issue/282: 算子贴错导致推理问题修复

src/infiniop/ops/causal_softmax/maca/causal_softmax_kernel.h

@@ -18,7 +18,7 @@ INFINIOP_MACA_KERNEL causalSoftmax(
 
     // [Reduce] Find max value in each row and store in shared memory
     __shared__ Tdata max_;
-    Tdata max_0 = op::common_maca::reduce_op::max<BLOCK_SIZE, Tdata>(x, width);
+    Tdata max_0 = op::common_maca::reduce_op::max<BLOCK_SIZE, Tdata>(x, width - height + 1 + blockIdx.x);
     if (threadIdx.x == 0) {
         max_ = max_0;
     }