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

permute/unpermute kernel for moe optimization (#14568)

Browse Source 
Signed-off-by: Caleb_Du <Caleb_Du@zju.edu.cn>
This commit is contained in:
Caleb_Du
2025-05-03 02:31:55 +08:00
committed by GitHub
parent 0f87d8f7b2
commit 3e887d2e0c
19 changed files with 1474 additions and 28 deletions
Download Patch File Download Diff File Expand all files Collapse all files

53
csrc/moe/permute_unpermute_kernels/dispatch.h Normal file
View File

@@ -0,0 +1,53 @@
#pragma once
#include <cuda_fp8.h>
#define MOE_SWITCH(TYPE, ...) \
at::ScalarType _st = ::detail::scalar_type(TYPE); \
switch (_st) { \
__VA_ARGS__ \
default: \
TORCH_CHECK(false, "[moe permute]data type dispatch fail!") \
}
#define MOE_DISPATCH_CASE(enum_type, ...) \
case enum_type: { \
using scalar_t = ScalarType2CudaType<enum_type>::type; \
__VA_ARGS__(); \
break; \
}
#define MOE_DISPATCH_FLOAT_CASE(...) \
MOE_DISPATCH_CASE(at::ScalarType::Float, __VA_ARGS__) \
MOE_DISPATCH_CASE(at::ScalarType::Half, __VA_ARGS__) \
MOE_DISPATCH_CASE(at::ScalarType::BFloat16, __VA_ARGS__) \
MOE_DISPATCH_CASE(at::ScalarType::Float8_e5m2, __VA_ARGS__) \
MOE_DISPATCH_CASE(at::ScalarType::Float8_e4m3fn, __VA_ARGS__)
#define MOE_DISPATCH(TYPE, ...) \
MOE_SWITCH(TYPE, MOE_DISPATCH_FLOAT_CASE(__VA_ARGS__))
template <at::ScalarType type>
struct ScalarType2CudaType;
template <>
struct ScalarType2CudaType<at::ScalarType::Float> {
using type = float;
};
template <>
struct ScalarType2CudaType<at::ScalarType::Half> {
using type = half;
};
template <>
struct ScalarType2CudaType<at::ScalarType::BFloat16> {
using type = __nv_bfloat16;
};
// #if __CUDA_ARCH__ >= 890
// fp8
template <>
struct ScalarType2CudaType<at::ScalarType::Float8_e5m2> {
using type = __nv_fp8_e5m2;
};
template <>
struct ScalarType2CudaType<at::ScalarType::Float8_e4m3fn> {
using type = __nv_fp8_e4m3;
};
// #endif