| 1 | #pragma once |
| 2 | #include <hip/hip_fp8.h> |
| 3 | |
| 4 | #include <hip/hip_fp16.h> |
| 5 | #include <hip/hip_bf16.h> |
| 6 | #include <hip/hip_bfloat16.h> |
| 7 | |
| 8 | #include "../../../../attention/attention_dtypes.h" |
| 9 | #include <torch/headeronly/core/ScalarType.h> |
| 10 | |
| 11 | namespace vllm { |
| 12 | #ifdef USE_ROCM |
| 13 | |
| 14 | namespace fp8 { |
| 15 | #ifdef ENABLE_FP8 |
| 16 | |
| 17 | // Use hardware cvt instruction for fp8 on rocm |
| 18 | template <typename fp8_type> |
| 19 | __device__ __forceinline__ fp8_type cvt_c10(float const r) { |
| 20 | return {}; |
| 21 | } |
| 22 | |
| 23 | // __hip_fp8_e4m3 only exists starting in ROCm 6.3. The macro |
| 24 | // HIP_FP8_TYPE_OCP comes from the hip_fp8.h header and also makes |
| 25 | // its first appearance in ROCm 6.3. Since VLLM_DISPATCH_FP8_TYPES |
| 26 | // on ROCm instantiates both OCP and FNUZ kernels, we need to replace |
| 27 | // the new HW cvt with something reasonable that doesn't rely on the |
| 28 | // ROCm 6.3 feature. This allows compiling on ROCm 6.2 or newer. |
| 29 | template <> |
| 30 | __device__ __forceinline__ c10::Float8_e4m3fn cvt_c10(float const r) { |
| 31 | #if HIP_FP8_TYPE_OCP |
| 32 | return c10::Float8_e4m3fn( |
| 33 | __hip_cvt_float_to_fp8(r, __hip_fp8_e4m3::__default_saturation, |
| 34 | __hip_fp8_e4m3::__default_interpret), |
| 35 | c10::Float8_e4m3fn::from_bits()); |
| 36 | #else |
| 37 | // Cast implemented by pytorch. Uses bit manipulation instead of HW cvt. |
| 38 | // HW cvt above is faster when it is available (ROCm 6.3 or newer). |
| 39 | return static_cast<c10::Float8_e4m3fn>(r); |
| 40 | #endif |
| 41 | } |
| 42 | |
| 43 | template <> |
| 44 | __device__ __forceinline__ c10::Float8_e4m3fnuz cvt_c10(float const r) { |
| 45 | return c10::Float8_e4m3fnuz( |
| 46 | __hip_cvt_float_to_fp8(r, __hip_fp8_e4m3_fnuz::__default_saturation, |
| 47 | __hip_fp8_e4m3_fnuz::__default_interpret), |
| 48 | c10::Float8_e4m3fnuz::from_bits()); |
| 49 | } |
| 50 | |
| 51 | template <typename Tout, typename Tin> |
| 52 | __inline__ __device__ Tout vec_conversion(const Tin& x) { |
| 53 | return x; |
| 54 | } |
| 55 | |
| 56 | template <typename Tout, typename Tin> |
| 57 | __inline__ __device__ Tout scaled_vec_conversion(const Tin& x, |
| 58 | const float scale) { |
| 59 | return x; |
| 60 | } |
| 61 | |
| 62 | #if HIP_FP8_TYPE_OCP |
| 63 | using fp8_type = __hip_fp8_e4m3; |
| 64 | using fp8x2_type = __hip_fp8x2_e4m3; |
| 65 | #else |
| 66 | using fp8_type = __hip_fp8_e4m3_fnuz; |
| 67 | using fp8x2_type = __hip_fp8x2_e4m3_fnuz; |
| 68 | #endif |
| 69 | |
| 70 | // fp8x2 -> float2 |
| 71 | template <> |
| 72 | __inline__ __device__ float2 |
| 73 | vec_conversion<float2, uint16_t>(const uint16_t& a) { |
