fix cpu kv kernel bug

Author: wangzaijun

lightllm/common/basemodel/triton_kernel/kv_cache_offload.py

@@ -252,7 +252,7 @@ def offload_gpu_kv_to_cpu(
         layer_num=gpu_kv_cache.shape[0],
         head_dim=head_dim,
         block_num=page_num,
-        cpu_k_start_head_index=cpu_v_start_head_index,
+        cpu_k_start_head_index=cpu_k_start_head_index,
         cpu_k_head_num=cpu_k_head_num,
         gpu_k_start_head_index=gpu_k_start_head_index,
         gpu_k_head_num=gpu_k_head_num,
@@ -509,7 +509,7 @@ def load_cpu_kv_to_gpu(
         cpu_stride4=cpu_kv_cache.stride(4),
         layer_num=gpu_kv_cache.shape[0],
         head_dim=head_dim,
-        cpu_k_start_head_index=cpu_v_start_head_index,
+        cpu_k_start_head_index=cpu_k_start_head_index,
         cpu_k_head_num=cpu_k_head_num,
         gpu_k_start_head_index=gpu_k_start_head_index,
         gpu_k_head_num=gpu_k_head_num,