[Core] Include the LoRA path in prefix-cache block hashes (#59335)

Signed-off-by: Nick Hill <nickhill123@gmail.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>

tests/v1/core/test_kv_cache_utils.py

@@ -1022,7 +1022,7 @@ def test_generate_block_hash_extra_keys_lora():
     )
 
     extra_keys, _ = generate_block_hash_extra_keys(request, 0, 3, 0)
-    assert extra_keys == (("lora", "test_lora_adapter"),)
+    assert extra_keys == (("lora", "test_lora_adapter", "/path/to/lora"),)
 
     request.lora_request = None
     extra_keys, _ = generate_block_hash_extra_keys(request, 0, 3, 0)
@@ -1048,6 +1048,28 @@ def test_lora_name_and_cache_salt_block_hashes_do_not_collide(hash_fn):
     assert lora_req.block_hashes[0] != salted_req.block_hashes[0]
 
 
+@pytest.mark.parametrize("hash_fn", [sha256, sha256_cbor])
+def test_lora_path_change_changes_block_hashes(hash_fn):
+    """Re-pointing a LoRA name at another adapter must not reuse its blocks."""
+
+    def make_lora_request(lora_path: str) -> Request:
+        return Request(
+            request_id=lora_path,
+            prompt_token_ids=[0, 1, 2],
+            sampling_params=SamplingParams(max_tokens=1),
+            pooling_params=None,
+            lora_request=LoRARequest(
+                lora_name="foo", lora_int_id=1, lora_path=lora_path
+            ),
+            block_hasher=get_request_block_hasher(3, hash_fn),
+        )
+
+    first = make_lora_request("/path/to/lora_a")
+    second = make_lora_request("/path/to/lora_b")
+
+    assert first.block_hashes[0] != second.block_hashes[0]
+
+
 def test_to_event_extra_keys_keeps_untagged_event_format():
     """KV events keep publishing the extra-key shapes consumers already parse."""
     request = make_request(
@@ -1064,7 +1086,7 @@ def test_to_event_extra_keys_keeps_untagged_event_format():
     extra_keys, _ = generate_block_hash_extra_keys(request, 0, 5, 0)
 
     assert extra_keys == (
-        ("lora", "adapter"),
+        ("lora", "adapter", "/path/to/lora"),
         ("mm", "hash1", 2),
         ("cache_salt", "salt"),
     )

vllm/entrypoints/openai/models/serving.py

@@ -212,6 +212,16 @@ class OpenAIServingModels:
                     status_code=HTTPStatus.INTERNAL_SERVER_ERROR,
                 )
 
+            previous = self.lora_requests.get(lora_name)
+            if previous is not None and previous.lora_path == lora_path:
+                logger.warning(
+                    "Reloaded LoRA adapter '%s' in place from the same path '%s'. "
+                    "Prefix-cache blocks computed with its previous weights are "
+                    "not invalidated; load changed weights from a new path to "
+                    "avoid reusing them.",
+                    lora_name,
+                    lora_path,
+                )
             self.lora_requests[lora_name] = lora_request
             logger.info(
                 "Loaded new LoRA adapter: name '%s', path '%s'", lora_name, lora_path

vllm/v1/core/kv_cache_utils.py

@@ -565,20 +565,24 @@ def _gen_mm_extra_hash_keys(
     return extra_keys, curr_mm_idx
 
 
-def _gen_lora_extra_hash_keys(request: Request) -> list[tuple[str, str]]:
+def _gen_lora_extra_hash_keys(request: Request) -> list[tuple[str, str, str]]:
     """Generate extra keys related to LoRA for block hash computation.
 
+    The adapter path is included so that re-pointing a LoRA name at a different
+    adapter does not reuse KV computed with the previous one.
+
     Args:
         request: The request object.
 
     Returns:
-        Return LoRA name of the request if it is a LoRA request. Return empty
-        list otherwise.
+        Return the LoRA name and path of the request if it is a LoRA request.
+        Return empty list otherwise.
 
     """
-    if not request.lora_request:
+    lora_request = request.lora_request
+    if not lora_request:
         return []
-    return [("lora", request.lora_request.lora_name)]
+    return [("lora", lora_request.lora_name, lora_request.lora_path)]
 
 
 def _gen_prompt_embeds_extra_hash_keys(