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 13, 04:30 UTC
[ROCm][Bugfix] Fix compilation errors with fused_qknorm_rope_kernel.cu (#28682)
Signed-off-by: Sage Moore <sage@neuralmagic.com>
Sage Moore committed
Nov 14, 2025 at 11:06 UTC
8977ffb5e6428a3e682d47d9ca8342ccab9916f8
1 file changed
+10
csrc/fused_qknorm_rope_kernel.cu
+10
@@ -37,6 +37,16 @@
37
38
#ifdef USE_ROCM
39
#define FINAL_MASK 0xffffffffffffffffULL
40
+
41
+ #if defined(HIP_VERSION) && HIP_VERSION < 70000000
42
+// On ROCm versions before 7.0, __syncwarp isn't defined. The below
43
+// implementation is copy/pasted from the implementation in ROCm 7.0
44
+__device__ inline void __syncwarp() {
45
+ __builtin_amdgcn_fence(__ATOMIC_RELEASE, "wavefront");
46
+ __builtin_amdgcn_wave_barrier();
47
+ __builtin_amdgcn_fence(__ATOMIC_ACQUIRE, "wavefront");
48
+}
49
+ #endif
50
#else
51
#define FINAL_MASK 0xffffffff
52
#endif