| 1 | /* |
| 2 | * Adapted from |
| 3 | * https://github.com/NVIDIA/FasterTransformer/blob/release/v5.3_tag/src/fastertransformer/kernels/decoder_masked_multihead_attention/decoder_masked_multihead_attention_template.hpp |
| 4 | * and |
| 5 | * https://github.com/NVIDIA/FasterTransformer/blob/release/v5.3_tag/src/fastertransformer/kernels/decoder_masked_multihead_attention_utils.h |
| 6 | * Copyright (c) 2023, The vLLM team. |
| 7 | * Copyright (c) 2020-2023, NVIDIA CORPORATION. All rights reserved. |
| 8 | * |
| 9 | * Licensed under the Apache License, Version 2.0 (the "License"); |
| 10 | * you may not use this file except in compliance with the License. |
| 11 | * You may obtain a copy of the License at |
| 12 | * |
| 13 | * http://www.apache.org/licenses/LICENSE-2.0 |
| 14 | * |
| 15 | * Unless required by applicable law or agreed to in writing, software |
| 16 | * distributed under the License is distributed on an "AS IS" BASIS, |
| 17 | * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. |
| 18 | * See the License for the specific language governing permissions and |
| 19 | * limitations under the License. |
| 20 | */ |
| 21 | #pragma once |
| 22 | |
| 23 | #include "attention_generic.cuh" |
| 24 | |
| 25 | #include <stdint.h> |
| 26 | |
| 27 | namespace vllm { |
| 28 | |
| 29 | // Define custom FP32 vector data types. |
| 30 | struct Float4_ { |
| 31 | float2 x; |
| 32 | float2 y; |
| 33 | }; |
| 34 | |
| 35 | struct Float8_ { |
| 36 | float2 x; |
| 37 | float2 y; |
| 38 | float2 z; |
| 39 | float2 w; |
| 40 | }; |
| 41 | |
| 42 | // FP32 vector types for Q, K, V. |
| 43 | template <> |
| 44 | struct Vec<float, 1> { |
| 45 | using Type = float; |
| 46 | }; |
| 47 | template <> |
| 48 | struct Vec<float, 2> { |
| 49 | using Type = float2; |
| 50 | }; |
| 51 | template <> |
| 52 | struct Vec<float, 4> { |
| 53 | using Type = float4; |
| 54 | }; |
| 55 | |
| 56 | // FP32 accumulator vector types corresponding to Vec. |
| 57 | template <> |
| 58 | struct FloatVec<float> { |
| 59 | using Type = float; |
| 60 | }; |
| 61 | template <> |
| 62 | struct FloatVec<float2> { |
| 63 | using Type = float2; |
| 64 | }; |
| 65 | template <> |
| 66 | struct FloatVec<float4> { |
| 67 | using Type = float4; |
| 68 | }; |
| 69 | |
| 70 | // Vector addition. |
| 71 | inline __device__ float add(float a, float b) { return a + b; } |
| 72 | |
| 73 | inline __device__ float2 add(float2 a, float2 b) { |
| 74 | float2 c; |
| 75 | c.x = add(a.x, b.x); |
| 76 | c.y = add(a.y, b.y); |
| 77 | return c; |
| 78 | } |
| 79 | |
| 80 | inline __device__ float4 add(float4 a, float4 b) { |
| 81 | float4 c; |
| 82 | c.x = add(a.x, b.x); |
| 83 | c.y = add(a.y, b.y); |
| 84 | c.z = add(a.z, b.z); |
| 85 | c.w = add(a.w, b.w); |
| 86 | return c; |
| 87 | } |
| 88 | |
| 89 | // Vector multiplication. |
| 90 | template <> |
| 91 | inline __device__ float mul<float, float>(float a, float b) { |
| 92 | return a * b; |
| 93 | } |
| 94 | |
| 95 | template <> |
| 96 | inline __device__ float2 mul(float2 a, float2 b) { |
| 97 | float2 c; |
| 98 | c.x = a.x * b.x; |
| 99 | c.y = a.y * b.y; |
| 100 | return c; |
| 101 | } |
| 102 | |
| 103 | template <> |
| 104 | inline __device__ float2 mul(float a, float2 b) { |
| 105 | float2 c; |
| 106 | c.x = a * b.x; |
| 107 | c.y = a * b.y; |
| 108 | return c; |
| 109 | } |
| 110 | |
| 111 | template <> |
| 112 | inline __device__ float4 mul(float4 a, float4 b) { |
| 113 | float4 c; |
| 114 | c.x = a.x * b.x; |
| 115 | c.y = a.y * b.y; |
| 116 | c.z = a.z * b.z; |
| 117 | c.w = a.w * b.w; |
| 118 | return c; |
| 119 | } |
| 120 | |
| 121 | template <> |
| 122 | inline __device__ float4 mul(float a, float4 b) { |
| 123 | float4 c; |
| 124 | c.x = a * b.x; |
| 125 | c.y = a * b.y; |
| 126 | c.z = a * b.z; |
| 127 | c.w = a * b.w; |
| 128 | return c; |
| 129 | } |
| 130 | |
| 131 | // Vector fused multiply-add. |
