Remove annoying warnings in sgl kernel build (#9905)

sgl-kernel/csrc/attention/cutlass_mla_kernel.cu

@@ -162,7 +162,7 @@ typename T::Fmha::Arguments args_from_options(
       // TODO(trevor-m): Change split_kv back to -1 when
       // https://github.com/NVIDIA/cutlass/issues/2274 is fixed. Split_kv=1 will
       // perform worse with larger context length and smaller batch sizes.
-      num_kv_splits, // split_kv
+      static_cast<int>(num_kv_splits), // split_kv
       nullptr,       // is_var_split_kv
   };
   // TODO(kaixih@nvidia): When split_kv=-1 and is_var_split_kv=false, we compute
@@ -259,7 +259,7 @@ int64_t cutlass_mla_get_workspace_size(int64_t max_seq_len, int64_t num_batches,
   // Assumes device 0 when getting sm_count.
   arguments.hw_info.sm_count =
       sm_count <= 0 ? cutlass::KernelHardwareInfo::query_device_multiprocessor_count(/*device_id=*/0) : sm_count;
-  arguments.split_kv = num_kv_splits;
+  arguments.split_kv = static_cast<int>(num_kv_splits);
   MlaSm100Type::Fmha::set_split_kv(arguments);
 
   return MlaSm100Type::Fmha::get_workspace_size(arguments);

sgl-kernel/csrc/gemm/dsv3_fused_a_gemm.cu

@@ -131,6 +131,7 @@ __device__ bool try_wait_barrier(uint64_t* smem_ptr, int phase_bit) {
       : "r"(smem_int_ptr), "r"(phase_bit));
   return static_cast<bool>(wait_complete);
 #endif
+  return false;
 }
 
 // Barrier arrive

sgl-kernel/csrc/gemm/nvfp4_expert_quant.cu

@@ -541,6 +541,11 @@ void quant_impl(
   }
 }
 
+// Avoid redefinition warnings
+#undef CHECK_CONTIGUOUS
+#undef CHECK_TH_CUDA
+#undef CHECK_INPUT
+
 /*Quantization entry for fp4 experts quantization*/
 #define CHECK_TH_CUDA(x, m) TORCH_CHECK(x.is_cuda(), m, "must be a CUDA tensor")
 #define CHECK_CONTIGUOUS(x, m) TORCH_CHECK(x.is_contiguous(), m, "must be contiguous")