【Inference】fix blha bug (#70466)

* fix blha bug

* fix blha bug

Author: gzy19990617

paddle/phi/kernels/fusion/gpu/block_attn.h

@@ -892,7 +892,9 @@ __global__ __launch_bounds__(THREADS_PER_BLOCK) void gqa_block_attention_kernel(
   float qk_maxs[GQA_SUB_PARTITION_SIZE];
 #pragma unroll
   for (int i = 0; i < GQA_SUB_PARTITION_SIZE; i++) {
-    qk_maxs[i] = -FLT_MAX;
+    // qk_maxs[i] = -FLT_MAX;
+    // initialize qk_maxs!!!
+    qk_maxs[i] = qk_smem[act_time_step * GQA_SUB_PARTITION_SIZE + i];
   }
 
   // threads in one block can process 'K_PER_ITER' keys