d,
input.data_ptr<scalar_t>(),
out.data_ptr<scalar_t>());
CPU_KERNEL_GUARD_OUT(gelu_and_mul_impl)
});
VLLM_DISPATCH_FLOATING_TYPES(input.scalar_type(), "gelu_and_mul_impl", [&] {
CPU_KERNEL_GUARD_IN(gelu_and_mul_impl)
activation_kernel<scalar_t, gelu_act, true>(
num_tokens, d, input.data_ptr<scalar_t>(), out.data_ptr<scalar_t>());
CPU_KERNEL_GUARD_OUT(gelu_and_mul_impl)
});
}
void gelu_tanh_and_mul(torch::Tensor &out, // [..., d]
torch::Tensor &input) // [..., 2 * d]
void gelu_tanh_and_mul(torch::Tensor& out, // [..., d]
torch::Tensor& input) // [..., 2 * d]
{
int num_tokens = input.numel() / input.size(-1);
int d = input.size(-1) / 2;
@@ -123,7 +119,7 @@ void gelu_tanh_and_mul(torch::Tensor &out, // [..., d]
});
}
void gelu_new(torch::Tensor &out, torch::Tensor &input) {
void gelu_new(torch::Tensor& out, torch::Tensor& input) {
int num_tokens = input.numel() / input.size(-1);
int d = input.size(-1);
@@ -135,7 +131,7 @@ void gelu_new(torch::Tensor &out, torch::Tensor &input) {
});
}
void gelu_fast(torch::Tensor &out, torch::Tensor &input) {
void gelu_fast(torch::Tensor& out, torch::Tensor& input) {
int num_tokens = input.numel() / input.size(-1);
int d = input.size(-1);