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 11, 08:45 UTC
[NIXL][Bugfix] Fix NIXL/RDMA registration failure over CuMemAllocator (#29569)
Signed-off-by: Somoku <linsh0@protonmail.com> Co-authored-by: Nicolò Lucchesi <nlucches@redhat.com>
Sheng Lin committed
Dec 17, 2025 at 17:52 UTC
f4e884f2224a25612eaeaeac2a854c1dd330c144
1 file changed
+10
csrc/cumem_allocator.cpp
+10
@@ -107,6 +107,16 @@ void create_and_map(unsigned long long device, ssize_t size, CUdeviceptr d_mem,
107
prop.location.id = device;
108
prop.allocFlags.compressionType = CU_MEM_ALLOCATION_COMP_NONE;
109
110
+#ifndef USE_ROCM
111
+ int flag = 0;
112
+ CUDA_CHECK(cuDeviceGetAttribute(
113
+ &flag, CU_DEVICE_ATTRIBUTE_GPU_DIRECT_RDMA_WITH_CUDA_VMM_SUPPORTED,
114
+ device));
115
+ if (flag) { // support GPUDirect RDMA if possible
116
+ prop.allocFlags.gpuDirectRDMACapable = 1;
117
+ }
118
+#endif
119
+
120
#ifndef USE_ROCM
121
// Allocate memory using cuMemCreate
122
CUDA_CHECK(cuMemCreate(p_memHandle, size, &prop, 0));