Convert IPC dataclasses to msgspec.Struct with opt-in msgpack transport (#28688)

Co-authored-by: Lianmin Zheng <lianminzheng@gmail.com>

test/registered/core/test_srt_endpoint.py

@@ -31,6 +31,8 @@ from sglang.test.test_utils import (
 register_cuda_ci(est_time=134, stage="base-b", runner_config="1-gpu-small")
 register_amd_ci(est_time=130, suite="stage-b-test-1-gpu-small-amd")
 
+SERVER_ENV = {"SGLANG_USE_PICKLE_IPC": "0"}
+
 
 class TestSRTEndpoint(CustomTestCase):
     @classmethod
@@ -41,6 +43,7 @@ class TestSRTEndpoint(CustomTestCase):
             cls.model,
             cls.base_url,
             timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
+            env=SERVER_ENV,
             other_args=(
                 "--enable-custom-logit-processor",
                 "--mem-fraction-static",
@@ -469,6 +472,12 @@ class TestSRTEndpoint(CustomTestCase):
             response = requests.post(self.base_url + "/flush_cache")
             assert response.status_code == 200
 
+        server_info = requests.get(self.base_url + "/server_info").json()
+        page_size = server_info.get("page_size") or 1
+
+        def align_down(num_tokens):
+            return num_tokens // page_size * page_size
+
         def send_and_check_cached_tokens(input_ids):
             response = requests.post(
                 self.base_url + "/generate",
@@ -483,10 +492,14 @@ class TestSRTEndpoint(CustomTestCase):
             return response_json["meta_info"]["cached_tokens"]
 
         self.assertEqual(send_and_check_cached_tokens(range(0, 100)), 0)
-        self.assertEqual(send_and_check_cached_tokens(range(0, 10000)), 100)
-        self.assertEqual(send_and_check_cached_tokens(range(0, 10000)), 9999)
-        self.assertEqual(send_and_check_cached_tokens(range(0, 1000)), 999)
-        self.assertEqual(send_and_check_cached_tokens(range(0, 11000)), 10000)
+        self.assertEqual(send_and_check_cached_tokens(range(0, 10000)), align_down(100))
+        self.assertEqual(
+            send_and_check_cached_tokens(range(0, 10000)), align_down(9999)
+        )
+        self.assertEqual(send_and_check_cached_tokens(range(0, 1000)), align_down(999))
+        self.assertEqual(
+            send_and_check_cached_tokens(range(0, 11000)), align_down(10000)
+        )
 
     def test_get_server_info(self):
         response = requests.get(self.base_url + "/server_info")
@@ -648,6 +661,7 @@ class TestTokenizeDetokenize(CustomTestCase):
             cls.model,
             cls.base_url,
             timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
+            env=SERVER_ENV,
         )
         cls.tokenizer = get_tokenizer(cls.model)
 

test/registered/unit/entrypoints/test_server_info.py

@@ -22,10 +22,12 @@ Current coverage:
 
 import asyncio
 import dataclasses
+import json
 import unittest
 from types import SimpleNamespace
 
 from sglang.srt.entrypoints import http_server
+from sglang.srt.lora.lora_registry import LoRARef
 from sglang.srt.server_args import ServerArgs
 from sglang.test.ci.ci_register import register_cpu_ci
 from sglang.test.test_utils import CustomTestCase
@@ -33,7 +35,9 @@ from sglang.test.test_utils import CustomTestCase
 register_cpu_ci(est_time=5, suite="base-a-test-cpu")
 
 
-def _call_server_info_with(server_args: ServerArgs) -> dict:
+def _call_server_info_with(
+    server_args: ServerArgs, internal_states: list[dict] | None = None
+) -> dict:
     """Invoke `http_server.server_info()` against a stub global state.
 
