Merge remote-tracking branch 'origin/0.6.2-zhagnshao' into v0.6.2-dev

csrc/attention/attention_kernels_opt_tc.cu

@@ -235,7 +235,7 @@ __device__ void paged_attention_kernel_TC(
   
   __shared__ half4x2 q_vecs[REUSE_KV_TIMES][16];
   //if(thread_idx==0)printf("blockIdx.x==%d,q_boundary=%d,head_idx=%d,kv_head_idx=%d\n",blockIdx.x,q_boundary,head_idx,kv_head_idx);
-  for(int i=0;i<REUSE_KV_TIMES;i++){
+    for(int i=0;i<q_boundary;i++){
     if(thread_idx<16){
       q_vecs[i][thread_idx]=*reinterpret_cast<const half4x2*>(q_ptr+i*HEAD_SIZE+thread_idx*8);
     }