[major] fix build on windows

src/kernels/gemm_w4a4.cu

@@ -1449,6 +1449,7 @@ public:
                 const int laneId = threadIdx.x % WARP_SIZE;
                 const int warpId = threadIdx.x / WARP_SIZE;
 
+                if constexpr (rank > 0) {
                     lora_act16_warp lora_act = load_lora_act(act + warpId * (LORA_M_TILES * LORA_R_TILES * 8 * WARP_SIZE), scales);
                     lora_wgt_warp lora_wgt   = load_lora_wgt(wgt);
                     for (int m = 0; m < LORA_M_TILES; m++) {
@@ -1463,6 +1464,7 @@ public:
                         }
                     }
                 }
+            }
 
             __device__ __forceinline__
             void operator()(const BlockInfo binfo, fpsum_warp &fpsum, half_t *out, int M, int N, int K, Arguments args) {
@@ -1498,6 +1500,7 @@ public:
                 const int laneId = threadIdx.x % WARP_SIZE;
                 const int warpId = threadIdx.x / WARP_SIZE;
 
+                if constexpr (rank > 0) {
                     lora_act_warp lora_act;
                     lora_act.fill(packed_f32psum_t::zeros());
 
@@ -1528,12 +1531,10 @@ public:
                         // }
                     }
 
-                
-
                     reduce_lora_act(act + warpId * (LORA_M_TILES * LORA_R_TILES * 8 * WARP_SIZE), lora_act);
 
-
                     // unused_var(dummy, alwaysfalse);
+                }
 
             }