diff --git a/tests/v1/core/test_single_type_kv_cache_manager.py b/tests/v1/core/test_single_type_kv_cache_manager.py
index 0e3e8879359a..2c491e9bd52e 100644
--- a/tests/v1/core/test_single_type_kv_cache_manager.py
+++ b/tests/v1/core/test_single_type_kv_cache_manager.py
@@ -14,9 +14,14 @@
 )
 from vllm.v1.core.single_type_kv_cache_manager import (
     ChunkedLocalAttentionManager,
+    MambaManager,
     SlidingWindowManager,
 )
-from vllm.v1.kv_cache_interface import ChunkedLocalAttentionSpec, SlidingWindowSpec
+from vllm.v1.kv_cache_interface import (
+    ChunkedLocalAttentionSpec,
+    MambaSpec,
+    SlidingWindowSpec,
+)
 
 pytestmark = pytest.mark.cpu_test
 
@@ -190,6 +195,40 @@ def run_one_case(block_is_cached, expect_length):
     )
 
 
+def test_mamba_possible_cached_prefix_with_eagle_drop():
+    block_size = 2
+    mamba_spec = MambaSpec(
+        block_size=block_size,
+        shapes=(1, 1),
+        dtypes=(torch.float32,),
+    )
+
+    block_pool = BlockPool(
+        num_gpu_blocks=100, enable_caching=True, hash_block_size=block_size
+    )
+    block_hash_list = [BlockHash(str(i).encode()) for i in range(3)]
+
+    for i, block_hash in enumerate(block_hash_list):
+        block_pool.cached_block_hash_to_block.insert(
+            make_block_hash_with_group_id(block_hash, 0),
+            block_pool.blocks[i + 10],
+        )
+
+    computed_blocks = MambaManager.find_longest_cache_hit(
+        block_hashes=block_hash_list,
+        max_length=len(block_hash_list) * block_size,
+        kv_cache_group_ids=[0],
+        block_pool=block_pool,
+        kv_cache_spec=mamba_spec,
+        drop_eagle_block=True,
+        alignment_tokens=block_size,
+    )[0][0]
+
+    assert len(computed_blocks) == 2
+    assert computed_blocks[0] == block_pool.null_block
+    assert computed_blocks[1].block_id == 11
+
+
 def test_chunked_local_attention_remove_skipped_blocks():
     attention_spec = ChunkedLocalAttentionSpec(
         block_size=2,
diff --git a/vllm/v1/core/single_type_kv_cache_manager.py b/vllm/v1/core/single_type_kv_cache_manager.py
index 478effcd7469..0b6f1e2a547c 100644
--- a/vllm/v1/core/single_type_kv_cache_manager.py
+++ b/vllm/v1/core/single_type_kv_cache_manager.py
@@ -991,6 +991,12 @@ def find_longest_cache_hit(
 
         block_size = kv_cache_spec.block_size
         max_num_blocks = max_length // block_size
+        if drop_eagle_block and max_num_blocks > 0:
+            # Full-attention cache hit lookup matches one extra block and then
+            # drops that final block as it's only partially accepted. Mamba/GDN state
+            # blocks are [null, ..., state] so popping after a match removes the state
+            # Instead, we can only search up to the boundary (not include final)
+            max_num_blocks -= 1
         # Search from right to left and early stop when a match is found.
         for i in range(max_num_blocks - 1, -1, -1):
             if cached_block := block_pool.get_cached_block(