[Spec] Remove the dead spec V1 scheduler paths (#27977)

test/registered/spec/eagle/test_spec_eagle_topk.py

@@ -27,7 +27,7 @@ class TestEagle3Topk16(Eagle3Base, SpecCorrectnessKit, SpecAccuracyKit, SpecLogp
 
     spec_topk = 16
     spec_tokens = 64
-    disable_overlap = True  # topk>1 -> spec v1
+    disable_overlap = True  # synchronous baseline; SpecV2 subclass flips overlap on
     cuda_graph_max_bs = 5
     acc_length_thres = 3.1
     batch_accept_len_thres = 1.75

test/registered/unit/hardware_backend/mlx/test_attention_patching.py

@@ -1125,6 +1125,8 @@ class TestMlxOverlapScheduler(unittest.TestCase):
         # (deferred input materialization) before launching the forward.
         # Without resolve_forward_inputs in _launch_fresh, input_ids stays
         # None and async_forward_batch_generation_mlx dereferences a None.
+        from sglang.srt.speculative.spec_info import SpeculativeAlgorithm
+
         class _StopLoop(Exception):
             pass
 
@@ -1151,7 +1153,8 @@ class TestMlxOverlapScheduler(unittest.TestCase):
             prefill_input_ids_cpu=torch.tensor([1, 2, 3], dtype=torch.int64),
             input_ids=None,
             mix_running_indices=None,
-            is_spec_v2=False,
+            enable_overlap=True,
+            spec_algorithm=SpeculativeAlgorithm.NONE,
             device="cpu",
         )
         scheduler.get_next_batch_to_run = lambda: batch

test/registered/unit/spec/test_decode_bookkeeping_ownership.py

@@ -69,10 +69,6 @@ _OWNER_SITES = {
         "DFlashDraftInputV2.prepare_for_decode",
         "kv_allocated_len",
     ): 1,
-    # spec v1: each verify path owns its own settlement
-    ("speculative/eagle_info.py", "EagleVerifyInput.verify", "kv_committed_len"): 1,
-    ("speculative/eagle_info.py", "EagleVerifyInput.verify", "kv_allocated_len"): 1,
-    ("speculative/eagle_info.py", "EagleVerifyInput.verify", "spec_verify_ct"): 1,
     # disaggregation decode prealloc
     (
         "disaggregation/decode.py",

test/registered/unit/spec/test_spec_registry.py

@@ -162,16 +162,22 @@ class TestCustomSpecAlgoInterface(_RegistryIsolated):
         self.assertEqual(self.algo.is_some(), not self.algo.is_none())
         self.assertEqual(SpeculativeAlgorithm.EAGLE.is_some(), self.algo.is_some())
 
-    def test_supports_spec_v2_follows_supports_overlap(self):
-        # Plugin registered with supports_overlap=False -> not spec_v2.
-        self.assertFalse(self.algo.supports_spec_v2())
+    def test_supports_overlap_false_warns_deprecation(self):
+        # supports_overlap=False plugins run the V2 schema synchronously; the
+        # removed V1 path is surfaced as a deprecation warning at create time.
+        server_args = MagicMock()
+        server_args.disable_overlap_schedule = True
+        with self.assertLogs("sglang.srt.speculative.spec_registry", "WARNING") as logs:
+            self.algo.create_worker(server_args)
+        self.assertTrue(any("deprecated" in line for line in logs.output))
 
         @SpeculativeAlgorithm.register("MY_V2", supports_overlap=True)
         def _factory(server_args):
             return MagicMock
 
         v2 = SpeculativeAlgorithm.from_string("MY_V2")
-        self.assertTrue(v2.supports_spec_v2())
+        server_args.disable_overlap_schedule = False
+        self.assertIs(v2.create_worker(server_args), MagicMock)
 
     def test_create_worker_calls_factory(self):
         server_args = MagicMock()