| 74 | fp8x2_type f8x2; |
| 75 | f8x2.__x = a; |
| 76 | return static_cast<float2>(f8x2); |
| 77 | } |
| 78 | |
| 79 | // fp8x4 -> float4 |
| 80 | template <> |
| 81 | __inline__ __device__ Float4_ |
| 82 | vec_conversion<Float4_, uint32_t>(const uint32_t& a) { |
| 83 | Float4_ res; |
| 84 | res.x = vec_conversion<float2, uint16_t>((uint16_t)a); |
| 85 | res.y = vec_conversion<float2, uint16_t>((uint16_t)(a >> 16U)); |
| 86 | return res; |
| 87 | } |
| 88 | |
| 89 | // fp8x4 -> float4 |
| 90 | template <> |
| 91 | __inline__ __device__ float4 |
| 92 | vec_conversion<float4, uint32_t>(const uint32_t& a) { |
| 93 | Float4_ tmp = vec_conversion<Float4_, uint32_t>(a); |
| 94 | float4 res = make_float4(tmp.x.x, tmp.x.y, tmp.y.x, tmp.y.y); |
| 95 | return res; |
| 96 | } |
| 97 | |
| 98 | /* Scaled and vectorized conversions, for data exchange between high and low |
| 99 | precision domains |
| 100 | |
| 101 | Convention of the scale in API, e.g: FP8_data = Quantization( |
| 102 | High_Precision_data / scale ) s.t. Quantize(HP / scale) => FP8 Dequant(FP8) * |
| 103 | scale => HP |
| 104 | |
| 105 | */ |
| 106 | |
| 107 | using __nv_bfloat16 = __hip_bfloat16; |
| 108 | using __nv_bfloat162 = __hip_bfloat162; |
| 109 | |
| 110 | // fp8 -> __nv_bfloat16 |
| 111 | template <> |
| 112 | __inline__ __device__ __nv_bfloat16 |
| 113 | scaled_vec_conversion<__nv_bfloat16, uint8_t>(const uint8_t& a, float scale) { |
| 114 | fp8_type f8; |
| 115 | f8.__x = a; |
| 116 | return __float2bfloat16(static_cast<float>(f8) * scale); |
| 117 | } |
| 118 | |
| 119 | // fp8x2 -> __nv_bfloat162 |
| 120 | template <> |
| 121 | __inline__ __device__ __nv_bfloat162 |
| 122 | scaled_vec_conversion<__nv_bfloat162, uint16_t>(const uint16_t& a, |
| 123 | float scale) { |
| 124 | __nv_bfloat162 res; |
| 125 | res.x = scaled_vec_conversion<__nv_bfloat16, uint8_t>((uint8_t)a, scale); |
| 126 | res.y = |
| 127 | scaled_vec_conversion<__nv_bfloat16, uint8_t>((uint8_t)(a >> 8U), scale); |
| 128 | return res; |
| 129 | } |
| 130 | |
| 131 | // fp8x4 -> bf16_4_t |
| 132 | template <> |
| 133 | __inline__ __device__ bf16_4_t |
| 134 | scaled_vec_conversion<bf16_4_t, uint32_t>(const uint32_t& a, float scale) { |
| 135 | bf16_4_t res; |
| 136 | res.x = scaled_vec_conversion<__nv_bfloat162, uint16_t>((uint16_t)a, scale); |
| 137 | res.y = scaled_vec_conversion<__nv_bfloat162, uint16_t>((uint16_t)(a >> 16U), |
| 138 | scale); |
| 139 | return res; |
| 140 | } |
| 141 | |
| 142 | // fp8x8 -> bf16_8_t |
| 143 | template <> |
| 144 | __inline__ __device__ bf16_8_t |
| 145 | scaled_vec_conversion<bf16_8_t, uint2>(const uint2& a, float scale) { |
| 146 | bf16_4_t tmp1, tmp2; |
| 147 | tmp1 = scaled_vec_conversion<bf16_4_t, uint32_t>(a.x, scale); |
