refactor(runner): move kernel warmup into the shared runner lifecycle (warmup()) (#28739)

Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

python/sglang/test/kits/attention_unittest/attention_methods/dense_attention.py

@@ -325,6 +325,11 @@ class MockModelRunner(ModelRunner):
         self.pp_size = 1
         self.is_draft_worker = False
         self.spec_algorithm = SpeculativeAlgorithm.NONE
+        # The runner lifecycle warms up kernels in capture() / first execute()
+        # via BaseRunner.warmup(); this mock never calls init_backends and has no
+        # real kernels to warm up, so mark it done (warmup becomes a no-op for
+        # the runner-mode attention tests that drive capture directly).
+        self._kernel_warmed_up = True
         speculative_num_draft_tokens = (
             max(case.input_lens)
             if case.forward_mode.is_target_verify()

python/sglang/test/kits/attention_unittest/attention_methods/dsa_attention.py

@@ -306,6 +306,7 @@ class DSAMockModelRunner(ModelRunner):
         self.page_size = case.page_size
         self.model_config = model_config
         self.tp_size = 1
+        self._kernel_warmed_up = True
         self.dp_size = 1
         self.pp_size = 1
         self.server_args = make_mock_server_args(

python/sglang/test/kits/attention_unittest/attention_methods/dsv4_attention.py

@@ -412,6 +412,7 @@ class MockDSV4ModelRunner:
         self.sliding_window_size = DSV4_SWA_WINDOW
         self.use_mla_backend = True
         self.is_draft_worker = False
+        self._kernel_warmed_up = True
 
     @property
     def hybrid_gdn_config(self):

python/sglang/test/kits/attention_unittest/attention_methods/dual_chunk_attention.py

@@ -321,6 +321,7 @@ class DualChunkMockModelRunner(ModelRunner):
         self.page_size = case.page_size
         self.model_config = model_config
         self.tp_size = 1
+        self._kernel_warmed_up = True
         self.dp_size = 1
         self.pp_size = 1
         self.server_args = make_mock_server_args(

python/sglang/test/kits/attention_unittest/attention_methods/gdn_attention.py

@@ -304,6 +304,7 @@ class MockGDNModelRunner(ModelRunner):
         self.sliding_window_size = None
         self.use_mla_backend = False
         self.is_draft_worker = False
+        self._kernel_warmed_up = True
 
     @property
     def hybrid_gdn_config(self):

python/sglang/test/kits/attention_unittest/attention_methods/kda_attention.py

@@ -310,6 +310,7 @@ class MockKDAModelRunner(ModelRunner):
         self.sliding_window_size = None
         self.use_mla_backend = False
         self.is_draft_worker = False
+        self._kernel_warmed_up = True
 
     @property
     def hybrid_gdn_config(self):

python/sglang/test/kits/attention_unittest/attention_methods/lightning_attention.py

@@ -319,6 +319,7 @@ class MockLightningModelRunner(ModelRunner):
         self.sliding_window_size = None
         self.use_mla_backend = False
         self.is_draft_worker = False
+        self._kernel_warmed_up = True
 
     @property
     def hybrid_gdn_config(self):

python/sglang/test/kits/attention_unittest/attention_methods/mamba2_attention.py

@@ -454,6 +454,7 @@ class MockMamba2ModelRunner(ModelRunner):
         self.sliding_window_size = None
         self.use_mla_backend = False
         self.is_draft_worker = False
+        self._kernel_warmed_up = True
 
     @property
     def hybrid_gdn_config(self):

python/sglang/test/kits/attention_unittest/attention_methods/mla_attention.py

@@ -306,6 +306,7 @@ class MockMLAModelRunner(ModelRunner):
         self.sliding_window_size = None
         self.use_mla_backend = True
         self.is_draft_worker = False
+        self._kernel_warmed_up = True
 
     @property
     def hybrid_gdn_config(self):