     Bypasses the FastAPI HTTP layer (no TestClient): the handler is an
@@ -44,7 +48,7 @@ def _call_server_info_with(server_args: ServerArgs) -> dict:
     """
 
     async def _fake_internal_state():
-        return [{"max_req_input_len": 1024}]
+        return internal_states or [{"max_req_input_len": 1024}]
 
     stub_state = SimpleNamespace(
         tokenizer_manager=SimpleNamespace(
@@ -291,6 +295,31 @@ class TestServerInfoExistingFieldsPreserved(CustomTestCase):
         self.assertIn("kv_events", info)
         self.assertIsNotNone(info["kv_events"])
 
+    def test_lora_refs_are_json_serializable_dicts(self):
+        lora_ref = LoRARef(
+            lora_id="lora-id",
+            lora_name="adapter",
+            lora_path="/tmp/adapter",
+            pinned=True,
+        )
+        args = ServerArgs(model_path="dummy")
+        args.lora_paths = [lora_ref]
+
+        info = _call_server_info_with(
+            args,
+            internal_states=[{"lora_paths": [lora_ref]}],
+        )
+
+        expected = {
+            "lora_id": "lora-id",
+            "lora_name": "adapter",
+            "lora_path": "/tmp/adapter",
+            "pinned": True,
+        }
+        self.assertEqual(info["lora_paths"], [expected])
+        self.assertEqual(info["internal_states"][0]["lora_paths"], [expected])
+        json.dumps(info)
+
 
 if __name__ == "__main__":
     unittest.main()

test/registered/unit/managers/test_tokenizer_manager_rid_cleanup.py

@@ -13,10 +13,11 @@ Covers:
 """
 
 import asyncio
-import dataclasses
 import unittest
 from unittest.mock import AsyncMock, MagicMock, Mock
 
+import msgspec
+
 from sglang.test.ci.ci_register import register_cpu_ci
 from sglang.test.test_utils import CustomTestCase, maybe_stub_sgl_kernel
 
@@ -35,7 +36,7 @@ _NOT_FINISHED = object()  # Sentinel: request has not finished yet
 # Categorised by value shape so that _make_batch_str_output can assign
 # type-appropriate defaults without hardcoding every field name.
 # When a field is renamed upstream, the old name simply won't appear in
-# dataclasses.fields() and the new name will fall through to the
+# msgspec.structs.fields() and the new name will fall through to the
 # pattern-matching or safe fallback — no test breakage.
 # ---------------------------------------------------------------------------
 
@@ -145,7 +146,7 @@ def _make_abort_req(rid: str, abort_message: str = "Aborted") -> AbortReq:
 def _make_batch_str_output(rid: str, finished_reason=None) -> BatchStrOutput:
     """Create a minimal BatchStrOutput for a single request.
 
-    Uses dataclass field introspection so that new or renamed fields in
+    Uses struct field introspection so that new or renamed fields in
     BatchStrOutput don't break this test.  Only the fields that matter for
     test logic (rids, finished_reasons, output_strs) are set explicitly;
     all others receive type-appropriate defaults based on naming patterns.
@@ -159,7 +160,7 @@ def _make_batch_str_output(rid: str, finished_reason=None) -> BatchStrOutput:
         fr = finished_reason
 
     kwargs = {}
-    for f in dataclasses.fields(BatchStrOutput):
+    for f in msgspec.structs.fields(BatchStrOutput):
         if f.name == "rids":
             kwargs[f.name] = [rid]
         elif f.name == "finished_reasons":
@@ -176,8 +177,8 @@ def _make_batch_str_output(rid: str, finished_reason=None) -> BatchStrOutput:
             kwargs[f.name] = [None]
         # Fields with class defaults — skip, let the default be used
         elif (
-            f.default is not dataclasses.MISSING
-            or f.default_factory is not dataclasses.MISSING
+            f.default is not msgspec.NODEFAULT
+            or f.default_factory is not msgspec.NODEFAULT
         ):
             continue
         # Unknown required field — provide a safe per-request default.