| 148 | tmp2 = scaled_vec_conversion<bf16_4_t, uint32_t>(a.y, scale); |
| 149 | bf16_8_t res; |
| 150 | res.x = tmp1.x; |
| 151 | res.y = tmp1.y; |
| 152 | res.z = tmp2.x; |
| 153 | res.w = tmp2.y; |
| 154 | return res; |
| 155 | } |
| 156 | |
| 157 | // fp8 -> float |
| 158 | template <> |
| 159 | __inline__ __device__ float scaled_vec_conversion<float, uint8_t>( |
| 160 | const uint8_t& a, float scale) { |
| 161 | fp8_type f8; |
| 162 | f8.__x = a; |
| 163 | return static_cast<float>(f8) * scale; |
| 164 | } |
| 165 | |
| 166 | // fp8x2 -> float2 |
| 167 | template <> |
| 168 | __inline__ __device__ float2 |
| 169 | scaled_vec_conversion<float2, uint16_t>(const uint16_t& a, float scale) { |
| 170 | fp8x2_type f8x2; |
| 171 | f8x2.__x = a; |
| 172 | return static_cast<float2>(f8x2) * scale; |
| 173 | } |
| 174 | |
| 175 | // fp8x4 -> float4 |
| 176 | template <> |
| 177 | __inline__ __device__ Float4_ |
| 178 | scaled_vec_conversion<Float4_, uint32_t>(const uint32_t& a, const float scale) { |
| 179 | Float4_ res; |
| 180 | res.x = scaled_vec_conversion<float2, uint16_t>((uint16_t)a, scale); |
| 181 | res.y = scaled_vec_conversion<float2, uint16_t>((uint16_t)(a >> 16U), scale); |
| 182 | return res; |
| 183 | } |
| 184 | |
| 185 | // fp8x4 -> float4 |
| 186 | template <> |
| 187 | __inline__ __device__ float4 |
| 188 | scaled_vec_conversion<float4, uint32_t>(const uint32_t& a, float scale) { |
| 189 | Float4_ res = scaled_vec_conversion<Float4_, uint32_t>(a, scale); |
| 190 | return {res.x.x, res.x.y, res.y.x, res.y.y}; |
| 191 | } |
| 192 | |
| 193 | // fp8x8 -> float8 |
| 194 | template <> |
| 195 | __inline__ __device__ Float8_ |
| 196 | scaled_vec_conversion<Float8_, uint2>(const uint2& a, float scale) { |
| 197 | Float4_ tmp1, tmp2; |
| 198 | tmp1 = scaled_vec_conversion<Float4_, uint32_t>(a.x, scale); |
| 199 | tmp2 = scaled_vec_conversion<Float4_, uint32_t>(a.y, scale); |
| 200 | Float8_ res; |
| 201 | res.x = tmp1.x; |
| 202 | res.y = tmp1.y; |
| 203 | res.z = tmp2.x; |
| 204 | res.w = tmp2.y; |
| 205 | return res; |
| 206 | } |
| 207 | |
| 208 | // fp8 -> half |
| 209 | template <> |
| 210 | __inline__ __device__ uint16_t |
| 211 | scaled_vec_conversion<uint16_t, uint8_t>(const uint8_t& a, float scale) { |
| 212 | __half_raw res; |
| 213 | res.data = scaled_vec_conversion<float, uint8_t>(a, scale); |
| 214 | return res.x; |
| 215 | } |
| 216 | |
| 217 | // fp8x2 -> half2 |
| 218 | template <> |
| 219 | __inline__ __device__ uint32_t |
| 220 | scaled_vec_conversion<uint32_t, uint16_t>(const uint16_t& a, float scale) { |
| 221 | union { |
| 222 | __half2_raw h2r; |
| 223 | uint32_t ui32; |
| 224 | } tmp; |
| 225 | tmp.h2r = __hip_cvt_fp8x2_to_halfraw2(a, fp8_type::__default_interpret); |
