[bug] Fix cache salt and extra keys for prefix cache isolation (#23300)

python/sglang/srt/managers/scheduler.py

@@ -613,7 +613,6 @@ class Scheduler(
         self.require_mlp_sync = require_mlp_sync(self.server_args)
 
     def init_tp_model_worker(self):
-
         worker_kwargs = dict(
             server_args=self.server_args,
             gpu_id=self.gpu_id,
@@ -826,7 +825,6 @@ class Scheduler(
 
                 self.tree_cache = SWAChunkCache(params)
         else:
-
             if envs.SGLANG_EXPERIMENTAL_CPP_RADIX_TREE.get():
                 # lazy import to avoid JIT overhead
                 from sglang.srt.mem_cache.radix_cache_cpp import RadixCacheCpp
@@ -1622,7 +1620,6 @@ class Scheduler(
         ):
             recv_reqs, abort_reqs = self.mm_receiver.process_waiting_requests(recv_reqs)
             for req, error_msg, error_code in abort_reqs:
-
                 status_code = (
                     HTTPStatus.BAD_REQUEST
                     if error_code == 400
@@ -1878,6 +1875,7 @@ class Scheduler(
                     self.metrics_collector if self.enable_metrics else None
                 ),
                 routing_key=recv_req.routing_key,
+                extra_key=recv_req.extra_key,
                 http_worker_ipc=recv_req.http_worker_ipc,
                 dllm_config=self.dllm_config,
                 time_stats=recv_req.time_stats,

python/sglang/srt/session/session_controller.py

@@ -235,6 +235,7 @@ class Session:
             return_routed_experts=req.return_routed_experts,
             priority=req.priority,
             routing_key=req.routing_key,
+            extra_key=req.extra_key,
             http_worker_ipc=req.http_worker_ipc,
             time_stats=req.time_stats,
         )

test/manual/openai_server/features/test_cache_report.py

@@ -24,6 +24,7 @@ class TestCacheReport(CustomTestCase):
             timeout=300,
             other_args=[
                 "--chunked-prefill-size=40",
+                "--attention-backend=triton",
                 "--enable-cache-report",
             ],
         )
@@ -206,6 +207,14 @@ class TestCacheReport(CustomTestCase):
 
     #     asyncio.run(run_test())
 
+    @staticmethod
+    def _get_cached_tokens(response) -> int:
+        """Extract cached_tokens from response, returning 0 if prompt_tokens_details is None."""
+        details = response.usage.prompt_tokens_details
+        if details is None:
+            return 0
+        return int(details.cached_tokens)
+
     def test_cache_salt_effectiveness(self):
         print("=" * 100)
         print("Testing cache_salt effectiveness")
@@ -221,7 +230,7 @@ class TestCacheReport(CustomTestCase):
             max_tokens=10,
             extra_body={"cache_salt": "salt1"},
         )
-        cached_tokens_1_first = int(response1.usage.prompt_tokens_details.cached_tokens)
+        cached_tokens_1_first = self._get_cached_tokens(response1)
         prompt_tokens_1 = int(response1.usage.prompt_tokens)
         print(
             f"First request with salt1 - cached_tokens: {cached_tokens_1_first}, prompt_tokens: {prompt_tokens_1}"
@@ -235,9 +244,7 @@ class TestCacheReport(CustomTestCase):
             max_tokens=10,
             extra_body={"cache_salt": "salt1"},
         )
-        cached_tokens_1_second = int(
+        cached_tokens_1_second = self._get_cached_tokens(response2)
-            response2.usage.prompt_tokens_details.cached_tokens
-        )
         print(
             f"Second request with salt1 - cached_tokens: {cached_tokens_1_second}, prompt_tokens: {prompt_tokens_1}"
         )
@@ -258,7 +265,7 @@ class TestCacheReport(CustomTestCase):
             max_tokens=10,
             extra_body={"cache_salt": "salt2"},
         )
-        cached_tokens_2_first = int(response3.usage.prompt_tokens_details.cached_tokens)
+        cached_tokens_2_first = self._get_cached_tokens(response3)
         print(f"First request with salt2 - cached_tokens: {cached_tokens_2_first}")
 
         # Verify no cache hit for different salt (should be similar to first request with salt1)
@@ -274,14 +281,12 @@ class TestCacheReport(CustomTestCase):
             max_tokens=10,
             extra_body={"cache_salt": "salt2"},
         )
-        cached_tokens_2_second = int(
+        cached_tokens_2_second = self._get_cached_tokens(response4)
-            response4.usage.prompt_tokens_details.cached_tokens
-        )
         print(f"Second request with salt2 - cached_tokens: {cached_tokens_2_second}")
 
         # Verify cache hit for salt2
         assert (
-            cached_tokens_2_second == cached_tokens_2_first
+            cached_tokens_2_second > cached_tokens_2_first
         ), "Should have cache hit with same cache_salt for salt2"