EngineX-Hygon/sglang
5
0
Fork 0
Code Issues Pull Requests Actions 7 Projects Releases Wiki Activity

适配v0.5.4

Browse Source
This commit is contained in:
maxiao
2025-10-25 12:16:25 +08:00
parent 1053e1be17
commit 251235c229
8 changed files with 213 additions and 49 deletions
Download Patch File Download Diff File Expand all files Collapse all files

38
sgl-kernel/include/utils.h
View File

@@ -369,25 +369,25 @@ __device__ __forceinline__ dstDtype castFromFloat(float val) {
#endif
// add FP8 support
#ifndef USE_ROCM
#include <c10/util/Float8_e4m3fn.h>
using FP8_TYPE = c10::Float8_e4m3fn;
C10_HOST_DEVICE constexpr auto FP8_E4M3_MAX = std::numeric_limits<FP8_TYPE>::max();
#else // USE_ROCM
#if HIP_FP8_TYPE_FNUZ
#include <c10/util/Float8_e4m3fnuz.h>
using FP8_TYPE = c10::Float8_e4m3fnuz;
constexpr auto FP8_E4M3_MAX = 224.0f;
#else
#if HIP_FP8_TYPE_E4M3
#include <c10/util/Float8_e4m3fn.h>
using FP8_TYPE = c10::Float8_e4m3fn;
C10_HOST_DEVICE constexpr auto FP8_E4M3_MAX = std::numeric_limits<FP8_TYPE>::max();
#else
#error "fp8 is not supported in this processor (arch < gfx942)."
#endif // HIP_FP8_TYPE_E4M3
#endif // HIP_FP8_TYPE_FNUZ
#endif // USE_ROCM
// #ifndef USE_ROCM
// #include <c10/util/Float8_e4m3fn.h>
// using FP8_TYPE = c10::Float8_e4m3fn;
// C10_HOST_DEVICE constexpr auto FP8_E4M3_MAX = std::numeric_limits<FP8_TYPE>::max();
// #else // USE_ROCM
// #if HIP_FP8_TYPE_FNUZ
// #include <c10/util/Float8_e4m3fnuz.h>
// using FP8_TYPE = c10::Float8_e4m3fnuz;
// constexpr auto FP8_E4M3_MAX = 224.0f;
// #else
// #if HIP_FP8_TYPE_E4M3
// #include <c10/util/Float8_e4m3fn.h>
// using FP8_TYPE = c10::Float8_e4m3fn;
// C10_HOST_DEVICE constexpr auto FP8_E4M3_MAX = std::numeric_limits<FP8_TYPE>::max();
// #else
// #error "fp8 is not supported in this processor (arch < gfx942)."
// #endif // HIP_FP8_TYPE_E4M3
// #endif // HIP_FP8_TYPE_FNUZ
// #endif // USE_ROCM
#define FULL_MASK 0xffffffff