| 226 | tmp.h2r.x.data *= scale; |
| 227 | tmp.h2r.y.data *= scale; |
| 228 | return tmp.ui32; |
| 229 | } |
| 230 | |
| 231 | // fp8x4 -> half2x2 |
| 232 | template <> |
| 233 | __inline__ __device__ uint2 |
| 234 | scaled_vec_conversion<uint2, uint32_t>(const uint32_t& a, float scale) { |
| 235 | union { |
| 236 | uint2 u32x2; |
| 237 | uint32_t u32[2]; |
| 238 | } tmp; |
| 239 | tmp.u32[0] = scaled_vec_conversion<uint32_t, uint16_t>((uint16_t)a, scale); |
| 240 | tmp.u32[1] = |
| 241 | scaled_vec_conversion<uint32_t, uint16_t>((uint16_t)(a >> 16U), scale); |
| 242 | return tmp.u32x2; |
| 243 | } |
| 244 | |
| 245 | // fp8x8 -> half2x4 |
| 246 | template <> |
| 247 | __inline__ __device__ uint4 scaled_vec_conversion<uint4, uint2>(const uint2& a, |
| 248 | float scale) { |
| 249 | union { |
| 250 | uint4 u64x2; |
| 251 | uint2 u64[2]; |
| 252 | } tmp; |
| 253 | tmp.u64[0] = scaled_vec_conversion<uint2, uint32_t>(a.x, scale); |
| 254 | tmp.u64[1] = scaled_vec_conversion<uint2, uint32_t>(a.y, scale); |
| 255 | return tmp.u64x2; |
| 256 | } |
| 257 | |
| 258 | // half -> fp8 |
| 259 | template <> |
| 260 | __inline__ __device__ uint8_t |
| 261 | scaled_vec_conversion<uint8_t, uint16_t>(const uint16_t& a, float scale) { |
| 262 | __half_raw tmp; |
| 263 | tmp.x = a; |
| 264 | tmp.data /= scale; |
| 265 | return __hip_cvt_halfraw_to_fp8(tmp, fp8_type::__default_saturation, |
| 266 | fp8_type::__default_interpret); |
| 267 | } |
| 268 | |
| 269 | // halfx2 -> fp8x2 |
| 270 | template <> |
| 271 | __inline__ __device__ uint16_t |
| 272 | scaled_vec_conversion<uint16_t, uint32_t>(const uint32_t& a, float scale) { |
| 273 | union { |
| 274 | uint32_t ui32; |
| 275 | __half2_raw h2r; |
| 276 | } tmp; |
| 277 | tmp.ui32 = a; |
| 278 | tmp.h2r.x.data /= scale; |
| 279 | tmp.h2r.y.data /= scale; |
| 280 | return __hip_cvt_halfraw2_to_fp8x2(tmp.h2r, fp8_type::__default_saturation, |
| 281 | fp8_type::__default_interpret); |
| 282 | } |
| 283 | |
| 284 | // half2x2 -> fp8x4 |
| 285 | template <> |
| 286 | __inline__ __device__ uint32_t |
| 287 | scaled_vec_conversion<uint32_t, uint2>(const uint2& a, float scale) { |
| 288 | union { |
| 289 | uint16_t ui16[2]; |
| 290 | uint32_t ui32; |
| 291 | } tmp; |
| 292 | tmp.ui16[0] = scaled_vec_conversion<uint16_t, uint32_t>(a.x, scale); |
| 293 | tmp.ui16[1] = scaled_vec_conversion<uint16_t, uint32_t>(a.y, scale); |
| 294 | return tmp.ui32; |
| 295 | } |
| 296 | |
| 297 | // half2x4 -> fp8x8 |
| 298 | template <> |
| 299 | __inline__ __device__ uint2 scaled_vec_conversion<uint2, uint4>(const uint4& a, |
| 300 | float scale) { |
| 301 | union { |
| 302 | uint2 ui2[2]; |
| 303 | uint4 ui4; |
| 304 | } tmp; |
| 305 | tmp.ui4 = a; |
| 306 | uint2 res; |
