[Fix] Try to fix nvcc compilation error (#21246)

python/sglang/jit_kernel/all_reduce.py

@@ -131,6 +131,8 @@ def get_custom_all_reduce_cls() -> type[CustomAllReduceObj]:
 
     @tvm_ffi.register_object("sgl.CustomAllReduce")
     class CustomAllReduceObjReal(tvm_ffi.Object):
+        __slots__ = ("__dict__",)
+
         def __init__(
             self,
             rank: int,

python/sglang/jit_kernel/include/sgl_kernel/distributed/custom_all_reduce.cuh

@@ -194,8 +194,8 @@ struct CustomAllReduceBase : public tvm::ffi::Object {
         for (const auto j : irange(new_registered_count)) {
           /// NOTE: structural binding will cause intern compiler error...
           const auto elem = array[j];
-          const auto offset = get<0>(elem);
-          const auto ipc_handle = get<1>(elem);
+          const auto offset = elem.get<0>();
+          const auto ipc_handle = elem.get<1>();
           data[j].input[i] = pointer::offset(open_cached(ipc_handle), offset);
         }
       }

python/sglang/jit_kernel/tests/test_custom_all_reduce.py

@@ -34,15 +34,13 @@ from sglang.srt.distributed.device_communicators.custom_all_reduce_v2 import (
 from sglang.test.ci.ci_register import register_cuda_ci
 
 register_cuda_ci(
-    est_time=120,
-    suite="stage-b-kernel-unit-1-gpu-large",
-    disabled="requires multi-GPU distributed setup",
+    est_time=500,
+    suite="stage-b-kernel-unit-8-gpu-h200",
 )
 register_cuda_ci(
-    est_time=120,
-    suite="nightly-kernel-1-gpu",
+    est_time=500,
+    suite="nightly-kernel-8-gpu-h200",
     nightly=True,
-    disabled="requires multi-GPU distributed setup",
 )
 
 # ---------------------------------------------------------------------------