From 528e9b14900fc8a012f2599172e2a4576caafe1a Mon Sep 17 00:00:00 2001
From: Kebe <mail@kebe7jun.com>
Date: Mon, 2 Feb 2026 23:55:46 +0900
Subject: [PATCH] [Feature][Core] Support Fabric detection to adapt the MNNVL
 protocol for the GB series (#33540)

Signed-off-by: Kebe <mail@kebe7jun.com>
Signed-off-by: youkaichao <youkaichao@gmail.com>
Co-authored-by: Thomas Vegas <tvegas@nvidia.com>
Co-authored-by: youkaichao <youkaichao@gmail.com>
---
 csrc/cumem_allocator.cpp | 19 ++++++++++++++++++-
 1 file changed, 18 insertions(+), 1 deletion(-)

diff --git a/csrc/cumem_allocator.cpp b/csrc/cumem_allocator.cpp
index 6c2c18a66..58ce8f71a 100644
--- a/csrc/cumem_allocator.cpp
+++ b/csrc/cumem_allocator.cpp
@@ -115,11 +115,28 @@ void create_and_map(unsigned long long device, ssize_t size, CUdeviceptr d_mem,
   if (flag) {  // support GPUDirect RDMA if possible
     prop.allocFlags.gpuDirectRDMACapable = 1;
   }
+  int fab_flag = 0;
+  CUDA_CHECK(cuDeviceGetAttribute(
+      &fab_flag, CU_DEVICE_ATTRIBUTE_HANDLE_TYPE_FABRIC_SUPPORTED, device));
+  if (fab_flag) {  // support fabric handle if possible
+    prop.requestedHandleTypes = CU_MEM_HANDLE_TYPE_FABRIC;
+  }
 #endif
 
 #ifndef USE_ROCM
   // Allocate memory using cuMemCreate
-  CUDA_CHECK(cuMemCreate(p_memHandle, size, &prop, 0));
+  CUresult ret = (CUresult)cuMemCreate(p_memHandle, size, &prop, 0);
+  if (ret) {
+    if (fab_flag &&
+        (ret == CUDA_ERROR_NOT_PERMITTED || ret == CUDA_ERROR_NOT_SUPPORTED)) {
+      // Fabric allocation may fail without multi-node nvlink,
+      // fallback to POSIX file descriptor
+      prop.requestedHandleTypes = CU_MEM_HANDLE_TYPE_POSIX_FILE_DESCRIPTOR;
+      CUDA_CHECK(cuMemCreate(p_memHandle, size, &prop, 0));
+    } else {
+      CUDA_CHECK(ret);
+    }
+  }
   if (error_code != 0) {
     return;
   }