| 307 | res.x = scaled_vec_conversion<uint32_t, uint2>(tmp.ui2[0], scale); |
| 308 | res.y = scaled_vec_conversion<uint32_t, uint2>(tmp.ui2[1], scale); |
| 309 | return res; |
| 310 | } |
| 311 | |
| 312 | // bf16 -> fp8 |
| 313 | template <> |
| 314 | __inline__ __device__ uint8_t scaled_vec_conversion<uint8_t, __nv_bfloat16>( |
| 315 | const __nv_bfloat16& a, float scale) { |
| 316 | return __hip_cvt_float_to_fp8(__bfloat162float(a) / scale, |
| 317 | fp8_type::__default_saturation, |
| 318 | fp8_type::__default_interpret); |
| 319 | } |
| 320 | |
| 321 | // bf16x2 -> fp8x2 |
| 322 | template <> |
| 323 | __inline__ __device__ uint16_t scaled_vec_conversion<uint16_t, __nv_bfloat162>( |
| 324 | const __nv_bfloat162& a, float scale) { |
| 325 | union { |
| 326 | uint8_t ui8[2]; |
| 327 | uint16_t ui16; |
| 328 | } tmp; |
| 329 | tmp.ui8[0] = scaled_vec_conversion<uint8_t, __nv_bfloat16>(a.x, scale); |
| 330 | tmp.ui8[1] = scaled_vec_conversion<uint8_t, __nv_bfloat16>(a.y, scale); |
| 331 | return tmp.ui16; |
| 332 | } |
| 333 | |
| 334 | // bf16x4 -> fp8x4 |
| 335 | template <> |
| 336 | __inline__ __device__ uint32_t |
| 337 | scaled_vec_conversion<uint32_t, bf16_4_t>(const bf16_4_t& a, float scale) { |
| 338 | union { |
| 339 | uint16_t ui16[2]; |
| 340 | uint32_t ui32; |
| 341 | } tmp; |
| 342 | tmp.ui16[0] = scaled_vec_conversion<uint16_t, __nv_bfloat162>(a.x, scale); |
| 343 | tmp.ui16[1] = scaled_vec_conversion<uint16_t, __nv_bfloat162>(a.y, scale); |
| 344 | return tmp.ui32; |
| 345 | } |
| 346 | |
| 347 | // bf16x8 -> fp8x8 |
| 348 | template <> |
| 349 | __inline__ __device__ uint2 |
| 350 | scaled_vec_conversion<uint2, bf16_8_t>(const bf16_8_t& a, float scale) { |
| 351 | uint2 res; |
| 352 | res.x = scaled_vec_conversion<uint32_t, bf16_4_t>({a.x, a.y}, scale); |
| 353 | res.y = scaled_vec_conversion<uint32_t, bf16_4_t>({a.z, a.w}, scale); |
| 354 | return res; |
| 355 | } |
| 356 | |
| 357 | // float -> fp8 |
| 358 | template <> |
| 359 | __inline__ __device__ uint8_t |
| 360 | scaled_vec_conversion<uint8_t, float>(const float& a, float scale) { |
| 361 | return __hip_cvt_float_to_fp8(a / scale, fp8_type::__default_saturation, |
| 362 | fp8_type::__default_interpret); |
| 363 | } |
| 364 | |
| 365 | // floatx2 -> fp8x2 |
| 366 | template <> |
| 367 | __inline__ __device__ uint16_t |
| 368 | scaled_vec_conversion<uint16_t, float2>(const float2& a, float scale) { |
| 369 | return __hip_cvt_float2_to_fp8x2(a / scale, fp8_type::__default_saturation, |
| 370 | fp8_type::__default_interpret); |
| 371 | } |
| 372 | |
| 373 | // floatx4 -> fp8x4 |
| 374 | template <> |
| 375 | __inline__ __device__ uint32_t |
| 376 | scaled_vec_conversion<uint32_t, float4>(const float4& a, float scale) { |
| 377 | union { |
