xiezhongtao
2bd9bd4cc2
将分散在各文件中的CUDA/HIP/MUSA硬件相关头文件引用统一到vendors目录下的对应头文件中，提高代码可维护性。移除重复的头文件引用，优化构建配置。
31 lines
838 B
Plaintext
31 lines
838 B
Plaintext
#pragma once
|
|
/**
|
|
* __device__ datatypes vectorized by 4
|
|
*/
|
|
|
|
// Include both AMD and NVIDIA fp8 types to avoid circular import
|
|
#include "../vendors/functions.h"
|
|
|
|
namespace vllm {
|
|
|
|
// Vectorization containers
|
|
template <typename scalar_t, size_t vec_size>
|
|
struct __align__(vec_size * sizeof(scalar_t)) vec_n_t {
|
|
scalar_t val[vec_size];
|
|
};
|
|
|
|
template <typename quant_type_t, size_t vec_size>
|
|
struct __align__(vec_size * sizeof(quant_type_t)) q8_n_t {
|
|
static_assert(std::is_same_v<quant_type_t, int8_t> ||
|
|
std::is_same_v<quant_type_t, c10::Float8_e4m3fn> ||
|
|
std::is_same_v<quant_type_t, c10::Float8_e4m3fnuz>);
|
|
quant_type_t val[vec_size];
|
|
};
|
|
|
|
template <typename scalar_t>
|
|
using vec4_t = vec_n_t<scalar_t, 4>;
|
|
template <typename quant_type_t>
|
|
using q8x4_t = q8_n_t<quant_type_t, 4>;
|
|
|
|
} // namespace vllm
|