diff --git a/csrc/api/dense_decode.h b/csrc/api/dense_decode.h
index 7df178a..a69faef 100644
--- a/csrc/api/dense_decode.h
+++ b/csrc/api/dense_decode.h
@@ -58,7 +58,7 @@ dense_attn_decode_interface(
     const int num_heads_q = sizes[2];
     const int head_size_k = sizes[3];
     TORCH_CHECK(head_size_k == 576 || head_size_k == 512, "Only head_size_k == 576 or 512 is supported");
-    TORCH_CHECK(head_size_v == 512, "Only head_size_v == 576 is supported");
+    TORCH_CHECK(head_size_v == 512, "Only head_size_v == 512 is supported");
     
     const int max_num_blocks_per_seq = block_table.size(1);
     const int num_blocks = kcache.size(0);
diff --git a/csrc/utils.h b/csrc/utils.h
index 8de676f..6423c45 100644
--- a/csrc/utils.h
+++ b/csrc/utils.h
@@ -46,14 +46,6 @@ do { \
 } while (0)
 #endif
 
-#ifndef TRAP_ONLY_DEVICE_ASSERT
-#define TRAP_ONLY_DEVICE_ASSERT(cond) \
-do { \
-    if (not (cond)) \
-        asm("trap;"); \
-} while (0)
-#endif
-
 
 struct RingBufferState {
     uint32_t cur_block_idx = 0u;