| 132 | inline __device__ float fma(float a, float b, float c) { return a * b + c; } |
| 133 | |
| 134 | inline __device__ float2 fma(float2 a, float2 b, float2 c) { |
| 135 | float2 d; |
| 136 | d.x = fma(a.x, b.x, c.x); |
| 137 | d.y = fma(a.y, b.y, c.y); |
| 138 | return d; |
| 139 | } |
| 140 | |
| 141 | inline __device__ float2 fma(float a, float2 b, float2 c) { |
| 142 | float2 d; |
| 143 | d.x = fma(a, b.x, c.x); |
| 144 | d.y = fma(a, b.y, c.y); |
| 145 | return d; |
| 146 | } |
| 147 | |
| 148 | inline __device__ float4 fma(float4 a, float4 b, float4 c) { |
| 149 | float4 d; |
| 150 | d.x = fma(a.x, b.x, c.x); |
| 151 | d.y = fma(a.y, b.y, c.y); |
| 152 | d.z = fma(a.z, b.z, c.z); |
| 153 | d.w = fma(a.w, b.w, c.w); |
| 154 | return d; |
| 155 | } |
| 156 | |
| 157 | inline __device__ float4 fma(float a, float4 b, float4 c) { |
| 158 | float4 d; |
| 159 | d.x = fma(a, b.x, c.x); |
| 160 | d.y = fma(a, b.y, c.y); |
| 161 | d.z = fma(a, b.z, c.z); |
| 162 | d.w = fma(a, b.w, c.w); |
| 163 | return d; |
| 164 | } |
| 165 | |
| 166 | inline __device__ Float4_ fma(float a, Float4_ b, Float4_ c) { |
| 167 | Float4_ d; |
| 168 | d.x = fma(a, b.x, c.x); |
| 169 | d.y = fma(a, b.y, c.y); |
| 170 | return d; |
| 171 | } |
| 172 | |
| 173 | inline __device__ Float8_ fma(float a, Float8_ b, Float8_ c) { |
| 174 | Float8_ d; |
| 175 | d.x = fma(a, b.x, c.x); |
| 176 | d.y = fma(a, b.y, c.y); |
| 177 | d.z = fma(a, b.z, c.z); |
| 178 | d.w = fma(a, b.w, c.w); |
| 179 | return d; |
| 180 | } |
| 181 | |
| 182 | // Vector sum. |
| 183 | template <> |
| 184 | inline __device__ float sum(float v) { |
| 185 | return v; |
| 186 | } |
| 187 | |
| 188 | template <> |
| 189 | inline __device__ float sum(float2 v) { |
| 190 | return v.x + v.y; |
| 191 | } |
| 192 | |
| 193 | template <> |
| 194 | inline __device__ float sum(float4 v) { |
| 195 | return v.x + v.y + v.z + v.w; |
| 196 | } |
| 197 | |
| 198 | template <> |
| 199 | inline __device__ float sum(Float4_ v) { |
| 200 | return v.x.x + v.x.y + v.y.x + v.y.y; |
| 201 | } |
| 202 | |
| 203 | template <> |
| 204 | inline __device__ float sum(Float8_ v) { |
| 205 | return v.x.x + v.x.y + v.y.x + v.y.y + v.z.x + v.z.y + v.w.x + v.w.y; |
| 206 | } |
| 207 | |
| 208 | // Vector dot product. |
| 209 | inline __device__ float dot(float a, float b) { return a * b; } |
| 210 | |
| 211 | inline __device__ float dot(float2 a, float2 b) { |
| 212 | float2 c = mul<float2, float2, float2>(a, b); |
| 213 | return c.x + c.y; |
| 214 | } |
| 215 | |
| 216 | inline __device__ float dot(Float4_ a, Float4_ b) { |
| 217 | float2 acc = mul<float2, float2, float2>(a.x, b.x); |
| 218 | acc = fma(a.y, b.y, acc); |
| 219 | return acc.x + acc.y; |
| 220 | } |
| 221 | |
| 222 | inline __device__ float dot(Float8_ a, Float8_ b) { |
| 223 | float2 acc = mul<float2, float2, float2>(a.x, b.x); |
| 224 | acc = fma(a.y, b.y, acc); |
| 225 | acc = fma(a.z, b.z, acc); |
| 226 | acc = fma(a.w, b.w, acc); |
| 227 | return acc.x + acc.y; |
| 228 | } |
| 229 | |
| 230 | // From float to float. |
| 231 | inline __device__ void from_float(float& dst, float src) { dst = src; } |
| 232 | |
| 233 | inline __device__ void from_float(float2& dst, float2 src) { dst = src; } |
| 234 | |
| 235 | inline __device__ void from_float(float4& dst, float4 src) { dst = src; } |
| 236 | |
| 237 | // From float to float. |
| 238 | inline __device__ float to_float(float u) { return u; } |
| 239 | |
| 240 | inline __device__ float2 to_float(float2 u) { return u; } |
| 241 | |
| 242 | inline __device__ float4 to_float(float4 u) { return u; } |
| 243 | |
| 244 | inline __device__ Float4_ to_float(Float4_ u) { return u; } |
| 245 | |
| 246 | inline __device__ Float8_ to_float(Float8_ u) { return u; } |
| 247 | |
| 248 | // Zero-out a variable. |
| 249 | inline __device__ void zero(float& dst) { dst = 0.f; } |
| 250 | |
| 251 | } // namespace vllm |