A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 17, 03:00 UTC
[Hotfix][CI/Build][Kernel] CUDA 11.8 does not support layernorm optimizations (#3782)
Matt Wong committed
Apr 8, 2024 at 14:31 UTC
59a6abf3c99ee4fed5312d357f6ecbf857f24433
2 files changed
+6
-2
cmake/utils.cmake
+2
@@ -100,6 +100,8 @@ function (get_torch_gpu_compiler_flags OUT_GPU_FLAGS GPU_LANG)
100
101
if (CUDA_VERSION VERSION_GREATER_EQUAL 11.8)
102
list(APPEND GPU_FLAGS "-DENABLE_FP8_E5M2")
103
+ endif()
104
+ if (CUDA_VERSION VERSION_GREATER_EQUAL 12.0)
105
list(REMOVE_ITEM GPU_FLAGS
106
"-D__CUDA_NO_HALF_OPERATORS__"
107
"-D__CUDA_NO_HALF_CONVERSIONS__"
csrc/layernorm_kernels.cu
+4
-2
@@ -59,6 +59,8 @@ __global__ void rms_norm_kernel(
59
template<typename torch_type>
60
struct _typeConvert { static constexpr bool exists = false; };
61
62
+#if defined(USE_ROCM) || (defined(CUDA_VERSION) && (CUDA_VERSION >= 12000))
63
+// CUDA < 12.0 runs into issues with packed type conversion
64
template<>
65
struct _typeConvert<c10::Half> {
66
static constexpr bool exists = true;
@@ -85,8 +87,8 @@ struct _typeConvert<c10::BFloat16> {
87
__device__ static inline hip_type convert(float x) { return __float2bfloat16(x); }
88
__device__ static inline packed_hip_type convert(float2 x) { return __float22bfloat162_rn(x); }
89
};
88
-#endif
89
-
90
+#endif // defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 800
91
+#endif // defined(USE_ROCM) || (defined(CUDA_VERSION) && (CUDA_VERSION >= 12000))
92
93
/* Vector POD struct to generate vectorized and packed FP16/BF16 ops
94
for appropriate specializations of fused_add_rms_norm_kernel.