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 4, 02:30 UTC
use HIP_VERSION variables to guard against duplicate atomicAdd definitions (#41802)
Signed-off-by: Philip Maybank <pmaybank@amd.com> Co-authored-by: TJian <tunjian.tan@embeddedllm.com>
pmaybank committed
May 8, 2026 at 18:44 UTC
6881c754e1ce4b154ad48e8532d4680e3a54096c
1 file changed
+2
-2
csrc/quantization/gptq/compat.cuh
+2
-2
@@ -42,8 +42,8 @@ __device__ __forceinline__ void atomicAdd_half2(half2* address, half2 val) {
42
}
43
44
//
45
-
46
-#if defined(__CUDA_ARCH__) || defined(USE_ROCM)
45
+#if defined(__CUDA_ARCH__) || \
46
+ (defined(USE_ROCM) && (HIP_VERSION_MAJOR * 100 + HIP_VERSION_MINOR) < 713)
47
#if __CUDA_ARCH__ < 700 || defined(USE_ROCM)
48
49
__device__ __forceinline__ void atomicAdd(half* address, half val) {