diff --git a/mxrec_add_ons/rec_for_torch/operators/disetangle_attention/READMD.md b/mxrec_add_ons/rec_for_torch/operators/disetangle_attention/READMD.md
index f8a43f0a1579798625a313bb5d40dbaf23af2c34..fabdb712107f5efa483fed17fa0eec7a6d2f2e2b 100644
--- a/mxrec_add_ons/rec_for_torch/operators/disetangle_attention/READMD.md
+++ b/mxrec_add_ons/rec_for_torch/operators/disetangle_attention/READMD.md
@@ -73,10 +73,10 @@ c) 算子约束说明：
 
 * 支持的型号：Atlas A2系列产品;
 * 支持的CANN版本：8.2.RC1.alpha001及之后版本；
-* b: 最大不超过100
-* n: 最大不超过64
-* s: 当前要求是256
-* d: 当前要求是16的倍数,最大不超过512
+* b: 取值范围[0, 100]
+* n: 取值范围[0, 64]
+* s: 取值为256
+* d: 取值范围[1, 512]
 * 数据类型: fp16
 
 ## 算子逻辑
diff --git a/mxrec_add_ons/rec_for_torch/torch_plugin/torch_library/2.6.0/disetangle_attention/DisetangleAttenFusion.cpp b/mxrec_add_ons/rec_for_torch/torch_plugin/torch_library/2.6.0/disetangle_attention/DisetangleAttenFusion.cpp
index 1a7dbc444b0d047d425ba8885e4e1b8c519c8238..dac428408be21aae17f63ddcc1f9ddc551b639d5 100644
--- a/mxrec_add_ons/rec_for_torch/torch_plugin/torch_library/2.6.0/disetangle_attention/DisetangleAttenFusion.cpp
+++ b/mxrec_add_ons/rec_for_torch/torch_plugin/torch_library/2.6.0/disetangle_attention/DisetangleAttenFusion.cpp
@@ -18,6 +18,7 @@ constexpr uint32_t CONST_4 = 4;
 constexpr uint32_t MAX_BATCH = 100;
 constexpr uint32_t MAX_NUM = 64;
 constexpr uint32_t MAX_DIM = 512;
+constexpr uint32_t SUPPORT_SEQ = 256;
 
 bool pos_attr_safe_get(const std::string pos_attr_type, int& pos_attr)
 {
@@ -34,6 +35,26 @@ bool pos_attr_safe_get(const std::string pos_attr_type, int& pos_attr)
     return true;
 }
 
+void tensor_dtype_check(const at::Tensor& query_layer, const at::Tensor& key_layer, const at::Tensor& value_layer,
+                        const at::Tensor& pos_key_layer, const at::Tensor& pos_query_layer,
+                        const at::Tensor& relative_pos, const at::Tensor& attn_mask)
+{
+    TORCH_CHECK(query_layer.scalar_type() == at::kHalf,
+                "float16 query_layer tensor expected but got dtype: ", query_layer.scalar_type());
+    TORCH_CHECK(key_layer.scalar_type() == at::kHalf,
+                "float16 key_layer tensor expected but got dtype: ", key_layer.scalar_type());
+    TORCH_CHECK(value_layer.scalar_type() == at::kHalf,
+                "float16 value_layer tensor expected but got dtype: ", value_layer.scalar_type());
+    TORCH_CHECK(pos_key_layer.scalar_type() == at::kHalf,
+                "float16 pos_key_layer tensor expected but got dtype: ", pos_key_layer.scalar_type());
+    TORCH_CHECK(pos_query_layer.scalar_type() == at::kHalf,
+                "float16 pos_query_layer tensor expected but got dtype: ", pos_query_layer.scalar_type());
+    TORCH_CHECK(relative_pos.scalar_type() == at::kLong,
+                "int64 relative_pos tensor expected but got dtype: ", relative_pos.scalar_type());
+    TORCH_CHECK(attn_mask.scalar_type() == at::kHalf,
+                "float16 attn_mask tensor expected but got dtype: ", attn_mask.scalar_type());
+}
+
 void tensor_format_check(const at::Tensor& query_layer, const at::Tensor& key_layer, const at::Tensor& value_layer,
                          const at::Tensor& pos_key_layer, const at::Tensor& pos_query_layer,
                          const at::Tensor& relative_pos, const at::Tensor& attn_mask)
@@ -101,15 +122,17 @@ std::tuple<at::Tensor, at::Tensor, at::Tensor> DisetangleAttentionPTA(
     auto relative_pos_conti = relative_pos.contiguous();
     auto mask_conti = attn_mask.contiguous();
 
+    tensor_dtype_check(query_layer, key_layer, value_layer, pos_key_layer, pos_query_layer, relative_pos, attn_mask);
     tensor_format_check(query_layer, key_layer, value_layer, pos_key_layer, pos_query_layer, relative_pos, attn_mask);
 
     auto batch = query_layer.size(0);
     auto head = query_layer.size(1);
     auto seq = query_layer.size(2);
     auto dim = query_layer.size(3);
-    TORCH_CHECK(batch <= MAX_BATCH, "current max batch is 100 but get", batch);
-    TORCH_CHECK(head <= MAX_NUM, "current max head is 64 but get", head);
-    TORCH_CHECK(dim <= MAX_DIM, "current max dim is 512 but get", dim);
+    TORCH_CHECK(batch <= MAX_BATCH && batch >= 0, "current batch range [0, 100] but get ", batch);
+    TORCH_CHECK(head <= MAX_NUM && head >= 0, "current head range [0, 64] but get ", head);
+    TORCH_CHECK(dim <= MAX_DIM && dim >= CONST_1, "current dim range [1, 512] but get ", dim);
+    TORCH_CHECK(seq == SUPPORT_SEQ, "current seq only support 256 but get ", seq);
 
     at::Tensor attn_output = at::empty_like(query_layer_conti);
     at::Tensor attn_probs = at::empty({batch, head, seq, seq}, query_layer_conti.options());