[CI/Build] Enable FP8 on NVIDIA Thor (#39712)

Signed-off-by: DarkLight1337 <tlleungac@connect.ust.hk>
This commit is contained in:
Cyrus Leung
2026-04-29 09:48:52 -07:00
committed by GitHub
parent a80d6f150c
commit 4a42aba380
4 changed files with 13 additions and 38 deletions
+10 -35
View File
@@ -96,44 +96,14 @@ struct enable_sm90_or_later : Kernel {
};
template <typename Kernel>
struct enable_sm90_only : Kernel {
struct enable_sm100_to_sm120 : Kernel {
template <typename... Args>
CUTLASS_DEVICE void operator()(Args&&... args) {
#if defined __CUDA_ARCH__
#if __CUDA_ARCH__ == 900
#if (__CUDA_ARCH__ >= 1000 && __CUDA_ARCH__ < 1200)
Kernel::operator()(std::forward<Args>(args)...);
#else
printf("This kernel only supports sm90.\n");
asm("trap;");
#endif
#endif
}
};
template <typename Kernel>
struct enable_sm100f_only : Kernel {
template <typename... Args>
CUTLASS_DEVICE void operator()(Args&&... args) {
#if defined __CUDA_ARCH__
#if __CUDA_ARCH__ == 1000 || __CUDA_ARCH__ == 1030
Kernel::operator()(std::forward<Args>(args)...);
#else
printf("This kernel only supports sm100f.\n");
asm("trap;");
#endif
#endif
}
};
template <typename Kernel>
struct enable_sm100a_only : Kernel {
template <typename... Args>
CUTLASS_DEVICE void operator()(Args&&... args) {
#if defined __CUDA_ARCH__
#if __CUDA_ARCH__ == 1000
Kernel::operator()(std::forward<Args>(args)...);
#else
printf("This kernel only supports sm100a.\n");
printf("This kernel only supports sm[100, 120).\n");
asm("trap;");
#endif
#endif
@@ -148,7 +118,7 @@ struct enable_sm120_only : Kernel {
#if __CUDA_ARCH__ == 1200
Kernel::operator()(std::forward<Args>(args)...);
#else
printf("This kernel only supports sm120.\n");
printf("This kernel only supports sm120a.\n");
asm("trap;");
#endif
#endif
@@ -160,8 +130,13 @@ template <typename Kernel>
struct enable_sm120_family : Kernel {
template <typename... Args>
CUTLASS_DEVICE void operator()(Args&&... args) {
#if defined __CUDA_ARCH__ && (__CUDA_ARCH__ >= 1200 && __CUDA_ARCH__ < 1300)
#if defined __CUDA_ARCH__
#if (__CUDA_ARCH__ >= 1200 && __CUDA_ARCH__ < 1300)
Kernel::operator()(std::forward<Args>(args)...);
#else
printf("This kernel only supports sm120f.\n");
asm("trap;");
#endif
#endif
}
};
@@ -141,7 +141,7 @@ struct cutlass_3x_gemm_sm100 {
sizeof(typename CollectiveEpilogue::SharedStorage))>,
KernelSchedule>::CollectiveOp;
using GemmKernel = enable_sm100f_only<cutlass::gemm::kernel::GemmUniversal<
using GemmKernel = enable_sm100_to_sm120<cutlass::gemm::kernel::GemmUniversal<
Shape<int, int, int, int>, CollectiveMainloop, CollectiveEpilogue, void>>;
};
@@ -125,7 +125,7 @@ struct cutlass_3x_gemm_fp8_blockwise {
MainloopScheduler
>::CollectiveOp>;
using KernelType = enable_sm100f_only<cutlass::gemm::kernel::GemmUniversal<
using KernelType = enable_sm100_to_sm120<cutlass::gemm::kernel::GemmUniversal<
Shape<int, int, int, int>, CollectiveMainloop, CollectiveEpilogue>>;
struct GemmKernel : public KernelType {};
@@ -92,7 +92,7 @@ struct cutlass_3x_gemm_sm100_fp8 {
// -----------------------------------------------------------
// Kernel definition
// -----------------------------------------------------------
using GemmKernel = enable_sm100f_only<cutlass::gemm::kernel::GemmUniversal<
using GemmKernel = enable_sm100_to_sm120<cutlass::gemm::kernel::GemmUniversal<
Shape<int, int, int, int>, CollectiveMainloop, CollectiveEpilogue, void>>;
};