| 378 | uint16_t ui16[2]; |
| 379 | uint32_t ui32; |
| 380 | } tmp; |
| 381 | tmp.ui16[0] = scaled_vec_conversion<uint16_t, float2>({a.x, a.y}, scale); |
| 382 | tmp.ui16[1] = scaled_vec_conversion<uint16_t, float2>({a.z, a.w}, scale); |
| 383 | return tmp.ui32; |
| 384 | } |
| 385 | #endif // ENABLE_FP8 |
| 386 | |
| 387 | template <typename Tout, typename Tin, Fp8KVCacheDataType kv_dt> |
| 388 | __inline__ __device__ Tout scaled_convert(const Tin& x, const float scale) { |
| 389 | #ifdef ENABLE_FP8 |
| 390 | if constexpr (kv_dt == Fp8KVCacheDataType::kFp8E4M3) { |
| 391 | return scaled_vec_conversion<Tout, Tin>(x, scale); |
| 392 | } |
| 393 | #endif |
| 394 | assert(false); |
| 395 | return {}; // Squash missing return statement warning |
| 396 | } |
| 397 | |
| 398 | // The following macro is used to dispatch the conversion function based on |
| 399 | // the data type of the key and value cache. The FN is a macro that calls a |
| 400 | // function with template<typename scalar_t, typename cache_t, |
| 401 | // Fp8KVCacheDataType kv_dt>. |
| 402 | #define DISPATCH_BY_KV_CACHE_DTYPE(SRC_DTYPE, KV_DTYPE, FN) \ |
| 403 | vllm::Fp8KVCacheDataType KV_CACHE_DTYPE = \ |
| 404 | vllm::get_fp8_kv_cache_data_type(KV_DTYPE); \ |
| 405 | if (KV_CACHE_DTYPE == vllm::Fp8KVCacheDataType::kAuto) { \ |
| 406 | if (SRC_DTYPE == torch::headeronly::ScalarType::Float) { \ |
| 407 | FN(float, float, vllm::Fp8KVCacheDataType::kAuto); \ |
| 408 | } else if (SRC_DTYPE == torch::headeronly::ScalarType::Half) { \ |
| 409 | FN(uint16_t, uint16_t, vllm::Fp8KVCacheDataType::kAuto); \ |
| 410 | } else if (SRC_DTYPE == torch::headeronly::ScalarType::BFloat16) { \ |
| 411 | FN(__nv_bfloat16, __nv_bfloat16, vllm::Fp8KVCacheDataType::kAuto); \ |
| 412 | } else { \ |
| 413 | STD_TORCH_CHECK(false, \ |
| 414 | "Unsupported input type of kv cache: ", SRC_DTYPE); \ |
| 415 | } \ |
| 416 | } else if (KV_CACHE_DTYPE == vllm::Fp8KVCacheDataType::kFp8E4M3) { \ |
| 417 | if (SRC_DTYPE == torch::headeronly::ScalarType::Float) { \ |
| 418 | FN(float, uint8_t, vllm::Fp8KVCacheDataType::kFp8E4M3); \ |
| 419 | } else if (SRC_DTYPE == torch::headeronly::ScalarType::Half) { \ |
| 420 | FN(uint16_t, uint8_t, vllm::Fp8KVCacheDataType::kFp8E4M3); \ |
| 421 | } else if (SRC_DTYPE == torch::headeronly::ScalarType::BFloat16) { \ |
| 422 | FN(__nv_bfloat16, uint8_t, vllm::Fp8KVCacheDataType::kFp8E4M3); \ |
| 423 | } else { \ |
| 424 | STD_TORCH_CHECK(false, \ |
| 425 | "Unsupported input type of kv cache: ", SRC_DTYPE); \ |
| 426 | } \ |
| 427 | } else { \ |
| 428 | STD_TORCH_CHECK(false, "Unsupported data type of kv cache: ", KV_DTYPE); \ |
| 429 | } |
| 430 | |
| 431 | } // namespace fp8 |
| 432 | #endif // USE_ROCM |
| 433 | } // namespace vllm |