diff --git a/3rdparty/tvm b/3rdparty/tvm
index 6d47d37dfe..5400532c4b 160000
--- a/3rdparty/tvm
+++ b/3rdparty/tvm
@@ -1 +1 @@
-Subproject commit 6d47d37dfe0e8f7bd079859d2aa744531887dacb
+Subproject commit 5400532c4ba37e8a30fcaac488c2ecb05a307e4f
diff --git a/python/mlc_llm/nn/kv_cache.py b/python/mlc_llm/nn/kv_cache.py
index 206e5d4958..4a058c6e03 100644
--- a/python/mlc_llm/nn/kv_cache.py
+++ b/python/mlc_llm/nn/kv_cache.py
@@ -244,6 +244,7 @@ def __init__(  # pylint: disable=too-many-locals
             rx.extern("flashinfer.attention_kernel_decode_with_paged_kv_cache_end_forward"),
             rx.extern("flashinfer.merge_state_in_place"),
             bb.add_func(llama_rope_with_position_map(rope_theta, rope_scale, head_dim, num_attention_heads, num_key_value_heads, dtype, rotary_dim), "tir_split_rotary"),
+            bb.add_func(_copy_single_page(num_key_value_heads, page_size, head_dim, dtype, target), "kv_cache_copy_single_page"),
             bb.add_func(_kv_cache_debug_get_kv(num_hidden_layers, num_key_value_heads, head_dim, dtype), "kv_cache_debug_get_kv"),
             # fmt: on
             # pylint: enable=line-too-long
@@ -347,6 +348,7 @@ def __init__(  # pylint: disable=too-many-locals
             bb.add_func(_attention_prefill_ragged(num_key_value_heads, num_attention_heads, head_dim, dtype, target), "tir_attention_prefill_ragged"),
             bb.add_func(_merge_state_inplace(num_attention_heads, head_dim, dtype, target), "tir_attention_merge_state"),
             bb.add_func(llama_rope_with_position_map(rope_theta, rope_scale, head_dim, num_attention_heads, num_key_value_heads, dtype, rotary_dim), "tir_split_rotary"),
+            bb.add_func(_copy_single_page(num_key_value_heads, page_size, head_dim, dtype, target), "kv_cache_copy_single_page"),
             bb.add_func(_kv_cache_debug_get_kv(num_hidden_layers, num_key_value_heads, head_dim, dtype), "kv_cache_debug_get_kv"),
             # fmt: on
             # pylint: enable=line-too-long
@@ -1539,3 +1541,43 @@ def apply_to_md(sch, block):
 
     apply_to_md(sch, sch.get_block("lse_store"))
     return sch.mod["main"].with_attr("tir.is_scheduled", 1)
+
+
+def _copy_single_page(num_heads, page_size, head_dim, dtype, target: Target):
+    tx = get_max_num_threads_per_block(target)
+
+    @T.prim_func
+    def copy_single_page(
+        var_pages: T.handle,
+        src_page_id: T.int64,
+        tgt_page_id: T.int64,
+        copy_length: T.int64,
+    ):
+        T.func_attr({"tir.is_scheduled": 1})
+        num_pages = T.int32()
+        pages = T.match_buffer(var_pages, (num_pages, 2, num_heads, page_size, head_dim), dtype)
+
+        for b in T.thread_binding(
+            (copy_length * num_heads * head_dim + tx - 1) // tx, thread="blockIdx.x"
+        ):
+            for t in T.thread_binding(tx, thread="threadIdx.x"):
+                with T.block("copy"):
+                    vh = T.axis.spatial(
+                        num_heads,
+                        T.Cast("int32", (b * tx + t) // (copy_length * head_dim)),
+                    )
+                    vp = T.axis.spatial(
+                        copy_length,
+                        (b * tx + t) % (copy_length * head_dim) // head_dim,
+                    )
+                    vd = T.axis.spatial(
+                        head_dim,
+                        T.Cast(
+                            "int32",
+                            (b * tx + t) % head_dim,
+                        ),
+                    )
+                    pages[tgt_page_id, 0, vh, vp, vd] = pages[src_page_id, 0, vh, vp, vd]
+                    pages[tgt_page_id, 1, vh, vp, vd] = pages[src_page_id, 1, vh, vp, vd]
+
+    return copy_single_page