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

Merge branch 'main' into wye-refactor-quant-folder

Browse Source
This commit is contained in:
yewentao256
2025-09-13 07:38:47 -07:00
parent 1e3e56abfc 15b8fef453
commit e925187f6d
106 changed files with 2761 additions and 997 deletions
Download Patch File Download Diff File Expand all files Collapse all files

19
csrc/quantization/w8a8/fp8/nvidia/quant_utils.cuh
View File

@@ -12,13 +12,26 @@ namespace vllm {
namespace fp8 {
#ifdef ENABLE_FP8
#if 0 // Disable the following code to reduce the binary size.
template <typename Tout, typename Tin>
__inline__ __device__ Tout
vec_conversion(const Tin &x, const __nv_fp8_interpretation_t fp8_type) {
__inline__ __device__ Tout vec_conversion(
const Tin& x, const __nv_fp8_interpretation_t fp8_type = __NV_E4M3) {
return x;
}
// float -> c10::Float8_e4m3fn
template <>
__inline__ __device__ c10::Float8_e4m3fn
vec_conversion<c10::Float8_e4m3fn, float>(
const float& a, const __nv_fp8_interpretation_t fp8_type) {
#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ < 800
return static_cast<c10::Float8_e4m3fn>(a);
#else
return c10::Float8_e4m3fn(__nv_cvt_float_to_fp8(a, __NV_SATFINITE, fp8_type),
c10::Float8_e4m3fn::from_bits());
#endif
}
#if 0 // Disable the following code to reduce the binary size.
// fp8 -> half
template <>
__inline__ __device__ uint16_t vec_conversion<uint16_t, uint8_t>(